fix(test): prima parte vuol dire tracciato, non fuori da vendor

Il controllo sui riferimenti esterni scandiva anche i file che git non
conosce. Una copia di comodo di una libreria lasciata in static/js/ --
fabric-debug.js -- sembrava codice nostro e faceva fallire la prova con gli
URL nei propri commenti, ma non sta nel repository e in reparto non ci
arriva mai. Su un checkout pulito la prova passava: falliva solo su chi
aveva quel file, cioe' proprio su chi stava lavorando.

Il filtro ora e' la tracciabilita', non il percorso. Non i pattern di
.gitignore, che non dicono nulla su un file gia' tracciato. Se git non
risponde si scandisce tutto come prima: una prova-guardia che ammutolisce
quando perde l'appoggio e' peggio di una che grida al lupo.

Verificato che morda ancora: un fetch verso un CDN in un file tracciato la
fa fallire.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014BBnuACZSCJqXrMYC3LUMU

.env.example

@@ -24,6 +24,20 @@ API_SERVER_URL=http://localhost:8000
 # Each physical tablet/PC deployment must set this unique per-station value.
 # Leave empty only for a single-station all-in-one demo using ST-DEFAULT.
 STATION_CODE=ST-DEFAULT
+# Allow switching station from the URL (?station=CODE) during commissioning, so a
+# single PC can exercise several stations. Leave at 0 in production: on the shop
+# floor the station identity comes from the local install, and measuring against
+# another station's recipes would silently break traceability.
+STATION_SWITCH_ENABLED=0
+
+# --- Vision ---
+# Internal address of the vision worker. Never exposed outside tmflow-net.
+VISION_WORKER_URL=http://vision:8100
+# The pinned VisionSuite commit, stamped into the vision image at build time
+# (Dockerfile.vision ARG). Required: `docker compose build vision` fails
+# immediately without it. Set with:
+#   VISION_ENGINE_VERSION=$(git -C vendor/visionsuite rev-parse HEAD)
+VISION_ENGINE_VERSION=
 
 # --- File Storage ---
 # Resolved against the project root in src/backend/config.py.
@@ -35,6 +49,10 @@ MAX_UPLOAD_SIZE_MB=50
 # --- Setup Page ---
 SETUP_PASSWORD=  # Password per /api/setup, vuoto = disabilitato
 
+# --- AI (OpenRouter) ---
+OPENROUTER_API_KEY=  # API key per parsing schede tecniche, vuoto = disabilitato
+OPENROUTER_MODEL=anthropic/claude-sonnet-4  # Modello AI da utilizzare
+
 # --- Docker ---
 DB_ROOT_PASSWORD=root_password_change_me
 NGINX_PORT=80

.gitignore

@@ -39,6 +39,9 @@ uploads/pdfs/*
 uploads/logos/*
 uploads/reports/*
 uploads/general/
+# Statistics files emitted when a production is closed: shop-floor evidence, not
+# source. They live on the mounted volume in production.
+uploads/statistics/
 !uploads/images/.gitkeep
 !uploads/pdfs/.gitkeep
 !uploads/logos/.gitkeep
@@ -74,3 +77,6 @@ nul
 # Competitor analysis (local only)
 Concorrente/
 docker-compose.override.yml
+
+# Subagent-driven development scratch: ledger, briefs, review packages
+.superpowers/

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "vendor/visionsuite"]
+	path = vendor/visionsuite
+	url = ssh://git@git.tielogic.xyz:222/Adriano/visionsuite.git

CLAUDE.md

@@ -123,7 +123,7 @@ uv lock                                     # rigenera uv.lock
 ### i18n (Traduzioni)
 ```bash
 # Estrai stringhe
-cd src/frontend/flask_app && uv run pybabel extract -F babel.cfg -k _ -o translations/messages.pot .
+cd src/frontend/flask_app && uv run pybabel extract -F translations/babel.cfg -k _ -o translations/messages.pot .
 
 # Aggiorna catalogo
 cd src/frontend/flask_app && uv run pybabel update -i translations/messages.pot -d translations

Dockerfile.frontend

@@ -20,8 +20,14 @@ RUN uv sync --frozen --no-dev --extra client
 COPY src/frontend/flask_app/ ./flask_app/
 
 # Build TailwindCSS (one-shot; no watcher in production image).
+#
+# Versione esatta, non `tailwindcss@3`. Con il vincolo aperto ogni ricostruzione
+# dell'immagine poteva prendere una minor diversa e produrre un CSS diverso senza
+# che nessuno l'avesse validato: per un sistema che produce evidenze per audit
+# ISO 9001 / IATF 16949 è lo stesso problema che il punto 12 risolve per le
+# librerie del browser. Aggiornarla è una decisione, non un effetto collaterale.
 WORKDIR /app/flask_app
-RUN npm install tailwindcss@3 && \
+RUN npm install tailwindcss@3.4.19 && \
     npx tailwindcss -i static/css/input.css -o static/css/tailwind.css --minify
 
 # Compile Flask-Babel translation catalogs.

Dockerfile.vision

@@ -0,0 +1,41 @@
+FROM python:3.13-slim
+
+# opencv-python (pulled in transitively by vs-core/vs-task/vs-measure/vs-pm2d)
+# needs libGL and glib at runtime even though it is never imported by name here.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        libgl1 libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+
+WORKDIR /app
+COPY pyproject.toml uv.lock ./
+COPY vendor/ ./vendor/
+RUN uv sync --extra vision-worker --frozen --no-dev
+
+COPY src/vision/ ./src/vision/
+COPY src/vision_worker/ ./src/vision_worker/
+
+# Ruling R8: the version is stamped at build time, not discovered at runtime.
+# COPY vendor/ above does not bring vendor/visionsuite/.git along - it is a
+# gitlink file pointing at ../../.git/modules/vendor/visionsuite, which lives
+# outside the build context - so `git rev-parse HEAD` cannot work in here.
+# Pass the submodule's commit in from the host, which does have git:
+#   VISION_ENGINE_VERSION=$(git -C vendor/visionsuite rev-parse HEAD) \
+#     docker compose build vision
+ARG VISION_ENGINE_VERSION
+# I4: a build without the commit must fail now, at build time - not later, at
+# every request including /health, with a container Docker still reports as
+# healthy because no healthcheck existed to say otherwise.
+RUN test -n "$VISION_ENGINE_VERSION" || { \
+        echo "VISION_ENGINE_VERSION is required. Build with:"; \
+        echo '  VISION_ENGINE_VERSION=$(git -C vendor/visionsuite rev-parse HEAD) docker compose build vision'; \
+        exit 1; \
+    }
+ENV VISION_ENGINE_VERSION=${VISION_ENGINE_VERSION}
+
+EXPOSE 8100
+
+# Two workers, not four: each holds the vision stack in memory.
+CMD ["uv", "run", "uvicorn", "src.vision_worker.main:app", \
+     "--host", "0.0.0.0", "--port", "8100", "--workers", "2"]

README.md

@@ -2,8 +2,9 @@
 
 Sistema di gestione task per misurazioni con calibro manuale. Soluzione tablet-first, multi-ruolo, con statistiche SPC (Statistical Process Control) integrate, identità per-stazione e rate limiting per-tablet.
 
-> **Versione corrente:** V2.0.0 (in sviluppo) — branch default `V2.0.0`.
+> **Versione corrente:** V3.0.0 (in sviluppo) — branch `V3.0.0`.
 > Per stato dettagliato e prossimi passi vedi [`docs/architecture/STATO_PROGETTO.md`](docs/architecture/STATO_PROGETTO.md) e [`docs/architecture/ROADMAP.md`](docs/architecture/ROADMAP.md).
+> V3.0.0 lavora sui quindici punti di `TieMeasureFlow_modifiche_2026-07-28.md`: vedi [Novità di V3.0.0](#novità-di-v300).
 
 ---
 
@@ -15,6 +16,7 @@ Caratteristiche principali:
 
 - Recipe versioning condizionale (copy-on-write se esistono misurazioni, update in-place altrimenti)
 - Calcolo pass/fail/warning su quattro limiti di tolleranza (UTL, UWL, LWL, LTL)
+- Tracciamento del tempo di inserimento per ogni misura (`input_duration_ms`): il client misura quanto tempo l'operatore impiega su ciascun subtask, dall'attivazione al salvataggio, utile per analisi di tempo ciclo e produttività
 - SPC: Cp, Cpk, Pp, Ppk, control chart UCL/LCL, istogramma con curva normale — puro stdlib (no numpy)
 - Annotazioni grafiche su disegni tecnici (Fabric.js) con viewer sincronizzato all'esecuzione
 - Identità per-stazione (`STATION_CODE`): ogni tablet vede solo le ricette assegnate alla propria stazione, gestione assegnazioni via GUI admin
@@ -24,6 +26,32 @@ Caratteristiche principali:
 
 ---
 
+## Novità di V3.0.0
+
+V3.0.0 lavora sui quindici punti raccolti in `TieMeasureFlow_modifiche_2026-07-28.md`,
+dal sopralluogo del 28/07. Quanto segue è **fatto e in esercizio**.
+
+| # | Punto | Cosa cambia |
+|---|---|---|
+| 1 · 6 | **Produzione con una vita propria** | `production_runs` e `production_events` sul server: avvio, fermo linea, ripresa e fine produzione esistono come stato, non come pulsanti. La fine produzione emette il file di statistica dell'intera produzione |
+| 2 | **Tipo di task dichiarato** | `task_type` (`note`, `measure`, `drawing`, `xf_compare`, `camera_measure`) al posto di «ha subtask, quindi è una misura»: un task di misura senza quote non si comporta più come una nota |
+| 3 | **Ciclo di misura** | L'intervallo della ricetta vive sul server, sopravvive al cambio pagina, conta anche il ritardo (in rosso, oltre lo zero) e riporta l'operatore alla misura. Cicalino via WebAudio |
+| 5 | **Fuori tolleranza vincolante** | Una quota fuori tolleranza blocca l'avanzamento finché il capoturno non autorizza — o finché quella stessa quota non viene rimisurata dentro i limiti. L'autorizzazione resta scritta sulla misura e nel CSV |
+| 7 | **Stazioni** | Ricette per stazione, cambio stazione da URL per il collaudo, reset per riga |
+| 8 | **Tracciabilità obbligatoria** | `requires_lot` / `requires_serial` per ricetta, verificati sul server: nessuna porta d'ingresso li aggira, barcode compreso |
+| 9 | **Blocco inserimento manuale** | `allow_manual_input`: con la ricetta a solo calibro il tastierino non viene disegnato affatto, e il server rifiuta comunque un valore digitato a mano |
+| 10 | **Sequenza per l'operatore** | La ricetta si apre sul primo task, non su un elenco. La lista scende a secondo livello e dice quali task sono rimasti incompiuti (`2/3`). «Fine ciclo misura» è visibile da subito e spento finché mancano quote |
+| 11 | **Descrizioni formattate** | `**grassetto**` e a capo nelle descrizioni dei task, con marcatura e non HTML: la sanificazione è per costruzione |
+| 12 | **Funzionamento senza internet** | Tutte le librerie e i font nell'installazione, versioni congelate e verificate per impronta, Content-Security-Policy a sola origine locale |
+| 14 | **Layout stabile** | Una cornice sola per tutte le viste, allineata alla navbar; spazio della barra di scorrimento sempre riservato; schermata di misura su `100dvh`. Vedi [`docs/architecture/LAYOUT.md`](docs/architecture/LAYOUT.md) |
+| 15 | **Statistica separata** | Si registra sempre, si consulta a parte: nessun percorso dell'operatore porta alla statistica (verificato da test) |
+
+**Aperto**: il punto 4 (numero di tentativi prima del capoturno) attende le risposte
+del cliente su modalità di autorizzazione e numero di tentativi; il punto 13
+(generazione task dalla scheda tecnica con l'AI) è fuori offerta.
+
+---
+
 ## Architettura
 
 ```
@@ -66,13 +94,22 @@ Il frontend Flask non espone mai le credenziali al browser: ogni chiamata al bac
 | Flask | 3.x | Framework web server-side |
 | gunicorn | 21+ | WSGI server (5 workers × 4 thread gthread) |
 | Jinja2 | incluso in Flask | Template engine |
-| Alpine.js | 3.x (CDN) | Reattività leggera lato client |
-| TailwindCSS | 3.x | CSS utility-first |
-| Plotly.js | CDN | Grafici SPC interattivi |
-| Fabric.js | 5.3.1 (CDN) | Editor annotazioni disegni tecnici |
-| html5-qrcode | CDN | Scanner barcode/QR camera |
+| Alpine.js | 3.15.12 (locale) | Reattività leggera lato client |
+| TailwindCSS | 3.4.19 (build) | CSS utility-first, compilato nell'immagine |
+| Plotly.js | 2.32.0 (locale) | Grafici SPC interattivi |
+| PDF.js | 3.11.174 (locale) | Visualizzazione disegni PDF, worker incluso |
+| Fabric.js | 5.3.1 (locale) | Editor annotazioni disegni tecnici |
+| Inter + JetBrains Mono | woff2 locali | Font UI e numeri |
 | Flask-Babel | ultima stabile | i18n IT/EN |
 
+**Nessuna libreria arriva dalla rete.** Tutte stanno in
+[`src/frontend/flask_app/static/vendor/`](src/frontend/flask_app/static/vendor/VERSIONS.md)
+con versione nel nome e impronta SHA-256 verificata da un test: l'installazione a
+Tràfilo è su rete di produzione isolata, dove una pagina che aspetta un CDN è una
+pagina bianca. Una Content-Security-Policy a sola origine locale (`app.py`) fa sì che
+un tag verso l'esterno aggiunto in futuro venga rifiutato alla scrivania, non in
+reparto.
+
 ### Tooling
 
 | Componente | Ruolo |
@@ -210,7 +247,8 @@ I ruoli sono combinabili (array JSON per utente). Il flag `is_admin` è separato
 |---|---|
 | **Maker** | Crea e gestisce ricette di misurazione: caricamento disegni (PDF/immagini), annotazioni Fabric.js, definizione task/subtask, configurazione tolleranze, versioning copy-on-write |
 | **MeasurementTec** | Esegue misurazioni: scansione barcode per selezione ricetta, interfaccia task-driven, input da calibro USB HID o numpad touch, validazione real-time pass/warning/fail. Vede solo le ricette assegnate alla propria stazione (`STATION_CODE`) |
-| **Metrologist** | Analisi qualità: dashboard SPC (X-bar, R, Cp, Cpk, Pp, Ppk), filtri multi-dimensionali, export report PDF, analisi capability e control chart |
+| **Supervisor** (capoturno) | Autorizza ciò che l'operatore non può decidere da solo: una quota fuori tolleranza che deve restare, il fermo linea, la fine produzione. L'autorizzazione resta scritta sulla misura (`supervisor_id`, `authorised_at`) e finisce nel file di statistica |
+| **Metrologist** | Analisi qualità: dashboard SPC (X-bar, R, Cp, Cpk, Pp, Ppk), filtri multi-dimensionali, export report PDF, analisi capability e control chart. **L'operatore non raggiunge la statistica da nessun percorso** |
 | **Admin** (flag) | Gestione sistema: CRUD utenti, cambio password, attivazione/disattivazione account, **CRUD stazioni e assegnazioni ricette** |
 
 ---
@@ -228,6 +266,7 @@ TieMeasureFlow/
 ├── docker-compose.yml          # Produzione (Traefik, SSL)
 ├── nginx/                      # Config Nginx (dev)
 ├── uploads/                    # Volume Docker file caricati
+├── scripts/                    # Script del progetto (seed ricette di collaudo)
 ├── docs/                       # Documentazione (vedi indice docs/README.md)
 └── src/
     ├── backend/                # FastAPI Backend
@@ -244,11 +283,14 @@ TieMeasureFlow/
     │   │   │                   #   RecipeTask, RecipeSubtask, Measurement,
     │   │   │                   #   AccessLog, SystemSetting,
     │   │   │                   #   RecipeVersionAudit, Station,
-    │   │   │                   #   StationRecipeAssignment
+    │   │   │                   #   StationRecipeAssignment,
+    │   │   │                   #   ProductionRun, ProductionEvent
     │   │   └── api/            # Pydantic v2 schemas request/response
     │   ├── services/           # recipe_service, measurement_service,
     │   │                       #   spc_service, report_service,
-    │   │                       #   auth_service, station_service
+    │   │                       #   auth_service, station_service,
+    │   │                       #   production_service,
+    │   │                       #   production_export_service
     │   ├── migrations/         # Alembic (alembic.ini + env.py)
     │   ├── templates/          # Pagina setup (Jinja2)
     │   └── tests/              # pytest + httpx + aiosqlite
@@ -261,9 +303,13 @@ TieMeasureFlow/
             ├── services/       # APIClient (proxy verso FastAPI con XFF)
             ├── templates/      # Jinja2 + Alpine.js
             ├── static/
-            │   ├── css/        # TailwindCSS compilato
-            │   └── js/         # numpad, caliper, barcode, csv-export,
-            │                   #   spc-charts, annotation-editor/viewer
+            │   ├── css/        # TailwindCSS compilato + themes.css
+            │   │               #   (cornice pagina, temi, scrollbar)
+            │   ├── js/         # numpad, caliper, barcode, csv-export,
+            │   │               #   spc-charts, annotation-editor/viewer,
+            │   │               #   production-clock, rich-text
+            │   └── vendor/     # Alpine, Plotly, PDF.js (+worker), Fabric,
+            │                   #   font woff2 — vedi VERSIONS.md
             ├── translations/   # Flask-Babel .po/.mo IT/EN
             └── tests/
 ```
@@ -340,7 +386,98 @@ uv run pytest --cov src/backend
 uv run pytest src/frontend/flask_app/tests/
 ```
 
-Stato corrente: **171 pass, 4 fail pre-esistenti** (vedi `docs/architecture/STATO_PROGETTO.md`).
+Il modulo di visione (`src/vision/`) è un albero a parte, con un vincolo di
+versione diverso dal resto del monorepo: le librerie di VisionSuite da cui
+dipende richiedono Python 3.13 o superiore, mentre backend e frontend restano
+su Python 3.11. Per questo motivo l'esecuzione di `uv run pytest` con
+l'interprete di default mostra i test di `src/vision/tests/` come *skipped*,
+non come falliti: il pacchetto viene rilevato ma la sua esecuzione reale
+richiede un ambiente Python 3.13 dedicato, creato con il proprio extra `uv`.
+Per eseguirli davvero:
+
+```bash
+# Prepara l'ambiente Python 3.13 con le dipendenze di visione
+uv sync --extra vision --extra dev --python 3.13
+
+# Esegue i test del runner di visione in quell'ambiente
+uv run --python 3.13 --extra vision --extra dev pytest src/vision/tests
+```
+
+`src/vision_worker/tests/` eredita lo stesso vincolo su Python 3.13, perché
+il worker dipende da `src.vision.runner`, ma aggiunge dipendenze proprie -
+FastAPI, Uvicorn, Pillow, python-multipart - dichiarate in un extra
+separato, `vision-worker`, tenuto distinto da `vision` apposta: un
+consumatore che incorpora solo il runner (come un futuro agente di stazione,
+che non è un servizio web) non deve trascinarsi dietro un server che non gli
+serve. `vision-worker` include comunque `vision`, quindi un solo extra basta
+per avere un worker funzionante:
+
+```bash
+uv run --python 3.13 --extra vision-worker --extra dev pytest src/vision_worker/tests
+```
+
+Per eseguire entrambe le suite insieme, nominando esplicitamente entrambi
+gli extra:
+
+```bash
+uv run --python 3.13 --extra vision --extra vision-worker --extra dev pytest src/vision/tests src/vision_worker/tests
+```
+
+### Il worker di visione (`Dockerfile.vision`)
+
+Il worker gira in un container separato dal server FastAPI principale, così
+che l'immagine dell'API resti leggera e un aggiornamento di VisionSuite non
+richieda di riavviare il traffico di produzione. Il container non pubblica
+porte verso l'esterno: lo raggiunge solo il server, all'indirizzo interno
+`http://vision:8100` sulla rete `tmflow-net`.
+
+Ogni misura riporta il commit di VisionSuite che l'ha prodotta
+(`engine_version`), perché una stazione che misura con un motore diverso da
+quello atteso deve poter essere identificata. Il container, però, non ha
+accesso al repository Git del progetto principale — `vendor/visionsuite` è un
+submodule, e la copia `.git` che lo collega al repository ospitante non
+viene inclusa nel contesto di build — quindi la versione non può essere
+scoperta al volo dentro l'immagine. Va invece **stampata al momento del
+build**, leggendo il commit dalla macchina che esegue `docker compose build`,
+dove il repository Git è presente per intero:
+
+```bash
+VISION_ENGINE_VERSION=$(git -C vendor/visionsuite rev-parse HEAD) \
+  docker compose -f docker-compose.dev.yml build vision
+```
+
+La variabile viene passata come build argument (`ARG VISION_ENGINE_VERSION`
+in `Dockerfile.vision`) e fissata nell'immagine come variabile d'ambiente, in
+modo che il worker la trovi già pronta a ogni avvio senza doverla ricalcolare.
+Fuori da un container, in un checkout di sviluppo locale, la stessa funzione
+ricade su `git rev-parse` se la variabile non è impostata — comodo per
+lavorare sul runner senza Docker. Ma se un'immagine viene costruita senza
+passare `VISION_ENGINE_VERSION`, quel ripiego non ha nulla su cui appoggiarsi:
+il `.git` del submodule non arriva nel contesto di build, e il worker
+risponde con un errore esplicito su `/health` invece di indovinare una
+versione o restituire `"unknown"`. La build va quindi sempre lanciata con la
+variabile impostata, sia in sviluppo sia in produzione, con lo stesso comando
+mostrato sopra (sostituendo `docker-compose.dev.yml` con `docker-compose.yml`
+in produzione).
+
+Stato corrente su `V3.0.0`, con `uv run pytest` su Python 3.11 e senza
+l'extra `vision` (un conteggio senza il suo ambiente non dice nulla, regola
+che vale anche qui): **390 pass, 1 fail, 4 skip** — 243 in `src/backend/tests/`,
+147 pass + 2 skip in `src/frontend/flask_app/tests/`, più 2 skip in
+`src/vision/tests/` e `src/vision_worker/tests/` per il vincolo su Python
+3.13 descritto sopra. Il fallimento è preesistente e indipendente da questo
+lavoro: `test_offline.py::test_no_first_party_script_calls_out`, dovuto a una
+copia locale di Fabric.js non tracciata da git (`static/js/fabric-debug.js`)
+che nei propri commenti cita pagine di documentazione esterne.
+
+Alcuni test frontend non renderizzano niente e leggono i sorgenti, perché guardano
+proprietà che sopravvivono solo se qualcuno le controlla:
+
+| File | Cosa impedisce |
+|---|---|
+| `test_offline.py` | Una libreria caricata dalla rete, un worker PDF.js lasciato sul CDN, una libreria sostituita senza aggiornare l'impronta |
+| `test_layout_shell.py` | Una vista che torna a dichiararsi la propria larghezza |
+| `test_template_js_syntax.py` | Una traduzione con l'apostrofo dentro una stringa JS a virgolette singole, che spegne Alpine su tutta la pagina |
 
 ---
 
@@ -358,6 +495,7 @@ Copia `.env.example` in `.env` e configura:
 | `CLIENT_SECRET_KEY` | Chiave segreta Flask (sessioni, CSRF) |
 | `API_SERVER_URL` | URL del backend visto dal client (es. `http://server:8000`) |
 | `STATION_CODE` | **Per-tablet** — codice stazione (es. `ST-001`). Senza, il client mostra errore configurazione. |
+| `VISION_WORKER_URL` | Indirizzo interno del worker di visione (default: `http://vision:8100`, mai esposto fuori da `tmflow-net`) |
 | `UPLOAD_DIR` | Percorso upload file (default: `uploads`, project root) |
 | `MAX_UPLOAD_SIZE_MB` | Limite dimensione upload (default 50) |
 | `RATE_LIMIT_LOGIN` | Login req/min/IP (default 5) |
@@ -376,6 +514,7 @@ Indice completo: [`docs/README.md`](docs/README.md).
 
 | Documento | Contenuto |
 |---|---|
+| [`TieMeasureFlow_modifiche_2026-07-28.md`](TieMeasureFlow_modifiche_2026-07-28.md) | I quindici punti del 28/07: cosa deve cambiare, dove intervenire, decisioni in attesa del cliente (D-1…D-9) |
 | [`docs/architecture/STATO_PROGETTO.md`](docs/architecture/STATO_PROGETTO.md) | Snapshot V2.0.0: cosa funziona oggi, test status, decisioni architetturali |
 | [`docs/architecture/ROADMAP.md`](docs/architecture/ROADMAP.md) | Cosa resta da fare (Fasi 2-7 rev04, decisioni cliente aperte, stime) |
 
@@ -387,6 +526,8 @@ Indice completo: [`docs/README.md`](docs/README.md).
 | [`docs/DEPLOYMENT.md`](docs/DEPLOYMENT.md) | Guida deployment VPS: Docker, Traefik, SSL, DNS, firewall |
 | [`docs/USER_GUIDE.md`](docs/USER_GUIDE.md) | Manuale utente per ruolo (Maker, MeasurementTec, Metrologist) |
 | [`docs/I18N_SETUP.md`](docs/I18N_SETUP.md) | Setup e workflow traduzioni (Flask-Babel + Alpine.js) |
+| [`docs/architecture/LAYOUT.md`](docs/architecture/LAYOUT.md) | La cornice delle pagine: perché il layout si spostava e la regola che lo tiene fermo |
+| [`src/frontend/flask_app/static/vendor/VERSIONS.md`](src/frontend/flask_app/static/vendor/VERSIONS.md) | Librerie di terze parti in locale: versioni, impronte, come aggiornarle |
 
 ### Piani dettagliati
 

TieMeasureFlow_modifiche_2026-07-28.md

@@ -0,0 +1,614 @@
+# TieMeasureFlow — modifiche da apportare
+
+**Data:** 28 luglio 2026 (agg. serale: architettura d'installazione, punti 7-8 precisati, punti 14-15 nuovi, timer bidirezionale nel punto 3)
+**Riferimento codice:** branch `V2.0.0`, commit `2a56632`
+**Destinatari:** Parte 1 leggibile da cliente e rivenditore · Parte 2 per chi mette mano al codice
+
+---
+
+## Come leggere questo documento
+
+La **Parte 1** dice *cosa deve fare* il sistema e *perché*: si può mandare a Ricerca e
+Misure e a Tràfilo. La **Parte 2** dice *dove si interviene*, con file e criteri di
+completamento. La **Parte 3** elenca le domande la cui risposta non dipende da noi.
+
+Ogni punto porta lo **stato verificato sul codice**, non riferito: `già fatto`,
+`parziale`, `da fare`. Serve a evitare l'equivoco emerso in questi giorni — diverse
+funzioni risultano "già implementate" perché **l'interfaccia c'è**, ma dietro non
+succede nulla.
+
+**Fonti:** specifica *Schema sviluppo SW TieFlow rev. 04-2026*; *MODIFICHE TIEMEASURE
+v260519*; call del 28/07/2026; rettifica di Marco Menoncin del 28/07; lettura diretta
+del codice.
+
+---
+
+## Il punto che regge tutti gli altri
+
+**Lo stato della produzione oggi vive dentro una pagina del browser.**
+
+Il timer, il conteggio dei cicli, il flag "produzione avviata" e le misure in corso
+sono variabili del componente Alpine di `task_execute.html`. La navigazione fra un
+task e l'altro è un **ricaricamento completo di pagina** (`window.location.href`).
+
+Conseguenza diretta: **cambiando task si perde tutto.** Il timer smette di esistere,
+il conteggio dei cicli riparte, "produzione avviata" si dimentica.
+
+Questo spiega, senza bisogno di altre ipotesi:
+
+- perché il loop di misura non regge quando la ricetta ha altri task dopo la misura;
+- perché con più task di misura non si sa a quale tornare allo scadere del timer;
+- perché *fermo linea* e *fine produzione* non hanno nulla da fermare o da chiudere;
+- perché non esiste storico di cosa è successo durante una produzione.
+
+**Va risolto per primo**: i punti 2, 3, 6 e 8 dipendono da questa scelta, e affrontarli
+prima significa rifarli dopo.
+
+---
+
+## Architettura d'installazione (decisa il 28/07)
+
+Chiude la domanda D-3 lato Tielogic. L'installazione a Tràfilo è così composta:
+
+```
+SERVER (uno, del cliente)          STAZIONE (una per PC, installata in locale)
+┌─────────────────────────┐        ┌──────────────────────────────────────┐
+│ MySQL                   │◄──API──│ App di stazione:                     │
+│ Backend dati (API)      │        │  · interfaccia operatore (frontend)  │
+│ File dei disegni        │        │  · agente hardware: calibro USB,     │
+│ Migrazioni dello schema │        │    driver camera, colonnina/cicalino │
+│ Adattatore GAIA         │        │  · elaborazione visione (futura:     │
+└─────────────────────────┘        │    XF compare, pattern matching)     │
+                                   └──────────────────────────────────────┘
+```
+
+- **Sul server vive tutto ciò che è stato**: il database, i file dei disegni, le
+  migrazioni dello schema e l'unico punto di contatto col gestionale GAIA.
+  Aggiornare la logica dati = un deploy, in un posto solo.
+- **Sulla stazione vive tutto ciò che tocca l'hardware**: calibro, telecamera,
+  segnalazione luminosa/acustica, e in prospettiva l'elaborazione visione — che
+  così non carica il server (rilevante per il dimensionamento, D-4).
+  L'app di stazione è **senza stato**: non tocca il database direttamente, parla
+  col backend dati via API e gli manda risultati e immagini.
+- **L'identità della stazione è data dall'installazione** (configurata sul PC),
+  non più dal container sul server: il vincolo "una stazione = un container"
+  decade.
+- La **licenza per postazione** ha l'aggancio naturale nell'app di stazione
+  installata.
+
+**Correzione a quanto detto in call:** i disegni **non stanno nel database** —
+stanno su filesystem (`uploads/`), serviti dal backend. Con questa architettura
+restano sul server, in un posto solo: nessuna cartella da sincronizzare fra i PC.
+
+Conseguenza per lo sviluppo: lo stato della produzione (punto 1) va nel database
+**per obbligo architetturale**, non per scelta — con 20 app di stazione non esiste
+altro posto dove possa vivere.
+
+---
+
+# Parte 1 — Cosa deve fare il sistema
+
+## Quadro: cosa già funziona
+
+Da mettere a verbale, perché in parte era dato per mancante:
+
+| Funzione | Stato |
+|---|---|
+| Stazioni e assegnazione ricette alle stazioni | ✅ presente |
+| Ruolo **capoturno** con autorizzazione a login | ✅ presente |
+| Timer d'intervallo misura con **conto alla rovescia a video** e cicalino | ✅ presente |
+| Logout automatico per inattività, configurabile | ✅ presente |
+| Tolleranze con soglie di attenzione (UTL/UWL/LWL/LTL) ed esito per quota | ✅ presente |
+| Versionamento delle ricette con storico delle modifiche | ✅ presente |
+| Registrazione del **tempo di inserimento** di ogni misura | ✅ presente (28/07) |
+| Pulsanti *Fermo linea*, *Fine produzione*, *Avvio produzione* | ⚠️ solo interfaccia |
+
+L'ultima riga è la più importante: i tre pulsanti esistono, chiedono correttamente
+l'autorizzazione del capoturno, **e poi non fanno nulla**.
+
+---
+
+## 1. Memoria della produzione in corso `da fare` · prerequisito
+
+**Oggi:** lo stato di una produzione esiste solo finché l'operatore resta sulla stessa
+schermata.
+
+**Deve:** la produzione diventa un'entità con una vita propria — si apre quando parte,
+registra ciò che accade (avvio, cicli di misura, fermo linea, ripresa, chiusura) e si
+chiude quando il capoturno la chiude. Se l'operatore cambia task, esce e rientra, o il
+tablet si riavvia, la produzione è ancora lì con il suo timer.
+
+**Perché conta:** senza questo, *fermo linea* e *fine produzione* non possono
+funzionare, il loop di misura non è realizzabile e non esiste il dato storico su cui
+poggiano sia la statistica sia l'integrazione col gestionale.
+
+---
+
+## 2. Tipo di task esplicito `da fare`
+
+**Oggi:** il sistema distingue un task di misura da un task documentale **deducendolo**:
+se ha delle quote è una misura, altrimenti è una nota. Non esiste un tipo dichiarato.
+
+**Deve:** ogni task nasce con un tipo scelto da chi crea la ricetta — **nota**,
+**misura**, **disegno** — predisposto per i tipi futuri (confronto profilo/DXF, misura
+con camera).
+
+**Perché conta:** è la richiesta n. 1 di Menoncin, ed è ciò a cui si agganciano timer e
+loop. Con la deduzione attuale un task di misura a cui non sono ancora state inserite le
+quote viene trattato come una nota: il sistema si comporta in modo diverso a seconda di
+quanto è completa la ricetta.
+
+---
+
+## 3. Loop di misura e ripetizione `parziale`
+
+**Oggi:** allo scadere del timer suona il cicalino e il ciclo si riazzera — ma solo se
+l'operatore è rimasto su quella schermata.
+
+**Deve:**
+- l'operatore esegue i task in sequenza fino al primo task di misura;
+- da lì **resta in loop sulla misura** finché la produzione non viene chiusa;
+- il conto alla rovescia di **quanto manca alla prossima misura** è sempre visibile;
+- **arrivato a zero, il contatore riparte nell'altro senso**: mostra da quanto tempo
+  si è **oltre** l'intervallo di misura, in evidenza — così il ritardo si vede, non
+  si deduce (richiesta dell'operatore, 28/07);
+- allo scadere del timer la misura **si ripropone** ovunque si trovi l'operatore;
+- se la ricetta ha **più task di misura**, il timer riparte alla fine dell'**ultimo**;
+- dev'essere possibile **girare il pezzo e rimisurare** senza chiudere il ciclo.
+
+**Perché conta:** è la seconda richiesta di Menoncin e il comportamento descritto al
+punto 4.10 della specifica del 19/05. La logica c'è già in gran parte: manca che
+sopravviva al cambio di schermata (punto 1).
+
+---
+
+## 4. Limite di tentativi prima del capoturno `da fare`
+
+**Oggi:** l'operatore può ripetere la misura quante volte vuole.
+
+**Deve:** dopo un numero di tentativi definito nella ricetta, per proseguire serve
+l'autorizzazione del capoturno.
+
+**Perché conta:** è nella specifica ed è la contromisura al caso in cui si ripete finché
+non "viene bene". Il meccanismo di autorizzazione esiste già: manca il contatore e la
+soglia.
+
+---
+
+## 5. Avanzamento solo se in tolleranza `da verificare`
+
+**Deve:** si passa alla quota successiva in autonomia **solo se la quota è in
+tolleranza**; per confermare una quota fuori tolleranza serve il capoturno.
+
+**Nota:** il gate del capoturno per il fuori tolleranza è implementato. Va verificato
+sul campo che **blocchi davvero l'avanzamento** e non sia solo una richiesta di conferma.
+Se blocca, il punto si chiude senza sviluppo.
+
+---
+
+## 6. Fermo linea e Fine produzione: dare effetto `parziale`
+
+**Oggi:** entrambi chiedono l'autorizzazione del capoturno, poi non succede niente.
+
+**Deve:**
+- **Fermo linea** — sospende il timer e la produzione; il capoturno può riattivarla;
+- **Fine produzione** — chiude la produzione, ferma il timer definitivamente e **invia
+  i dati di misura dell'intera produzione** al file di statistica.
+
+**Perché conta:** il comportamento verso il gestionale è ancora da definire (Parte 3),
+ma **tutto ciò che sta prima del gestionale si può e si deve fare adesso**: sospendere,
+riprendere, chiudere, registrare. Consegnare i pulsanti funzionanti senza il gestionale
+è possibile; il contrario no.
+
+---
+
+## 7. Gestione delle stazioni `parziale`
+
+**Deve:**
+- la lista stazioni mostra, oltre a codice e postazione, **le ricette collegate**;
+- esiste un **reset della stazione**, con un pulsante **per riga** nella lista:
+  la stazione torna senza ricette associate e si riassegna (precisazione
+  dell'operatore, 28/07);
+- la **stazione corrente si può cambiare al volo**, per poter provare più stazioni da un
+  solo computer senza riconfigurare l'installazione.
+
+**Perché conta:** è la terza richiesta di Menoncin. Il cambio al volo non è un vezzo da
+sviluppatori: senza, la sessione di collaudo con Menoncin richiede tanti PC quante sono
+le stazioni da provare.
+
+---
+
+## 8. Tracciabilità obbligatoria `parziale`
+
+**Oggi:** numero di lotto e numero seriale sono facoltativi e si inseriscono nella lista
+task, cioè dopo aver iniziato.
+
+**Deve:** l'obbligatorietà di lotto e seriale si **decide alla creazione della
+ricetta** (obbligatori sì/no); l'operatore li inserisce **alla selezione della
+ricetta**, e finché mancano il pulsante *Avvia* **non si attiva** (precisazione
+dell'operatore, 28/07).
+
+**Perché conta:** una misura senza lotto non è tracciabile a posteriori, e la
+tracciabilità è metà del valore del sistema in un audit.
+
+---
+
+## 9. Blocco dell'inserimento manuale `da fare`
+
+**Oggi:** il sistema registra **come** è stata inserita una misura (calibro o tastiera),
+ma accetta sempre entrambi.
+
+**Deve:** un'impostazione della ricetta consente o vieta l'inserimento manuale, **con
+divieto come impostazione predefinita**: si misura col calibro.
+
+**Perché conta:** è il punto sollevato in call — senza questo vincolo un valore in
+tolleranza si può digitare. Il dato su *come* è stata inserita c'è già: manca la regola
+che lo impedisce.
+
+---
+
+## 10. Interfaccia operatore: sequenza e conferme `da fare`
+
+Richieste del 19/05, tutte di interfaccia:
+
+- un pulsante che **avvia i task in sequenza**, senza sceglierli a uno a uno;
+- la lista completa dei task retrocessa a **secondo livello**, per tornare a vedere i
+  task precedenti;
+- «inizia misure» rinominato **«visualizza singolo TASK»**;
+- dentro il task, «Riepilogo» sostituito da **«Completato»** per passare al successivo;
+- un task lasciato a metà **resta incompiuto** e si vede;
+- «fine ciclo misura» **cliccabile solo quando tutte le quote hanno un valore**.
+
+---
+
+## 11. Formattazione delle descrizioni `da fare`
+
+**Deve:** le descrizioni dei task accettano andate a capo e grassetto.
+
+**Perché conta:** chi crea le ricette fa **copia e incolla dal PDF della scheda
+tecnica**; oggi il testo arriva appiattito e va risistemato a mano ogni volta.
+
+---
+
+## 12. Funzionamento senza internet `da fare` · bloccante per l'installazione
+
+**Oggi:** l'applicazione **non funziona senza collegamento a internet**. Cinque librerie
+vengono scaricate al volo da servizi esterni ogni volta che si apre una pagina.
+
+**In una rete di produzione isolata — la norma in fabbrica — il risultato è una pagina
+bianca.** Non un degrado: l'interfaccia non parte proprio, e senza le altre non si vedono
+i disegni tecnici né i grafici statistici.
+
+**Deve:** tutte le librerie sono incluse nell'installazione e l'applicazione funziona a
+rete staccata.
+
+**Perché conta ora:** l'installazione a Tràfilo è on-premise e prevista per settembre.
+È poco lavoro, ma va fatto **prima**, non in fabbrica il giorno dell'installazione.
+
+**Beneficio collaterale non ovvio:** oggi una delle librerie è agganciata a una versione
+"qualunque della serie 3" — cioè **l'applicazione cambia da sola** quando gli autori
+pubblicano un aggiornamento, senza che nessuno l'abbia validata. Per un sistema che
+produce evidenze per audit ISO 9001 / IATF 16949 questo è di per sé un problema.
+Includendo le librerie le versioni si congelano: da difetto diventa argomento di vendita.
+
+---
+
+## 13. Generazione dei task dalla scheda tecnica con l'AI `fuori offerta`
+
+**Richiesta:** leggere il PDF della scheda tecnica e **creare un task per blocco**,
+invece del copia-incolla manuale.
+
+**Storia:** posta il **19/05/2026** nel documento delle modifiche, rimasta senza
+risposta; **rilanciata da Tràfilo il 28/07** come elaborazione massiva iniziale delle
+schede, «senza installare agenti nel sistema».
+
+**Stato:** non è in nessuna offerta. Prima di quotare servono tre informazioni: quante
+sono le schede, se il formato è standard, e se l'elaborazione è una-tantum in fase di
+avviamento o una funzione permanente del prodotto. Sono domande da fare, non da
+supporre — vedi Parte 3.
+
+---
+
+## 14. Stabilità del layout `da fare` · da circoscrivere
+
+**Oggi:** le dimensioni delle viste **cambiano a seconda del menu**: passando da una
+schermata all'altra la finestra non mantiene proporzioni stabili.
+
+**Deve:** il layout resta stabile nel passaggio fra le viste.
+
+**Nota:** segnalazione dell'operatore del 28/07, non ancora circoscritta sul codice —
+prima di intervenire va riprodotta e va stilato l'elenco delle viste interessate.
+
+---
+
+## 15. Statistica: si registra sempre, si consulta a parte `già fatto` · da confermare sul campo
+
+**Richiesta (28/07):** a fine misura l'operatore **non va portato nella pagina della
+statistica**. I dati **entrano comunque in statistica**: cambia solo chi la consulta —
+serve l'**utente con il ruolo adeguato**, che apre la pagina dedicata.
+
+**Verificato sul codice:** è già così. Tutte le pagine di statistica richiedono il
+ruolo **Metrologo** (`role_required("Metrologist")` su ogni route), e a fine ciclo
+l'operatore viene portato al **riepilogo**, non alla statistica.
+
+**Resta da fare:** niente sviluppo; il requisito entra come **criterio di collaudo**
+(l'operatore non deve poter raggiungere la statistica da nessun percorso) e va tenuto
+fermo quando il punto 10 ridisegna la navigazione a fine task.
+
+---
+
+# Parte 2 — Dove si interviene
+
+Riferimenti al branch `V2.0.0`, commit `2a56632`.
+
+## Ordine consigliato
+
+```
+12 (offline)  ──────────────►  indipendente, si può fare subito
+                               bloccante per l'installazione
+
+ 1 (stato produzione)  ──┬───►   3 (loop misura)
+                         ├───►   6 (fermo linea / fine produzione)
+                         └───►   4 (limite tentativi)
+
+ 2 (tipo task)  ─────────────►   3, 10
+
+ 7 (stazioni) · 8 (tracciabilità) · 9 (inserimento manuale) · 11 (formattazione)
+   indipendenti fra loro
+
+ 5 (avanzamento in tolleranza)  ──►  prima verificare, forse è già a posto
+
+14 (layout)  ────────────────►  prima riprodurre e circoscrivere le viste
+
+15 (statistica riservata)  ──►  già a posto: solo criterio di collaudo,
+                                da non rompere lavorando sul punto 10
+```
+
+La separazione **backend dati sul server / app di stazione in locale** (vedi
+*Architettura d'installazione*) non è un punto di questa lista: è il contesto in
+cui i punti 1, 3, 6 e 7 vanno progettati. In pratica: API senza stato, stato solo
+nel database, niente dipendenze dal container per l'identità della stazione.
+
+## 1 · Stato della produzione lato server
+
+**Problema tecnico:** `task_execute.html` tiene in variabili Alpine
+(`timerActive`, `timerRemaining`, `_timerInterval`, `cycleCount`, `cycleConfirmed`,
+`productionStarted`, `measurements`) uno stato che deve sopravvivere alla pagina.
+`goToNextTask()` fa `window.location.href` → il componente viene distrutto.
+Lato server la sessione conserva soltanto `lot_number` e `serial_number`.
+
+**Intervento:**
+- nuove tabelle `production_runs` e `production_events` (avvio, ciclo completato, fermo
+  linea, ripresa, chiusura), con `station_id`, `recipe_version_id`, `operator_id`,
+  `supervisor_id` dove serve, timestamp;
+- endpoint REST per aprire, interrogare e aggiornare la produzione corrente della
+  stazione;
+- il frontend legge lo stato all'apertura di ogni pagina invece di tenerlo in memoria;
+  il conto alla rovescia si **ricalcola dall'orario di scadenza** salvato lato server,
+  non da un contatore locale;
+- gli endpoint vanno progettati **senza stato in memoria di processo**: con l'app di
+  stazione installata su ogni PC (vedi *Architettura d'installazione*) il database è
+  l'unico posto condiviso.
+
+**File:** `src/backend/models/orm/` (nuovo modulo), `src/backend/migrations/versions/`
+(migrazione 005), `src/backend/api/routers/`, `src/frontend/flask_app/blueprints/measure.py`,
+`src/frontend/flask_app/templates/measure/task_execute.html`.
+
+**Fatto quando:** avviata una produzione, si naviga fra i task, si esce e si rientra, e
+il timer prosegue coerente; il riavvio del browser non azzera nulla.
+
+---
+
+## 2 · Tipo di task
+
+**Problema tecnico:** `RecipeTask` non ha campo tipo; il frontend decide con
+`subtasks.length > 0` (`task_execute.html`, `task_list.html`).
+
+**Intervento:** campo `type` su `recipe_tasks` con valori `note | measure | drawing`
+(predisposto per `xf_compare`, `camera_measure`), migrazione con valorizzazione dei dati
+esistenti secondo la regola attuale, selezione del tipo nell'editor ricetta,
+sostituzione dei controlli su `subtasks.length` con il tipo.
+
+**File:** `src/backend/models/orm/task.py`, nuova migrazione,
+`src/backend/models/api/`, `src/frontend/flask_app/templates/maker/task_editor.html`,
+`templates/measure/task_execute.html`, `templates/measure/task_list.html`.
+
+**Fatto quando:** una ricetta con un task di misura ancora privo di quote si comporta da
+task di misura.
+
+---
+
+## 3 · Loop di misura
+
+**Base già presente:** `confirmCycle()`, `startMeasurementTimer()`, `onTimerExpired()`,
+`timerDisplay`, `playBuzzer()` in `task_execute.html` (righe ~955-1050). La logica è
+corretta; il problema è la persistenza (punto 1) e il fatto che allo scadere non si può
+riportare l'operatore sul task giusto.
+
+**Intervento:** spostare la scadenza sul server; alla scadenza, **redirezione al task di
+misura** della produzione corrente; con più task di misura far ripartire il timer al
+completamento dell'**ultimo**; aggiungere «rimisura» che riapre il ciclo senza chiuderlo.
+
+**Timer bidirezionale:** oggi il contatore **solo decrementa e si ferma a zero**
+(`timerRemaining--`, poi `onTimerExpired()`); `timerDisplay` formatta minuti:secondi
+dal residuo. Va esteso: sotto zero il valore continua **in negativo** e la
+visualizzazione passa a "oltre da m:s", con stile in evidenza. Calcolando dal
+timestamp di scadenza lato server (come sopra), il ritardo è coerente su qualunque
+schermata e sopravvive al ricaricamento.
+
+**Fatto quando:** con una ricetta a due task di misura e task documentali in coda, allo
+scadere del timer l'operatore viene riportato alla misura da qualunque schermata.
+
+---
+
+## 4 · Limite di tentativi
+
+**Problema tecnico:** nessun `max_retries` nel codice.
+
+**Intervento:** campo sulla ricetta (accanto a `measurement_interval_minutes`, che segue
+lo stesso schema), contatore per quota nella produzione corrente, superata la soglia
+riuso del modale capoturno esistente con motivo `max_retries`.
+
+**File:** `src/backend/models/orm/recipe.py`, migrazione,
+`templates/maker/recipe_editor.html` (accanto al timer), `task_execute.html`.
+
+---
+
+## 5 · Avanzamento in tolleranza — prima verificare
+
+Il modale capoturno gestisce già `out_of_tolerance` (`task_execute.html`,
+`openSupervisorModal`, `validateSupervisor`, endpoint
+`measure.validate_supervisor`). **Prima di sviluppare, provare**: se il rifiuto blocca
+l'avanzamento, il punto è chiuso. Se è solo una conferma, va reso vincolante.
+
+---
+
+## 6 · Fermo linea e Fine produzione
+
+**Problema tecnico:** in `task_execute.html` i due pulsanti aprono il modale e poi
+ricadono su un commento: `fermo_linea and fine_produzione are handled by GAIA
+integration (future)`. `startProduction()` è un `TODO` con la chiamata commentata.
+
+**Intervento (senza gestionale):** scrivere gli eventi su `production_events`,
+sospendere e riprendere il timer, chiudere la produzione, ed **emettere il file di
+statistica** con tutte le misure della produzione. L'invio al gestionale resta un
+adattatore separato da riempire quando il protocollo sarà definito (Parte 3): va
+previsto il punto d'innesto, non l'implementazione.
+
+**Fatto quando:** *fermo linea* congela il timer e solo il capoturno lo riattiva; *fine
+produzione* chiude e produce il file.
+
+---
+
+## 7 · Stazioni
+
+**Base presente:** `Station` e `StationRecipeAssignment` (`models/orm/station.py`,
+migrazione 002).
+
+**Intervento:** ricette collegate nella lista stazioni; pulsante di reset **per riga**
+che rimuove le assegnazioni; selezione della stazione corrente da parametro URL con
+ricaduta sulla variabile d'ambiente, per il collaudo da una sola macchina.
+
+**Nota (28/07, chiude il dubbio che era in D-3):** in produzione l'identità della
+stazione è data dall'**installazione locale** dell'app di stazione, non dal container.
+Il cambio al volo via URL resta come strumento di **collaudo**.
+
+---
+
+## 8 · Tracciabilità
+
+**Base presente:** `lot_number` e `serial_number` su `Measurement`, salvataggio in
+sessione (`measure.save_traceability`), inserimento in `task_list.html`.
+
+**Intervento:** flag sulla ricetta (`richiede lotto`, `richiede seriale`); spostare
+l'inserimento sull'avvio produzione; *Avvia* disabilitato finché mancano.
+
+---
+
+## 9 · Inserimento manuale
+
+**Base presente:** `Measurement.input_method` (`manual | usb_caliper`), valorizzato dal
+frontend.
+
+**Intervento:** flag sulla ricetta `consente inserimento manuale`, **predefinito falso**;
+validazione **lato server** — un flag solo nel frontend non protegge da nulla; tastierino
+nascosto quando vietato.
+
+---
+
+## 10 · Interfaccia operatore
+
+Interventi su `templates/measure/task_list.html` e `task_execute.html`: pulsante di
+avvio in sequenza, retrocessione della lista a secondo livello, rinomina dei due
+pulsanti, stato «incompiuto» sui task abbandonati, «fine ciclo misura» abilitato solo a
+quote complete.
+
+---
+
+## 11 · Formattazione descrizioni
+
+`RecipeTask.description` è già `Text`. Serve un editor minimale (grassetto e a capo) e
+la resa corrispondente in esecuzione, con **sanificazione dell'HTML** in ingresso.
+
+**File:** `templates/maker/task_editor.html`, `templates/measure/task_execute.html`.
+
+---
+
+## 12 · Funzionamento senza internet
+
+**Verificato sul codice.** Cinque librerie esterne in sei template:
+
+| Libreria | Dove | Senza rete si perde |
+|---|---|---|
+| Alpine.js `3.x.x` | `base.html` | **tutta l'interfaccia** |
+| Plotly `2.32.0` | `statistics/dashboard.html` | carte di controllo e istogrammi |
+| PDF.js `3.11.174` | `task_execute`, `recipe_preview`, `task_drawing`, `task_editor` | visualizzazione dei disegni |
+| Fabric.js `5.3.1` | `maker/task_drawing.html` | editor delle annotazioni |
+| Google Fonts | `base.html` | estetica e attese al caricamento |
+
+**Intervento:** scaricare le librerie in `src/frontend/flask_app/static/vendor/` (la
+cartella **esiste già ed è vuota**), ripuntare i tag, chiudere la policy di sicurezza in
+`security_headers.py` da elenco-di-CDN a solo-origine-locale.
+
+⚠️ **Trappola da non mancare:** PDF.js ha una **seconda** referenza al CDN,
+`pdfjsLib.GlobalWorkerOptions.workerSrc`, presente in **quattro file**. Ripuntando solo
+lo script principale la libreria si carica in locale **e il worker continua a cercare
+internet**: sembra funzionare finché non si apre un disegno.
+
+**Fatto quando:** con la rete staccata si percorre login → scelta ricetta → esecuzione
+task → annotazione disegno → statistiche → report, senza errori in console.
+
+---
+
+# Parte 3 — Decisioni che non dipendono da noi
+
+Da chiudere **prima** che i punti collegati entrino in sviluppo. Vanno girate a Tràfilo
+tramite Menoncin.
+
+| # | Domanda | Blocca | Chi risponde |
+|---|---|---|---|
+| **D-1** | **Protocollo del gestionale GAIA**: come si scambiano i dati — servizi web, database condiviso, file? | avvio produzione, fermo linea, fine produzione verso il gestionale; lettura dei codici articolo per stazione | IT Tràfilo + fornitore GAIA |
+| **D-2** | **Rete e credenziali** per raggiungere GAIA dal server dove sarà installato | come sopra | IT Tràfilo |
+| ~~**D-3**~~ | ~~Una applicazione per stazione o una sola per tutte?~~ **Decisa il 28/07** lato Tielogic: backend dati sul server, app di stazione installata su ogni PC — vedi *Architettura d'installazione*. Resta la validazione con l'IT di Tràfilo (macchina e rete) | — | chiusa (noi); validazione in D-4 |
+| **D-4** | **Server**: quale macchina, quanto spazio disco. Sul server stanno database **e file dei disegni** (`uploads/`); l'elaborazione visione **non** è sul server (sta sull'app di stazione), quindi pesa lo spazio disco, non la potenza di calcolo | installazione | IT Tràfilo |
+| **D-5** | **Cicalino**: basta il suono del browser o serve una segnalazione luminosa? La specifica chiede luce **e** suono accesi per tutta la misura, visibili da lontano. Con l'architettura del 28/07 l'app di stazione **può pilotare una colonnina**: la domanda diventa *quale hardware* | punto 3 | Tràfilo |
+| **D-6** | **Autorizzazione capoturno**: username e password come oggi, o PIN rapido / badge? Venti volte al giorno la password è un attrito | punti 4, 5, 6 | Tràfilo |
+| **D-7** | **Numero di tentativi** consentiti prima del capoturno: quanti, e uguali per tutte le ricette? | punto 4 | Tràfilo |
+| **D-8** | **Schede tecniche**: quante sono, il formato è standard, e serve una conversione una-tantum o una funzione permanente? | punto 13 e la sua quotazione | Tràfilo |
+| **D-9** | Modificare i parametri di una ricetta (timer, tentativi) **crea una nuova versione** o no? Sono parametri di esercizio, non di prodotto | punti 4, 3 | noi, con conferma cliente |
+
+**Nota su D-1 e D-2:** finché non hanno risposta, del gestionale si può solo predisporre
+il punto d'innesto. Tutto il resto del punto 6 — sospendere, riprendere, chiudere,
+registrare, produrre il file — **si fa comunque e va fatto adesso**.
+
+---
+
+## Appendice — Come è stato verificato
+
+Ogni «già fatto» e ogni «da fare» viene dalla lettura del codice al commit `2a56632`,
+non dai documenti. In particolare:
+
+- il **tipo di task** è dedotto da `subtasks.length > 0` in `task_execute.html:117` e
+  `task_list.html:169`;
+- il **ruolo capoturno** esiste come `Supervisor`
+  (`api/middleware/api_key.py:71`, `blueprints/measure.py:350`);
+- **fermo linea / fine produzione / avvio produzione** hanno interfaccia e gate ma
+  nessun effetto (`task_execute.html:1017-1021`, `1118`);
+- il **timer** è un `setInterval` locale alla pagina (`task_execute.html:968-995`) e la
+  navigazione fra task è un ricaricamento (`task_execute.html:1054`);
+- **`max_retries`** e **`production_events`** non compaiono in nessun file;
+- le **cinque librerie da CDN** sono ai riferimenti citati al punto 12, e
+  `static/vendor/` contiene solo `.gitkeep`;
+- i **disegni stanno su filesystem**, non nel database: cartella `uploads/` servita
+  dal backend (`api/routers/files.py:201`, `FileResponse`) — verifica del 28/07 sera,
+  corregge quanto detto in call;
+- la **statistica è già riservata**: ogni route di `blueprints/statistics.py` porta
+  `@role_required("Metrologist")`; a fine ciclo il frontend va al riepilogo
+  (`task_execute.html:1126`), non alla statistica;
+- il **timer si ferma a zero**: decremento in `task_execute.html:974-975` con uscita
+  su `onTimerExpired()` — il conteggio del ritardo (punto 3) oggi non esiste;
+- lo stack attuale è a 4 container (`docker-compose.yml`): MySQL 8, backend FastAPI,
+  frontend Flask, nginx — base della sezione *Architettura d'installazione*.
+
+Il punto 14 (layout) è l'unico **non verificato sul codice**: è una segnalazione
+dell'operatore del 28/07, da riprodurre.

docker-compose.dev.yml

@@ -33,6 +33,7 @@ services:
     environment:
       DB_HOST: mysql
       UPLOAD_DIR: uploads
+      VISION_WORKER_URL: http://vision:8100
     volumes:
       - upload_data:/app/uploads
     depends_on:
@@ -41,6 +42,23 @@ services:
     networks:
       - tmflow-net
 
+  vision:
+    build:
+      context: .
+      dockerfile: Dockerfile.vision
+      args:
+        VISION_ENGINE_VERSION: ${VISION_ENGINE_VERSION:-}
+    container_name: tmflow-vision
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8100/health', timeout=5)"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 20s
+    networks:
+      - tmflow-net
+
   client:
     build:
       context: .

docker-compose.yml

@@ -33,6 +33,7 @@ services:
     environment:
       DB_HOST: mysql
       UPLOAD_DIR: uploads
+      VISION_WORKER_URL: http://vision:8100
     volumes:
       - upload_data:/app/uploads
     depends_on:
@@ -50,6 +51,23 @@ services:
       - tmflow-net
       - traefik-net
 
+  vision:
+    build:
+      context: .
+      dockerfile: Dockerfile.vision
+      args:
+        VISION_ENGINE_VERSION: ${VISION_ENGINE_VERSION:-}
+    container_name: tmflow-vision
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8100/health', timeout=5)"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 20s
+    networks:
+      - tmflow-net
+
   client:
     build:
       context: .
@@ -93,4 +111,4 @@ networks:
     driver: bridge
   traefik-net:
     external: true
-    name: root_default
+    name: traefik

docs/COLLAUDO.md

@@ -0,0 +1,191 @@
+# Scaletta di collaudo — V3.0.0
+
+Tredici dei quindici punti del 28/07 sono in esercizio e **nessuno li ha ancora
+percorsi**. Questa è la scaletta per farlo: ordinata in modo che ogni prova prepari
+la successiva, e scritta perché si possa spuntare riga per riga.
+
+Per ogni riga: cosa fare, e **come si vede che è giusto**. Se qualcosa non torna,
+annotare il testo esatto dell'eventuale errore in console — non «dava errore».
+
+Totale stimato: **circa 90 minuti**, di cui 30 sul solo punto 3 (il timer, che si
+misura aspettando).
+
+---
+
+## 0 · Preparazione (5 min)
+
+| | Cosa | Verifica |
+|---|---|---|
+| 0.1 | Aprire il sito e fare **ricarica forzata** (`Ctrl+Shift+R`) | CSS e JS sono in cache per 7 giorni: senza questo si collauda la versione di ieri |
+| 0.2 | Aprire la **console del browser** (`F12`) e lasciarla aperta per tutto il collaudo | Molte prove si leggono lì, non a schermo |
+| 0.3 | Verificare che esistano le ricette `COLLAUDO-A` e `COLLAUDO-B` | Sono in `/measure/select`. Se una sessione precedente le ha sporcate: `docker compose exec server uv run python /tmp/seed_collaudo.py --replace` (cancella anche le misure fatte su di esse) |
+
+**Credenziali del capoturno:** `capoturno` / `Collaudo2026!` — da cambiare prima di
+qualunque uso reale.
+
+**Le ricette:**
+
+- **COLLAUDO-A** — lotto obbligatorio, digitazione ammessa, intervallo **2 minuti**.
+  Cinque task: nota → misura (2 quote) → nota → misura (1 quota) → nota.
+- **COLLAUDO-B** — lotto **e** seriale obbligatori, **solo calibro**, una quota.
+
+**Le quote e i valori da usare** (tutte con nominale al centro):
+
+| Quota | Task | Nominale | Conforme | Attenzione | Fuori tolleranza |
+|---|---|---|---|---|---|
+| 1 · Diametro esterno | A, misura 1 | 10.00 | `10.00` | `10.30` | `12.00` |
+| 2 · Spessore parete | A, misura 1 | 25.00 | `25.00` | `25.30` | `27.00` |
+| 3 · Lunghezza totale | A, misura 2 | 100.00 | `100.00` | `100.70` | `105.00` |
+
+---
+
+## 1 · Layout — punto 14 (10 min) · **da monitor da scrivania**
+
+Va fatto **prima** e **non dal tablet**: in verticale su un tablet quasi tutte le
+larghezze coincidono e il difetto non si vede. Serve uno schermo largo.
+
+| | Cosa | Verifica |
+|---|---|---|
+| 1.1 | Percorrere Ricette → apri una ricetta → Task → Disegno → Anteprima, e tornare indietro | Il blocco di contenuto **non cambia larghezza** fra una schermata e l'altra |
+| 1.2 | Su ognuna, guardare il bordo sinistro del contenuto e quello della barra in alto | Sono **allineati**: il contenuto sta dentro la stessa cornice della navbar |
+| 1.3 | Passare da una pagina lunga (che scorre) a una corta (che non scorre) | Niente **salta di lato** di qualche pixel: lo spazio della barra di scorrimento è riservato sempre |
+| 1.4 | Entrare in un task di misura e uscirne | Stesso controllo: nessuno spostamento orizzontale |
+| 1.5 | Admin → Utenti, Stazioni, Impostazioni; poi il proprio Profilo | Utenti e Stazioni sono larghe, Impostazioni e Profilo strette — ma con **gli stessi margini esterni** |
+| 1.6 | **Dal tablet**, entrare in un task di misura | Il piede con «Fine ciclo misura» e il tastierino è visibile **senza scorrere**, anche quando la barra dell'indirizzo compare e sparisce |
+
+---
+
+## 2 · Niente arriva dalla rete — punto 12 (15 min)
+
+Da fare **adesso**, perché la Content-Security-Policy è appena entrata in vigore: se
+una libreria avesse bisogno di un permesso non concesso, il sintomo compare ora.
+
+> **Attenzione al metodo.** Il sistema è oggi su una VPS pubblica: staccare la rete
+> significherebbe perdere anche il server, quindi non prova niente. La prova
+> equivalente e più diretta è guardare **dove vanno le richieste**. Il giro a rete
+> davvero staccata si fa in reparto, sull'installazione on-premise.
+
+| | Cosa | Verifica |
+|---|---|---|
+| 2.1 | Console → scheda **Rete**, spuntare «Disabilita cache», ricaricare | Ordinando per dominio compare **un solo dominio**: `tieflow.tielogic.xyz`. Nessun `cdn`, `jsdelivr`, `cloudflare`, `googleapis`, `plot.ly` |
+| 2.2 | Ripetere su: login, selezione ricetta, esecuzione task, editor annotazioni (`/maker`), dashboard statistiche | Sempre un dominio solo. La dashboard è quella che carica Plotly, l'editor quella che carica Fabric |
+| 2.3 | Aprire un task che ha un **disegno PDF** allegato | Il disegno si vede. È la prova del *worker* di PDF.js: se cercasse la rete, tutto sembrerebbe a posto fino a questo momento |
+| 2.4 | Con la console aperta su ogni pagina del giro | **Zero** messaggi che iniziano con `Refused to load…` o `…violates the following Content Security Policy directive`. Se ce n'è uno, copiarlo per intero: dice quale permesso manca |
+| 2.5 | Generare un **report PDF** dalle statistiche | Il PDF esce. È generato sul server, non nel browser, ma vale la pena vederlo funzionare nel giro completo |
+
+---
+
+## 3 · Le regole della ricetta — punti 8 e 9 (10 min) · **COLLAUDO-B**
+
+| | Cosa | Verifica |
+|---|---|---|
+| 3.1 | `/measure/select`, **senza** compilare lotto e seriale | Sulla scheda di COLLAUDO-B il pulsante è spento e dice **«Compila prima: lotto, seriale»** |
+| 3.2 | Compilare **solo** il lotto | Il pulsante resta spento e ora chiede solo il **seriale** |
+| 3.3 | Provare ad andare a mano su `/measure/start/<id>` di COLLAUDO-B senza seriale | Rimanda alla selezione con il messaggio «Compila prima:». La regola non si aggira dall'indirizzo |
+| 3.4 | Compilare entrambi e avviare | Entra nel task |
+| 3.5 | Guardare il tastierino | I tasti numerici **non ci sono affatto** — non sono grigi, non ci sono. Restano «cancella» e «conferma» |
+| 3.6 | Con un **calibro USB** collegato, prendere una lettura | Il valore entra e si salva |
+| 3.7 | *(facoltativo, per chi vuole la prova dura)* Provare a salvare un valore digitato aggirando lo schermo | Il server risponde **422**: la regola vive lì, non nella pagina |
+
+---
+
+## 4 · Il primo ciclo e l'avvio produzione — punti 10 e 11 (10 min) · **COLLAUDO-A**
+
+> Da qui in avanti si lavora su COLLAUDO-A. **Il timer non parte subito**: prima si
+> fa un ciclo completo, poi compare «Avvio Produzione». È voluto — il primo pezzo è
+> l'attrezzaggio, non produzione.
+
+| | Cosa | Verifica |
+|---|---|---|
+| 4.1 | Selezione ricetta → COLLAUDO-A → lotto `LOT-TEST-01` → **AVVIA IN SEQUENZA** | Si apre **il primo task** (la nota «Preparazione del pezzo»), **non** l'elenco dei task |
+| 4.2 | Leggere la nota | «a sinistra» è in **grassetto**, le righe vanno a capo e la riga vuota si vede (punto 11) |
+| 4.3 | Guardare i due pulsanti in alto a destra | Sono **«Lista task»** e **«Completato»**. Su una nota «Completato» è attivo |
+| 4.4 | Premere «Completato» | Passa al task 2, «Misura 1 — diametro e spessore» |
+| 4.5 | Guardare il pulsante in basso a destra | **«Mancano 2 quote»**, grigio e non cliccabile |
+| 4.6 | Misurare la quota 1 con `10.00` e confermare | Verde (conforme). Il pulsante in basso ora dice **«Manca 1 quota»** |
+| 4.7 | **Prova chiave del punto 10**: premere «Lista task» adesso, a metà | Nel'elenco il task «Misura 1» porta **«Incompiuto 1/2»** con il bordo ambra. Gli altri non dicono niente |
+| 4.8 | Premere **«Visualizza singolo TASK»** su quel task | Ci si torna dentro, con la quota già presa |
+| 4.9 | Misurare la quota 2 con `25.00` | Il pulsante in basso diventa blu: **«Fine ciclo misura»** |
+| 4.10 | Premerlo | Compare il verde **«Task successivo»** (non un secondo «Completato») |
+| 4.11 | Guardare sotto il piede della pagina | Compare il pulsante verde **«Avvio Produzione»** |
+| 4.12 | Premere «Avvio Produzione» | Parte il conto alla rovescia da **2 minuti**. Da qui in avanti compaiono «Fermo linea» e «Fine produzione» |
+
+---
+
+## 5 · Il fuori tolleranza — punto 5 (15 min)
+
+La parte più importante del collaudo: fino a ieri questo gate era una modale che si
+poteva chiudere.
+
+| | Cosa | Verifica |
+|---|---|---|
+| 5.1 | Tornare su «Misura 1». Cliccare la **quota 1** in alto per rimisurarla, e inserire `12.00` | Si salva ed è **rossa**. La misura sbagliata resta a registro: è un dato, non un errore da nascondere |
+| 5.2 | Provare a passare alla **quota 2** e salvare un valore | **Rifiutato.** Compare l'avviso che serve l'autorizzazione |
+| 5.3 | Guardare «Completato» in alto | È **spento**, e passandoci sopra dice «Serve l'autorizzazione del capoturno» |
+| 5.4 | **Ricaricare la pagina** (`F5`) | Il blocco **è ancora lì**. Ricaricare non è una via d'uscita |
+| 5.5 | **La via d'uscita senza capoturno**: cliccare di nuovo la quota 1 e rimisurarla con `10.00` | Il blocco **si scioglie**. È voluto: il calibro scivola, il pezzo si riposiziona, e chiamare il capoturno per questo sarebbe assurdo |
+| 5.6 | Rifarlo: quota 1 con `12.00`, e stavolta **autorizzare** con `capoturno` / `Collaudo2026!` | Il blocco si scioglie e la misura resta segnata come autorizzata |
+| 5.7 | Provare ad autorizzare con una password sbagliata | Rifiutato, e il blocco **resta** |
+| 5.8 | Provare ad autorizzare con un utente **senza** ruolo Supervisor (es. l'operatore stesso) | Rifiutato: non basta essere collegati, serve il ruolo |
+| 5.9 | Rimettere `10.00` sulla quota 1 e `25.00` sulla 2, chiudere il ciclo | Si prosegue normalmente |
+
+---
+
+## 6 · Il ciclo di misura — punto 3 (30 min, quasi tutti di attesa)
+
+| | Cosa | Verifica |
+|---|---|---|
+| 6.1 | Con il conto alla rovescia in corso, **cambiare task** | Il tempo residuo **prosegue**, non riparte da 2 minuti |
+| 6.2 | **Ricaricare la pagina** (`F5`) | Idem: il tempo è lo stesso. Vive sul server, non nella pagina |
+| 6.3 | Fermarsi su una **nota** e lasciare scadere l'intervallo | A ~5 secondi dalla scadenza compare l'avviso e **suona il cicalino** |
+| 6.4 | Non toccare niente | Alla scadenza si viene portati **al primo task di misura** della ricetta, non a quello dove si era |
+| 6.5 | Fare un ciclo completo: misura 1 (2 quote) → nota → **misura 2** (quota 3, `100.00`) → «Fine ciclo misura» | Chiudendo il ciclo **sull'ultimo task di misura** l'intervallo **riparte da 2 minuti** |
+| 6.6 | Ripetere ma chiudendo il ciclo sulla **prima** misura | L'intervallo **non** riparte: il ciclo si chiude solo sull'ultimo task di misura |
+| 6.7 | **La prova più delicata**: stare **dentro** un task di misura, con una quota già presa e una da prendere, e lasciare scadere l'intervallo | **Non** si viene spostati. Il conteggio diventa **rosso** e prosegue **oltre lo zero**, mostrando il ritardo. Chi sta misurando va lasciato finire |
+| 6.8 | Usare **«Rimisura»** dopo aver completato le quote | Registra una seconda lettura dello stesso pezzo **senza** far ripartire l'intervallo. Girare il pezzo e rimisurare non è un pezzo nuovo |
+
+---
+
+## 7 · Fermo linea e fine produzione — punto 6 (10 min)
+
+| | Cosa | Verifica |
+|---|---|---|
+| 7.1 | Dentro un task di misura, premere **«Fermo linea»** → credenziali capoturno | Il conto alla rovescia **si ferma** |
+| 7.2 | Ricaricare la pagina | È ancora fermo: lo stato è sul server |
+| 7.3 | Premere **«Ripresa»** → credenziali | Riparte |
+| 7.4 | Premere **«Fine produzione»** → credenziali | La produzione si chiude |
+| 7.5 | Sul server: `docker compose exec server ls uploads/statistics/` | C'è un file `production_<id>_<data>.csv` |
+| 7.6 | Aprire il file | Contiene **tutte** le misure della produzione, compresi i `12.00`. Sulla riga autorizzata le colonne `authorised_by` e `authorised_at` sono **valorizzate**: il file mostra il guasto **e** la decisione |
+| 7.7 | Aprire il file in Excel | Le accentate si leggono, il separatore è quello configurato |
+
+---
+
+## 8 · La statistica resta separata — punto 15 (5 min)
+
+| | Cosa | Verifica |
+|---|---|---|
+| 8.1 | Collegarsi con un utente che ha **solo** il ruolo MeasurementTec | In barra **non c'è** la voce Statistiche |
+| 8.2 | Percorrere tutto il flusso operatore guardando ogni schermata | Nessun collegamento porta alla statistica, in nessun punto |
+| 8.3 | Andare a mano su `/statistics/dashboard` | **403**: il ruolo non basta |
+| 8.4 | Verificare che le misure ci siano comunque | Con un utente Metrologist, le misure appena fatte compaiono in dashboard. Si registra sempre, si consulta a parte |
+
+---
+
+## Cosa annotare
+
+Per ogni riga che non torna:
+
+1. **Numero della riga** di questa scaletta.
+2. **Cosa è successo** invece di quello che c'è scritto.
+3. Se c'è un errore in console, il **testo esatto**, per intero.
+4. Se è un problema di aspetto, uno **screenshot** e su **quale schermo** (tablet o
+   monitor, e in che orientamento) — per il punto 14 è l'informazione che manca.
+
+## Cosa questa scaletta non copre
+
+- Il **punto 4** (numero di tentativi prima del capoturno) non è implementato: è
+  fermo sulle risposte del cliente (D-6, D-7, D-9).
+- Il **cicalino** è il suono del browser. Se serva una colonnina luminosa è D-5.
+- Il giro a **rete davvero staccata** si fa in reparto, sull'installazione
+  on-premise: qui si prova l'equivalente (nessuna richiesta esce dal dominio).
+- Il **carico** con venti tablet insieme non è mai stato misurato.

docs/README.md

@@ -6,8 +6,11 @@ Indice della documentazione del progetto.
 
 | Documento | Scopo |
 |---|---|
-| [`architecture/STATO_PROGETTO.md`](architecture/STATO_PROGETTO.md) | Cosa è fatto oggi (V2.0.0). Snapshot del sistema, componenti e capacità. |
-| [`architecture/ROADMAP.md`](architecture/ROADMAP.md) | Cosa resta da fare. Fasi 2-7 della migrazione rev04 verso V1.1.0/M1 demo cliente. |
+| [`../TieMeasureFlow_modifiche_2026-07-28.md`](../TieMeasureFlow_modifiche_2026-07-28.md) | **Il piano di lavoro corrente**: i quindici punti del sopralluogo del 28/07 e le decisioni in attesa del cliente (D-1…D-9). |
+| [`architecture/STATO_PROGETTO.md`](architecture/STATO_PROGETTO.md) | Cosa è fatto oggi (V3.0.0), punto per punto, e cosa non è ancora stato provato sul campo. |
+| [`architecture/ROADMAP.md`](architecture/ROADMAP.md) | Cosa resta, in ordine: collaudo, punto 4, innesto GAIA, installazione a Tràfilo. Include la mappatura con le Fasi rev04 di aprile. |
+| [`architecture/LAYOUT.md`](architecture/LAYOUT.md) | La cornice delle pagine: perché il layout si spostava fra le viste e la regola che lo tiene fermo. |
+| [`../src/frontend/flask_app/static/vendor/VERSIONS.md`](../src/frontend/flask_app/static/vendor/VERSIONS.md) | Librerie di terze parti in locale: versioni, impronte SHA-256, come aggiornarle. |
 
 ## Riferimenti operativi
 
@@ -17,8 +20,12 @@ Indice della documentazione del progetto.
 | [`DEPLOYMENT.md`](DEPLOYMENT.md) | Guida deploy su VPS (Hostinger/Tielogic con Traefik + Let's Encrypt). |
 | [`USER_GUIDE.md`](USER_GUIDE.md) | Manuale utente (operatore, maker, metrologist, admin). |
 | [`I18N_SETUP.md`](I18N_SETUP.md) | Setup e workflow traduzioni (Flask-Babel + Alpine.js). |
+| [`COLLAUDO.md`](COLLAUDO.md) | **Scaletta di collaudo V3.0.0**: cosa provare, in che ordine, con quali valori, e come si vede che è giusto. |
 
-## Piani dettagliati TDD (rev04)
+## Piani dettagliati TDD (rev04) — storici
+
+> Il piano rev04 di aprile è **superato** dal documento del 28/07. La mappatura fra
+> le sue sette Fasi e i quindici punti correnti è in fondo a `architecture/ROADMAP.md`.
 
 | Documento | Scopo |
 |---|---|

docs/architecture/LAYOUT.md

@@ -0,0 +1,98 @@
+# Cornice delle pagine
+
+Nasce dal punto 14 del documento del 28/07: *«le dimensioni delle viste cambiano a
+seconda del menu; passando da una schermata all'altra la finestra non mantiene
+proporzioni stabili»*. La segnalazione era di un operatore e non era ancora stata
+circoscritta sul codice. Questo è l'esito della circoscrizione e la regola che ne
+è uscita.
+
+## Cosa succedeva
+
+Tre meccanismi distinti, tutti verificabili sul codice senza avere il tablet in mano.
+
+**1. Ogni vista si dichiarava la propria larghezza.** Sette valori diversi su
+diciassette template, e nessuno coincideva con quello della navbar, che sta a
+`max-w-7xl` su tutte le pagine. Il contenuto risultava quindi disallineato dalla
+barra sopra di sé, e il disallineamento cambiava da pagina a pagina.
+
+| vista | prima | dopo |
+|---|---|---|
+| `measure/select_recipe.html` | `max-w-7xl` `py-8` | `tmf-page` |
+| `measure/task_list.html` | `max-w-5xl` `py-8` | `tmf-page` |
+| `measure/task_complete.html` | `max-w-7xl` `py-6` | `tmf-page` |
+| `maker/recipe_list.html` | `max-w-6xl` `py-8` | `tmf-page` |
+| `maker/task_editor.html` | `max-w-5xl` `py-8` | `tmf-page` |
+| `maker/task_drawing.html` | `max-w-5xl` `py-8` | `tmf-page` |
+| `maker/recipe_preview.html` | `max-w-5xl` `py-8` | `tmf-page` |
+| `admin/stations.html` | `max-w-7xl` `py-6` | `tmf-page` |
+| `admin/users.html` | `max-w-7xl` `py-6` | `tmf-page` |
+| `statistics/dashboard.html` | `max-w-7xl` `py-6` | `tmf-page` |
+| `admin/settings.html` | `max-w-3xl` `py-6` | `tmf-page tmf-page-narrow` |
+| `auth/profile.html` | `max-w-4xl` `py-8` | `tmf-page tmf-page-narrow` |
+| `maker/recipe_editor.html` | `max-w-4xl` `py-8` | `tmf-page tmf-page-narrow` |
+| `maker/version_history.html` | `max-w-4xl` `py-8` | `tmf-page tmf-page-narrow` |
+| `errors/station_not_configured.html` | `max-w-2xl` `py-8` | `tmf-page tmf-page-narrow` |
+
+Il percorso dell'operatore — quello da cui è arrivata la segnalazione — faceva
+`1280px → 1024px → tutto schermo → 1280px` in quattro passaggi.
+
+**2. Il padding verticale non era lo stesso.** `py-8` e `py-6` mescolati fra viste
+consecutive: la prima card si trovava a un'altezza diversa a ogni cambio di
+schermata.
+
+**3. La barra di scorrimento appariva e spariva.** È una barra classica da 8px
+(`themes.css`, `::-webkit-scrollbar`), quindi occupa spazio nel layout. Una pagina
+lunga la mostrava, una corta no, e la schermata di misura la toglie sempre
+(`body { overflow: hidden }`): a ogni navigazione tutto il contenuto centrato —
+navbar compresa — si spostava di 8px in orizzontale.
+
+## La regola
+
+Due classi in `static/css/themes.css`, nessuna larghezza nei template.
+
+- **`.tmf-page`** — la cornice: `max-width: 80rem`, padding `1rem / 1.5rem / 2rem`
+  ai tre breakpoint, `padding-block: 1.5rem`. Sono esattamente i valori della
+  navbar (`max-w-7xl px-4 sm:px-6 lg:px-8`), così il contenuto ci si allinea.
+- **`.tmf-page-narrow`** — si aggiunge alla prima e stringe a `56rem`. È per i
+  moduli: un campo di testo largo 1280px non si compila meglio.
+- **`html { scrollbar-gutter: stable }`** — lo spazio della barra è riservato
+  sempre, anche dove la pagina non scorre.
+- **`body.h-screen { height: 100dvh }`** — la schermata di misura è alta quanto la
+  finestra vera. `100vh` su Android e iOS è l'altezza che la pagina avrebbe con la
+  barra dell'indirizzo nascosta: su un tablet il piede della schermata — dove
+  stanno «Fine ciclo misura» e il tastierino — finiva sotto il bordo, e ricompariva
+  quando la barra si ritraeva.
+
+Due larghezze in tutto il prodotto al posto di sette, e il criterio è il tipo di
+pagina: liste, tabelle e tele di disegno stanno larghe, i moduli stanno stretti.
+
+## Le due eccezioni
+
+Hanno geometria propria per un motivo, e sono elencate in
+`tests/test_layout_shell.py` perché non sembrino dimenticate.
+
+- **`auth/login.html`** — schermata piena senza navbar: non c'è niente a cui
+  allinearsi.
+- **`measure/task_execute.html`** — la schermata di misura è un pannello a tutta
+  altezza che non deve scorrere mentre un operatore sta misurando
+  (`h-screen overflow-hidden`, footer nascosto). È l'unica vista in cui il cambio
+  di forma è voluto, ed è anche l'unica in cui l'operatore si ferma a lavorare:
+  entrarci e uscirne resta un salto, ma ora è l'unico.
+
+## Cosa resta da verificare sul campo
+
+La segnalazione non è mai stata riprodotta su un dispositivo: quanto sopra viene
+dalla lettura del codice, non da una prova. Quello che è dimostrato è che il
+layout *poteva* muoversi per quattro motivi distinti e che ora non può più per
+nessuno dei quattro.
+
+Se l'operatore vedesse ancora spostamenti, il sospetto successivo è la **tastiera
+virtuale**: quando compare, il browser ridimensiona la finestra, e con `100dvh` la
+schermata di misura si ridimensiona con lei invece di scorrere sotto. È il
+comportamento giusto per il tastierino a schermo, ma va guardato con un calibro
+USB collegato, dove la tastiera di sistema non dovrebbe comparire affatto.
+
+Da chiedere all'operatore che ha aperto la segnalazione: **su quale schermo** l'ha
+vista. Su un tablet in verticale la maggior parte delle larghezze qui sopra
+collassa allo stesso valore e il difetto non si vede; su un monitor da scrivania
+si vede tutto.

docs/architecture/ROADMAP.md

@@ -1,95 +1,146 @@
-# Roadmap TieMeasureFlow — V2.0.0 → V1.1.0 (rev04 / M1 demo cliente)
+# Roadmap TieMeasureFlow — V3.0.0 → installazione a Tràfilo
 
-> Aggiornare ad ogni Fase chiusa.
+> Aggiornata al 2026-07-28. Aggiornare a ogni punto chiuso.
 
-## Riferimenti
+## Dove siamo
 
-- Master plan dettagliato: [`../superpowers/plans/2026-04-17-rev04-master-roadmap.md`](../superpowers/plans/2026-04-17-rev04-master-roadmap.md)
-- Spec sorgente: [`../specs/2026-04-16-schema-sviluppo-rev04.docx`](../specs/2026-04-16-schema-sviluppo-rev04.docx)
-- Stato corrente: [`STATO_PROGETTO.md`](STATO_PROGETTO.md)
+La roadmap rev04 di aprile (Fasi 1-7, milestone M1/M2) è **superata dai fatti**: il
+sopralluogo del 28/07 ha prodotto un elenco di quindici punti concreti che è oggi il
+piano di lavoro. Buona parte delle vecchie Fasi 2 e 4 è dentro quei punti ed è già
+fatta; il resto è confluito o decaduto. La mappatura è in fondo, per non perdere il
+filo con i documenti di aprile.
 
-## Strategia: due milestone
+- Piano corrente: [`../../TieMeasureFlow_modifiche_2026-07-28.md`](../../TieMeasureFlow_modifiche_2026-07-28.md)
+- Stato di dettaglio: [`STATO_PROGETTO.md`](STATO_PROGETTO.md)
+- Piano rev04 (storico): [`../superpowers/plans/2026-04-17-rev04-master-roadmap.md`](../superpowers/plans/2026-04-17-rev04-master-roadmap.md)
 
-| Milestone | Scope | Obiettivo |
+**Scadenza che comanda tutto: l'installazione on-premise a Tràfilo è prevista per
+settembre 2026.**
+
+## I quindici punti
+
+| # | Punto | Stato |
 |---|---|---|
-| **M1 — Demo cliente** | Fasi 1-5 + deploy "demo" | Sistema testabile end-to-end col cliente per raccogliere feedback |
-| **M2 — Produzione** | Fasi 6-7 + correzioni post-feedback + GAIA live | Rollout su tablet/PC reali |
+| 1 | Memoria della produzione in corso | ✅ fatto |
+| 2 | Tipo di task esplicito | ✅ fatto |
+| 3 | Loop di misura e ripetizione | ✅ fatto |
+| 4 | Limite di tentativi prima del capoturno | ⛔ **fermo** su D-6, D-7, D-9 |
+| 5 | Avanzamento solo se in tolleranza | ✅ fatto |
+| 6 | Fermo linea e Fine produzione con effetto | ✅ fatto (innesto GAIA fermo su D-1, D-2) |
+| 7 | Gestione delle stazioni | ✅ fatto |
+| 8 | Tracciabilità obbligatoria | ✅ fatto |
+| 9 | Blocco dell'inserimento manuale | ✅ fatto |
+| 10 | Interfaccia operatore: sequenza e conferme | ✅ fatto |
+| 11 | Formattazione delle descrizioni | ✅ fatto |
+| 12 | Funzionamento senza internet | ✅ fatto |
+| 13 | Generazione task dalla scheda tecnica con l'AI | — fuori offerta (quotazione su D-8) |
+| 14 | Stabilità del layout | ✅ fatto |
+| 15 | Statistica separata dalla registrazione | ✅ fatto |
 
-## Stato Fasi (M1)
+## Cosa resta, in ordine
 
-| Fase | Scope | Stato | Branch / Commit |
+### 1. Collaudo sul campo — *da fare adesso, non serve nessuna risposta*
+
+È il lavoro più urgente rimasto, ed è l'unico che non dipende da nessuno. Tredici
+punti sono in esercizio e **nessuno li ha percorsi su un tablet**.
+
+Le ricette `COLLAUDO-A` e `COLLAUDO-B` sono già sul sistema, con un utente
+`capoturno` (`scripts/seed_collaudo.py`). Da verificare, in quest'ordine:
+
+| Cosa | Come si vede che è giusto |
+|---|---|
+| Ciclo di misura (3) | Il conto alla rovescia sopravvive al cambio task, prosegue in rosso oltre lo zero, riporta alla misura |
+| Fuori tolleranza (5) | Con `12.00` sulla prima quota non si passa alla seconda finché il capoturno non autorizza, o finché non si rimisura dentro |
+| Tracciabilità (8) | Su `COLLAUDO-B` l'avvio non parte finché lotto e seriale non ci sono |
+| Inserimento manuale (9) | Su `COLLAUDO-B` il tastierino numerico non c'è affatto |
+| Sequenza (10) | La ricetta si apre sul primo task; un task lasciato a metà appare «Incompiuto 1/2» nella lista |
+| Senza rete (12) | Staccare la rete e percorrere login → ricetta → task → annotazione → statistiche → report, console aperta |
+| Layout (14) | **Da un monitor da scrivania**, non solo dal tablet: è lì che il difetto si vedeva |
+
+Il giro del punto 12 va fatto **ora**, perché la Content-Security-Policy è appena
+entrata in vigore: se una libreria avesse bisogno di un permesso non concesso, il
+sintomo compare adesso.
+
+### 2. Punto 4 — appena arrivano le risposte
+
+Numero di tentativi prima del capoturno. Serve sapere **D-7** (quanti, e se uguali
+per tutte le ricette), **D-6** (password, PIN o badge) e **D-9** (se cambiare i
+parametri crea una nuova versione). Il resto dell'impianto è pronto: il blocco per
+fuori tolleranza e l'autorizzazione del capoturno esistono già, manca il contatore.
+
+Stima, a decisioni chiuse: **2-3 giorni**.
+
+### 3. Innesto verso GAIA — fermo su D-1 e D-2
+
+Avvio produzione, fermo linea e fine produzione hanno già lo stato e gli eventi; la
+fine produzione emette già il file di statistica. Manca solo il canale verso il
+gestionale, e non si può nemmeno disegnare finché non si sa **come** si parla con
+GAIA (D-1) e **da dove** (D-2).
+
+Stima, a protocollo noto: **1-2 settimane**, molto dipendente dalla risposta.
+
+### 4. Installazione a Tràfilo — settembre
+
+| Cosa | Blocco |
+|---|---|
+| Macchina server, spazio disco | D-4 |
+| Immagini Docker portate in reparto, non repository da compilare sul posto | — |
+| Colonnina luminosa, se serve | D-5 |
+| Validazione rete e macchine con l'IT | D-4 |
+
+Nota: **la costruzione delle immagini richiede rete** (npm, apt, uv); è l'esecuzione
+a non richiederla. In reparto va portata l'immagine già costruita.
+
+## Decisioni in attesa del cliente
+
+Da girare a Tràfilo tramite Menoncin. Le prime due e la D-4 hanno l'orizzonte di
+settembre; D-6 e D-7 bloccano lavoro che sappiamo già fare.
+
+| ID | Decisione | Blocca | Chi risponde |
 |---|---|---|---|
-| **1** | Stazioni + identità per-tablet | ✅ **COMPLETATA** | `V2.0.0` (merge `ea8e468` da `feature/rev04-phase1-stations`) |
-| 2 | Ruolo Capoturno (Supervisor) + override token breve | ⏳ Da iniziare | — |
-| 3 | Editor ricetta a blocchi (preparation + measurement) | ⏳ Da iniziare | — |
-| 4 | Workflow operatore (retry/timer/autologout/avvio produzione) | ⏳ Da iniziare | — |
-| 5 (M1) | `ImportOnlyGaiaClient` + UI import dati cliente reali | ⏳ Da iniziare | — |
-| Deploy M1 | VPS demo (compose + Traefik + LE, no registry) | ⏳ Da iniziare | — |
+| **D-1** | Protocollo del gestionale GAIA: servizi web, database condiviso, file? | innesto GAIA (punto 6) | IT Tràfilo + fornitore GAIA |
+| **D-2** | Rete e credenziali per raggiungere GAIA | come sopra | IT Tràfilo |
+| ~~D-3~~ | ~~Una app per stazione o una sola?~~ **Decisa il 28/07**: dati sul server, app di stazione su ogni PC | — | chiusa; validazione in D-4 |
+| **D-4** | Server: quale macchina, quanto spazio disco (database **e** disegni) | installazione | IT Tràfilo |
+| **D-5** | Cicalino: basta il suono del browser o serve una colonnina luminosa? | punto 3 (completamento) | Tràfilo |
+| **D-6** | Autorizzazione capoturno: password come oggi, o PIN / badge? Venti volte al giorno la password è attrito | punti 4, 5, 6 | Tràfilo |
+| **D-7** | Quanti tentativi prima del capoturno, e uguali per tutte le ricette? | punto 4 | Tràfilo |
+| **D-8** | Schede tecniche: quante, formato standard, conversione una-tantum o funzione permanente? | punto 13 e la sua quotazione | Tràfilo |
+| **D-9** | Cambiare i parametri di una ricetta (timer, tentativi) crea una nuova versione? | punti 3, 4 | noi, con conferma cliente |
 
-## Stato Fasi (M2)
+Una decisione già presa che va **confermata**: la migrazione 009 ha impostato
+`allow_manual_input = 1` su tutte le ricette esistenti, per conservare il
+comportamento in essere invece di irrigidire di colpo ricette già in uso. Va deciso
+ricetta per ricetta quali devono passare a solo calibro.
 
-| Fase | Scope | Stato |
-|---|---|---|
-| 5 (M2) | GAIA reale (protocollo TBD, polling, comandi produzione) | ⏳ Bloccata da decisioni cliente (D-0.1, D-0.2) |
-| 6 | Deploy B industriale (registry privato + Watchtower + STATION_ID per-tablet + CI release) | ⏳ Pianificata |
-| 7 | Hardening, security review, E2E sito pilota, docs aggiornati, i18n delta | ⏳ Pianificata |
-
-## Decisioni aperte (bloccanti per M2 / future fasi)
-
-Da: master plan §0 "Precondizioni e Decisioni Aperte". Da risolvere col cliente prima della Fase 5/6.
-
-| ID | Decisione | Stato | Bloccante per |
-|---|---|---|---|
-| D-0.1 | Protocollo integrazione GAIA (REST / DB shared / OPC-UA / file) | **Aperta** | Fase 5 reale (M2) |
-| D-0.2 | Credenziali e rete GAIA (VPN / firewall / whitelist IP) | **Aperta** | Fase 5 reale (M2) |
-| D-0.3 | Target hardware "tablet" (Windows / Linux industriale / Android) | **Aperta** | Fase 6 (deploy B) |
-| D-0.4 | Cicalino/luce avviso (audio HTML5 / hardware USB / entrambi) | **Rimandata a M2** | Fase 4 finale |
-| D-0.5 | Parametri runtime modificabili vs versione immutabile | **Aperta** (raccomandato B: separare volatili) | Fase 3 |
-| D-0.6 | Auth capoturno durante override (modale / PIN / RFID) | **Aperta** | Fase 2 |
-| D-0.7 | Timeout auto-logout | **Risolta** | — |
-| D-0.8 | Naming ruolo capoturno | **Proposta:** `Supervisor` | Fase 2 |
-| D-0.9 | Tag versione immagine docker | **Proposta:** SemVer + `latest` | Fase 6 |
-| D-0.10 | Registry esposto su Internet o solo VPN | **Proposta:** solo VPN cliente | Fase 6 |
-
-## Tech debt da chiudere
+## Tech debt
 
 | Item | Priorità | Note |
 |---|---|---|
-| 3 test backend pre-esistenti rotti (`test_recipes`, `test_tasks`) | Media | Investigare prima di Fase 3 (toccano recipe + task router). |
-| 1 test client pre-esistente rotto (`test_save_measurement_proxy`) | Bassa | Probabilmente CSRF/payload. Risolvere con Fase 4. |
-| Pagina `task_complete` riepilogo: utente segnala riga vuota in alcuni scenari | Media | Da debuggare (rendering corretto via curl ma utente vede vuoto in browser, possibile interazione con sessione lot/serial). |
-| `.env` rename a convenzione spec (SERVICE_NAME, SERVICE_DOMAIN, API_KEY) | Bassa | Rinviato (impatto deploy). |
-| Header `X-API-Key` rename a `X-Api-Key` | Bassa | Vedere se M2 lo richiede. |
-| Envelope risposta `{success,data,error}` | Bassa | Eventuale API v2 in M2. |
-| `Dockerfile.frontend`: `pybabel compile` via `uv run` non testato in build reale | Alta | Verificare al primo `docker compose build`. |
-| Smoke test in container Docker (non solo locale uvicorn+gunicorn) | Alta | Validare che i Dockerfile riscritti con `uv` buildino e girino correttamente prima di chiudere V2.0.0. |
+| `components/barcode_scanner.html` dipende da `html5-qrcode`, mai caricata | Media | Codice morto: il componente non è incluso da nessuna parte e il lettore che l'operatore usa è un campo di testo. Da rimuovere, o da completare portando la libreria in `static/vendor/` |
+| Pagina `task_complete`: riga vuota segnalata in alcuni scenari | Media | Segnalazione di aprile mai riprodotta. Da verificare durante il collaudo, ora che il flusso è cambiato |
+| `.env` rename a convenzione spec (SERVICE_NAME, SERVICE_DOMAIN, API_KEY) | Bassa | Rinviato: impatta i deploy esistenti |
+| Header `X-API-Key` → `X-Api-Key` | Bassa | Breaking per i deploy esistenti |
+| Envelope risposta `{success,data,error}` | Bassa | Eventuale API v2 |
+| Test di carico a 20 tablet reali | Bassa | La capacità è dimensionata ma mai misurata sotto carico vero |
 
-## Open per scelta utente prima della prossima sessione
+I quattro test rotti tracciati nello snapshot V2.0.0 non risultano più: la suite è a
+360 pass, 0 fail.
 
-1. **Quale fase iniziare adesso?** Opzioni: Fase 2 (Supervisor — sblocca workflow), Fase 3 (block editor — independente), Fase 5 (import GAIA dati reali — sblocca demo).
-2. **Revisione decisioni aperte col cliente** — D-0.1 / D-0.2 / D-0.3 / D-0.6 prima di pianificare Fase 5 e 6.
-3. **Smoke test Docker** della nuova struttura V2.0.0 (`docker compose -f docker-compose.dev.yml up --build`) per validare i Dockerfile riscritti.
-4. **Test di carico** (k6/locust) a 20 VU su `/measure/save-measurement` per validare la scalatura worker (capacità annunciata: 20-30 tablet contemporanei).
+## Mappatura con la roadmap rev04 di aprile
 
-## Stima tempi residui M1 (post-Fase 1)
+Per chi torna sui documenti di aprile e non ritrova le Fasi.
 
-| Task | Stima full-time |
+| Fase rev04 | Che fine ha fatto |
 |---|---|
-| Fase 2 — Supervisor + override | 1 settimana |
-| Fase 3 — Block editor | 1.5 settimane |
-| Fase 4 — Workflow operatore | 2 settimane |
-| Fase 5 (M1) — Import-only GAIA | 1 settimana |
-| Deploy M1 demo | 0.5 settimane |
-| **Totale M1 residuo** | **~6 settimane** |
+| 1 — Stazioni per-tablet | Chiusa in V2.0.0, ampliata dal punto 7 |
+| 2 — Ruolo Capoturno + override | Assorbita dai punti 5 e 6: il ruolo `Supervisor` esiste e l'autorizzazione resta scritta sulla misura. L'override a token breve dipende da D-6 |
+| 3 — Editor ricetta a blocchi | Sostituita dal punto 2 (tipo di task dichiarato), che risolve il problema vero senza riscrivere l'editor |
+| 4 — Workflow operatore | Assorbita dai punti 3, 4, 10: timer e sequenza sono fatti, i tentativi sono il punto 4 |
+| 5 — Import GAIA | Diventata l'innesto del punto 6, ferma su D-1 e D-2 |
+| 6 — Deploy industriale (registry + Watchtower) | Ridimensionata: con una rete isolata l'aggiornamento automatico non ha senso. Restano immagini versionate portate a mano |
+| 7 — Hardening | Parzialmente assorbita: CSP, versioni congelate e impronte sono entrate col punto 12 |
 
-## Stima tempi M2 (dopo feedback)
-
-| Task | Stima |
-|---|---|
-| Aggiustamenti post-feedback | variabile (1-2 sett.) |
-| Fase 5 reale GAIA | 1-2 settimane |
-| Fase 6 deploy B | 1 settimana |
-| Fase 7 hardening | 1-2 settimane |
-| **Totale M2** | **~4-7 settimane** |
-
-**Totale fino a produzione:** ~10-13 settimane full-time da oggi (2026-04-25), assumendo decisioni aperte risolte in tempo utile.
+Le vecchie decisioni `D-0.x` sono confluite nelle `D-x` qui sopra: D-0.1→D-1,
+D-0.2→D-2, D-0.4→D-5, D-0.5→D-9, D-0.6→D-6. D-0.8 (nome del ruolo capoturno) è
+chiusa: si chiama `Supervisor`. D-0.7 (auto-logout) era già risolta.

docs/architecture/STATO_PROGETTO.md

@@ -1,145 +1,192 @@
-# Stato Progetto TieMeasureFlow — V2.0.0
+# Stato Progetto TieMeasureFlow — V3.0.0
 
-> Snapshot al 2026-04-27. Aggiornare ad ogni milestone.
+> Snapshot al 2026-07-28. Aggiornare ad ogni milestone.
+> Lo snapshot V2.0.0 è in [`../archive/STATO_PROGETTO_V2.0.0_2026-04-27.md`](../archive/STATO_PROGETTO_V2.0.0_2026-04-27.md).
 
 ## Versione corrente
 
-**V2.0.0** (in sviluppo, branch `V2.0.0` come default su `git.tielogic.xyz`).
+**V3.0.0** (in sviluppo, branch `V3.0.0`). Versione precedente: `V2.0.0`.
 
-Versione precedente di produzione: `V1.0.7`.
+In esercizio su `tieflow.tielogic.xyz` (Docker Compose + Traefik + Let's Encrypt),
+schema alla migrazione `010_meas_authorisation`.
 
 ## Sintesi esecutiva
 
-Il sistema base (V1.0.7) è completo e collaudato: ricette, task, misurazioni, SPC, report PDF, gestione utenti, dashboard metrologist. La V2.0.0 in corso aggiunge il primo blocco della migrazione **rev04** (stazioni per-tablet) e ristruttura l'intero monorepo secondo lo standard `python-project-spec-design.md` (uv + `src/backend/` + `src/frontend/flask_app/`).
+V2.0.0 aveva chiuso stazioni per-tablet, ristrutturazione monorepo con `uv` e la
+tenuta multi-tablet. **V3.0.0 nasce dal sopralluogo del 28/07 e lavora sui quindici
+punti** raccolti in [`TieMeasureFlow_modifiche_2026-07-28.md`](../../TieMeasureFlow_modifiche_2026-07-28.md).
 
-## Cosa funziona oggi (V2.0.0 — branch corrente)
+Il filo che li tiene insieme: fino a V2.0.0 il sistema *registrava* misure ma non
+*governava* la produzione. Le regole vivevano nell'interfaccia — una modale che si
+poteva chiudere, un timer che moriva al cambio pagina, un tipo di task dedotto dalla
+presenza di quote. V3.0.0 le sposta sul server, dove non si aggirano.
 
-### Funzionalità ereditate da V1.0.7
-- Autenticazione username/password + API key per-utente, ruoli combinabili (Maker, MeasurementTec, Metrologist) + flag `is_admin`.
-- Recipe versioning copy-on-write: una nuova versione si crea solo se la corrente ha già measurements; altrimenti update in-place.
-- Editor ricette (Maker) con annotation editor Fabric.js (~1200 LOC, collaudato su tablet).
-- Workflow operatore tablet: select_recipe → task_list → task_execute → task_complete, con barcode scanner e numpad touch (input USB calibro con burst detection).
-- Calcolo pass/fail con limiti UTL/UWL/LWL/LTL.
-- Dashboard SPC: capability (Cp/Cpk/Pp/Ppk), control chart (UCL/LCL = mean ± 3σ), istogramma con curva normale, calcoli puro stdlib (no numpy).
-- Report PDF (WeasyPrint + Kaleido SVG).
-- Setup page protetta da `SETUP_PASSWORD` per inizializzazione DB e seed.
-- i18n IT/EN (Flask-Babel + Alpine.js JSON).
-- Tema light/dark via `Alpine.store('theme')` + localStorage.
+Tredici punti su quindici sono fatti e in esercizio. Uno è fermo sulle risposte del
+cliente, uno è fuori offerta.
 
-### Aggiunte V2.0.0 (rev04 Fase 1 — Stazioni per-tablet)
-- Tabelle `stations` + `station_recipe_assignments` (Alembic migration `002_add_stations.py`).
-- Modelli ORM: `Station`, `StationRecipeAssignment` con vincolo unique `(station_id, recipe_id)`.
-- Schemas Pydantic: `StationCreate/Update/Response`, `StationRecipeAssignmentCreate/Response`, `RecipeSummary`.
-- Service `station_service` (CRUD + assegnazioni + cascade delete).
-- Router `/api/stations` con CRUD admin + endpoint operatore `GET /api/stations/by-code/{code}/recipes`.
-- Seed automatico `ST-DEFAULT` con tutte le ricette esistenti (idempotente).
-- Variabile env client `STATION_CODE` letta da `Config`, helper `APIClient.get_station_recipes()`.
-- Filtro `select_recipe`: il client mostra solo le ricette assegnate alla propria stazione, errore se `STATION_CODE` non configurato.
-- **GUI admin completa** in `/admin/stations`: tabella con search, modal create/edit, modal gestione assegnazioni ricette, conferma eliminazione, link in navbar (desktop + mobile).
-- 47 nuovi test (32 server + 15 client) tutti pass.
+## I quindici punti
 
-### Aggiunte V2.0.0 (performance + multi-utente)
-- Gunicorn 5 workers × 4 thread (gthread) — capacità ~20 richieste concorrenti Flask, regge 20+ tablet.
-- Uvicorn 4 workers + `--proxy-headers --forwarded-allow-ips='*'`.
-- Rate limit middleware: identificazione IP reale via `X-Forwarded-For` → `X-Real-IP` → `request.client.host`.
-- Rate limit general 100 → 300 req/min/IP (per-tablet ora, non più condiviso).
-- Flask `ProxyFix(x_for=1, x_proto=1, x_host=1)` per IP reale dietro Nginx.
-- `APIClient` propaga `X-Forwarded-For` + `X-Real-IP` (sia JSON che multipart).
-- 12 test aggiuntivi (7 server + 5 client).
+| # | Punto | Stato | Dove vive |
+|---|---|---|---|
+| 1 | Memoria della produzione in corso | ✅ | `production_runs`, `production_service` |
+| 2 | Tipo di task esplicito | ✅ | `RecipeTask.task_type`, migrazione 007 |
+| 3 | Loop di misura e ripetizione | ✅ | `production-clock.js`, `production_service`, migrazione 008 |
+| 4 | Limite di tentativi prima del capoturno | ⛔ **Fermo su D-6, D-7, D-9** | — |
+| 5 | Avanzamento solo se in tolleranza | ✅ | `measurement_service.pending_authorisation`, migrazione 010 |
+| 6 | Fermo linea e Fine produzione con effetto | ✅ | `production_service`, `production_export_service` |
+| 7 | Gestione delle stazioni | ✅ | `station_service`, `/admin/stations` |
+| 8 | Tracciabilità obbligatoria | ✅ | `Recipe.requires_lot/serial`, migrazione 009 |
+| 9 | Blocco dell'inserimento manuale | ✅ | `Recipe.allow_manual_input`, `numpad.js` |
+| 10 | Interfaccia operatore: sequenza e conferme | ✅ | `/measure/start`, `task-progress`, `task_list`, `task_execute` |
+| 11 | Formattazione delle descrizioni | ✅ | filtro `rich_text`, `rich-text.js` |
+| 12 | Funzionamento senza internet | ✅ | `static/vendor/`, CSP in `app.py` |
+| 13 | Generazione task dalla scheda tecnica con l'AI | — **Fuori offerta** | — |
+| 14 | Stabilità del layout | ✅ | `.tmf-page` in `themes.css`, [`LAYOUT.md`](LAYOUT.md) |
+| 15 | Statistica separata dalla registrazione | ✅ | `@role_required("Metrologist")`, test dedicato |
 
-### Aggiunte V2.0.0 (struttura monorepo)
-- `pyproject.toml` unico con extra `server`/`client`/`dev`. Niente più `requirements.txt`.
-- `uv.lock` (77 pacchetti) + `.python-version` (3.11) committati per build riproducibili.
-- Layout `src/backend/` + `src/frontend/flask_app/` (vedi sotto).
-- `Dockerfile` (root) + `Dockerfile.frontend` riscritti con `uv sync --frozen --no-dev --extra server|client`.
-- `docker-compose.{dev,}.yml` con build context `.`.
-- Alembic env.py aggiunge project root a `sys.path`; `script_location = %(here)s` resta valido.
-- `.dockerignore` aggiornato.
+## Cosa è entrato in V3.0.0
 
-### Hardening post-restructure (smoke test 2026-04-26)
+### Produzione come stato, non come pulsante (punti 1, 6)
 
-Sequenza di smoke test in locale (uvicorn + gunicorn + MySQL Docker) ha fatto emergere quattro regressioni che sarebbero rimaste invisibili al test suite:
+- Tabelle `production_runs` e `production_events`; una produzione si apre su una
+  stazione, accumula eventi (`start`, `cycle_completed`, `task_measured`,
+  `remeasure`, `line_stop`, `resume`, `close`) e si chiude.
+- Fermo linea, ripresa e fine produzione richiedono le credenziali del capoturno e
+  **cambiano lo stato**: prima erano modali che si chiudevano.
+- La fine produzione emette il file di statistica dell'intera produzione
+  (`production_export_service`), CSV con i delimitatori configurati, marcando le
+  misure come esportate. Il punto d'innesto verso GAIA è lì accanto, in attesa di
+  D-1 e D-2.
 
-- **`src/backend/config.py`**: `env_file` era cwd-relative (`../../.env`). Rotto fuori da `src/backend/`. Risolto con percorso assoluto `Path(__file__).resolve().parents[2] / ".env"`.
-- **`src/backend/models/orm/__init__.py`**: `Station` e `StationRecipeAssignment` non erano esportati, quindi `Base.metadata.create_all` non creava le tabelle stations. Aggiunti agli import.
-- **`.env.example`**: `UPLOAD_DIR=server/uploads` era residuo della vecchia struttura → file landavano fuori dall'albero di progetto. Aggiornato a `UPLOAD_DIR=uploads`.
-- **Apostrofi italiani in template Alpine** (`l'utente`, `nell'eliminazione`, `nell'assegnazione`): chiudevano prematuramente JS string literals dentro `x-text` e blocchi `<script>`. Riscritti con delimitatori `&quot;...&quot;` o riformulazione testuale.
+### Il ciclo di misura (punto 3)
 
-Inoltre **UX rework** della modale assegnazione ricette su `/admin/stations`: dropdown sostituita da layout a 2 colonne (disponibili / assegnate) con bottone inline `+ Assegna`, search filter, empty state esplicativo (mostrava silenziosamente lista vuota se tutte le ricette erano già assegnate).
+- L'intervallo della ricetta è calcolato dal server (`seconds_to_next_measurement`,
+  `overdue`, `server_time`): sopravvive al cambio task, che è un ricaricamento di
+  pagina, e non si azzera cambiando schermata.
+- Il conteggio **prosegue oltre lo zero**: il ritardo si vede, in rosso.
+- A scadenza l'operatore viene riportato al primo task di misura della ricetta, con
+  cinque secondi di preavviso; chi sta già misurando viene lasciato finire.
+- Cicalino via WebAudio, nessun file audio da caricare.
+- `remeasure` distingue una seconda lettura dello stesso pezzo da un ciclo nuovo:
+  girare il pezzo e rimisurare non deve far ripartire l'intervallo.
 
-Test guard aggiunto: `test_template_js_syntax.py` valida ogni inline `<script>` E ogni espressione Alpine (`x-*`, `@*`, `:*`) della pagina con `node --check`. Cattura automaticamente il bug-class apostrofo. Skip se Node non è installato.
+### Le regole della ricetta (punti 8, 9)
 
-## Layout repository (V2.0.0)
+- `requires_lot`, `requires_serial`, `allow_manual_input` per ricetta, verificati in
+  `measurement_service`: **ogni porta d'ingresso passa di lì**, barcode compreso.
+- Con `allow_manual_input` a falso il tastierino non viene disegnato — non nascosto:
+  il markup non esiste — e il server rifiuta comunque un valore digitato a mano.
+- Il rilevamento del calibro USB è stato reso più tollerante (Enter veloce), perché
+  con la regola stretta una lettura corta come `9.5` veniva scambiata per digitazione.
 
-```
-TieMeasureFlow/
-├── pyproject.toml + uv.lock + .python-version
-├── Dockerfile (backend) + Dockerfile.frontend
-├── docker-compose.dev.yml + docker-compose.yml
-├── nginx/
-├── uploads/                      # volume Docker
-├── docs/                         # raggruppata e indicizzata
-│   ├── README.md (indice)
-│   ├── API.md / DEPLOYMENT.md / USER_GUIDE.md / I18N_SETUP.md
-│   ├── architecture/             # questo file + ROADMAP.md
-│   ├── archive/                  # piani storici
-│   ├── specs/                    # spec esterne (.docx)
-│   └── superpowers/plans/        # piani TDD dettagliati
-└── src/
-    ├── backend/
-    │   ├── main.py / config.py / database.py
-    │   ├── api/{routers,middleware}/
-    │   ├── models/{orm,api}/
-    │   ├── services/
-    │   ├── migrations/
-    │   ├── templates/
-    │   └── tests/
-    └── frontend/
-        └── flask_app/
-            ├── app.py / config.py / compile_translations.py
-            ├── blueprints/ (auth, maker, measure, statistics, admin)
-            ├── services/ (api_client.py)
-            ├── templates/ + static/ + translations/
-            └── tests/
-```
+### Il fuori tolleranza (punto 5)
 
-## Smoke test status
+- Una quota fuori tolleranza e non autorizzata **blocca** il salvataggio della quota
+  successiva e la chiusura del ciclo (409).
+- Conta l'**ultima** lettura di ogni quota: rimisurare la stessa quota resta
+  possibile — il calibro scivola, il pezzo si riposiziona — e una lettura dentro i
+  limiti scioglie il blocco senza chiamare nessuno.
+- L'autorizzazione del capoturno resta scritta sulla misura (`supervisor_id`,
+  `authorised_at`) e finisce in due colonne del CSV di produzione: il file mostra il
+  guasto **e** la decisione.
+- Quanti tentativi siano ammessi è il punto 4, non questo.
 
-Validazione end-to-end in locale (2026-04-26):
+### L'interfaccia dell'operatore (punti 10, 11, 14, 15)
 
-- ✅ MySQL container Docker up, schema creato, alembic stamp head OK
-- ✅ uvicorn `--reload` su :8000, `/api/health` risponde
-- ✅ Seed `/api/setup/seed` con `SETUP_PASSWORD=adriano77` → admin + 4 utenti demo + DEMO-001 + ST-DEFAULT con assegnazione automatica
-- ✅ Login `admin/admin123` via web, sessione persistente
-- ✅ `/admin/stations`: tabella, modal create/edit, modal gestione assegnazioni a 2 colonne con search, eliminazione con cascade
-- ✅ `/admin/users`, `/maker/recipes`, `/measure/select` (filtrato per stazione), `/statistics/dashboard`
-- ✅ Workflow MeasurementTec end-to-end: select_recipe → task_list → task_execute → task_complete (riepilogo con misure)
-- ✅ Hot reload Flask + uvicorn `--reload` + Tailwind watch attivi durante lo sviluppo
+- La ricetta si apre sul primo task (`/measure/start/<id>`), non su un elenco. La
+  lista scende a secondo livello e mostra quali task sono rimasti incompiuti
+  (`GET /api/measurements/task-progress`, contato per quota e non per tentativo).
+- «Fine ciclo misura» è visibile da subito e spento finché mancano quote, con
+  scritto quante ne mancano.
+- Descrizioni con `**grassetto**` e a capo: marcatura, non HTML, così la
+  sanificazione è per costruzione.
+- Una cornice sola per tutte le viste, allineata alla navbar — vedi [`LAYOUT.md`](LAYOUT.md).
+- Nessun percorso dell'operatore porta alla statistica, ed è un test.
+
+### Fuori dalla rete (punto 12)
+
+- Alpine, Plotly, PDF.js (+ worker), Fabric e i font sono nell'installazione, con
+  versione nel nome e impronta SHA-256 verificata da un test.
+- Content-Security-Policy a sola origine locale servita **dal client Flask**: prima
+  esisteva solo sul backend, cioè sulle risposte API e non sulle pagine.
+- Tailwind fissato a `3.4.19` nel `Dockerfile.frontend`.
+
+## Migrazioni
+
+| # | Contenuto |
+|---|---|
+| 001–004 | image_path, stazioni, intervallo di misura + auto-logout, `input_duration_ms` |
+| 005–006 | `production_runs`, misure legate alla produzione |
+| 007 | `task_type` dichiarato |
+| 008 | tipi di evento del ciclo + `task_id` sull'evento |
+| 009 | `requires_lot`, `requires_serial`, `allow_manual_input` |
+| 010 | `supervisor_id`, `authorised_at` sulla misura |
+
+Le 008 e 010 usano il *batch mode* per compatibilità SQLite (i test) e sono state
+verificate anche in resa MySQL (`alembic upgrade X:Y --sql`) prima del deploy.
+
+La 009 imposta `allow_manual_input = 1` sulle ricette esistenti: conserva il
+comportamento in essere invece di irrigidire di colpo ricette già in uso. È una
+scelta, e va confermata dal cliente ricetta per ricetta.
 
 ## Test status
-| Backend (`src/backend/tests/`) | 127 | 3 | Fail pre-esistenti: `test_recipes` (2) + `test_tasks` (1). Nessuno introdotto dalla V2.0.0. |
-| Frontend (`src/frontend/flask_app/tests/`) | 46 | 1 | +2 test post-restructure (`test_template_js_syntax.py`). Fail pre-esistente: `test_save_measurement_proxy`. |
-| **Totale** | **173** | **4** | Tutti i fallimenti tracciati come tech debt da risolvere. |
 
-## Stack confermato
+| | Test | Fail |
+|---|---|---|
+| Backend (`src/backend/tests/`) | 212 | 0 |
+| Frontend (`src/frontend/flask_app/tests/`) | 148 | 0 |
+| **Totale** | **360** | **0** |
 
-- **Backend:** FastAPI + SQLAlchemy 2.0 async + MySQL 8 + Alembic + Pydantic v2 + WeasyPrint + Plotly/Kaleido.
-- **Frontend:** Flask + Jinja2 + Alpine.js + TailwindCSS + Fabric.js 5.3.1 + html5-qrcode + Plotly.js + Flask-Babel.
-- **Deploy:** Docker Compose. Dev = Nginx; Prod = Traefik + Let's Encrypt SSL.
-- **Tooling:** uv (package mgmt), pytest + pytest-asyncio + httpx + aiosqlite (test).
+I quattro fallimenti pre-esistenti tracciati nello snapshot V2.0.0 non ci sono più.
+
+Tre file di test non renderizzano niente e leggono i sorgenti, perché guardano
+proprietà che sopravvivono solo se qualcuno le controlla:
+
+| File | Cosa impedisce |
+|---|---|
+| `test_offline.py` | Una libreria caricata dalla rete, un worker PDF.js lasciato sul CDN, una libreria sostituita senza aggiornare l'impronta |
+| `test_layout_shell.py` | Una vista che torna a dichiararsi la propria larghezza |
+| `test_template_js_syntax.py` | Una traduzione con l'apostrofo dentro una stringa JS a virgolette singole, che spegne Alpine su tutta la pagina |
+
+## Cosa non è stato provato
+
+Va detto perché non si confonda «i test passano» con «funziona in reparto».
+
+- **Il collaudo sul campo non è mai stato percorso.** Le ricette `COLLAUDO-A` e
+  `COLLAUDO-B` sono state seminate sul sistema in esercizio (`scripts/seed_collaudo.py`)
+  insieme a un utente `capoturno`, ma nessuno ha ancora guidato il flusso su un
+  tablet vero.
+- **Il punto 14 non è stato riprodotto su un dispositivo**: è dimostrato che il
+  layout *poteva* muoversi per quattro motivi e che ora non può più, non che
+  l'operatore vedesse esattamente quelli.
+- **Il punto 12 non è stato provato a rete staccata.** È dimostrato che nessuna
+  risorsa esterna viene richiesta. Il giro con la rete staccata va fatto, ed è ora il
+  momento giusto perché la CSP è appena entrata in vigore.
+- **Il cicalino** è il suono del browser. Se serva una colonnina luminosa è D-5.
+
+## Stack
+
+- **Backend:** FastAPI + SQLAlchemy 2.0 async + MySQL 8 + Alembic + Pydantic v2 +
+  WeasyPrint + Plotly/Kaleido. 12 router.
+- **Frontend:** Flask + Jinja2 + Alpine.js 3.15.12 + TailwindCSS 3.4.19 +
+  Fabric.js 5.3.1 + PDF.js 3.11.174 + Plotly.js 2.32.0 + Flask-Babel —
+  **tutte copie locali**, nessun CDN.
+- **Deploy:** Docker Compose. Dev = Nginx; Prod = Traefik + Let's Encrypt.
+- **Tooling:** uv, pytest + pytest-asyncio + httpx + aiosqlite.
 
 ## Decisioni architetturali rilevanti
 
 | Decisione | Stato | Note |
 |---|---|---|
-| Frontend Flask invece di React (deroga vs spec §8) | **Confermata** | Tablet UX server-side, USB calipers/barcode, Fabric.js editor, i18n Babel collaudato. Vedi conversazione 2026-04-25. |
-| NATS messaging (spec §7) | **Skippato** | Monorepo single-host, no microservizi. Nessuno stub `nats_client/` creato. |
-| Envelope risposta `{success,data,error}` (spec §6) | **Rimandato** | Costo 4-5gg refactor + rotture client. Eventuale v2 API in M2. |
-| Header `X-API-Key` vs spec `X-Api-Key` | **Mantenuto attuale** | Rinominare costa 50+ punti di codice + breaking per deploy. Rivedere in M2. |
-| Variabili `.env` (DB_HOST, SERVER_PORT, ...) | **Mantenute attuali** | Rename a SERVICE_NAME/SERVICE_DOMAIN/API_KEY rinviato (impatta deploy esistenti). |
+| Le regole di misura stanno sul server, non sullo schermo | **Confermata (V3.0.0)** | Tracciabilità, inserimento manuale, fuori tolleranza: la schermata può nasconderle, il server le rifiuta. |
+| Una app di stazione per PC, dati sul server | **Decisa il 28/07** (D-3) | Validazione con l'IT di Tràfilo aperta come D-4. |
+| Descrizioni in marcatura, non in HTML | **Confermata** | Niente HTML accettato in ingresso: la sanificazione è per costruzione, non per filtro. |
+| Frontend Flask invece di React (deroga vs spec §8) | **Confermata** | Tablet UX server-side, calibri USB, editor Fabric.js, i18n Babel collaudato. |
+| NATS messaging (spec §7) | **Skippato** | Monorepo single-host, nessun microservizio. |
+| Envelope risposta `{success,data,error}` (spec §6) | **Rimandato** | Costo alto, rotture client. Eventuale API v2. |
+| Header `X-API-Key` vs spec `X-Api-Key` | **Mantenuto attuale** | Rinominare è breaking per i deploy esistenti. |
 
 ## Branch git
 
-- **Default:** `V2.0.0` (lavoro corrente)
-- **Mantenuti:** `V1.0.0` … `V1.0.7` (release branches storiche)
-- **Mergiato e chiuso:** `feature/rev04-phase1-stations` (in `V2.0.0` con commit `ea8e468`)
+- **Corrente:** `V3.0.0`
+- **Precedenti:** `V2.0.0`, `V1.0.0` … `V1.0.7` (release storiche)

docs/archive/STATO_PROGETTO_V2.0.0_2026-04-27.md

@@ -0,0 +1,145 @@
+# Stato Progetto TieMeasureFlow — V2.0.0
+
+> Snapshot al 2026-04-27. Aggiornare ad ogni milestone.
+
+## Versione corrente
+
+**V2.0.0** (in sviluppo, branch `V2.0.0` come default su `git.tielogic.xyz`).
+
+Versione precedente di produzione: `V1.0.7`.
+
+## Sintesi esecutiva
+
+Il sistema base (V1.0.7) è completo e collaudato: ricette, task, misurazioni, SPC, report PDF, gestione utenti, dashboard metrologist. La V2.0.0 in corso aggiunge il primo blocco della migrazione **rev04** (stazioni per-tablet) e ristruttura l'intero monorepo secondo lo standard `python-project-spec-design.md` (uv + `src/backend/` + `src/frontend/flask_app/`).
+
+## Cosa funziona oggi (V2.0.0 — branch corrente)
+
+### Funzionalità ereditate da V1.0.7
+- Autenticazione username/password + API key per-utente, ruoli combinabili (Maker, MeasurementTec, Metrologist) + flag `is_admin`.
+- Recipe versioning copy-on-write: una nuova versione si crea solo se la corrente ha già measurements; altrimenti update in-place.
+- Editor ricette (Maker) con annotation editor Fabric.js (~1200 LOC, collaudato su tablet).
+- Workflow operatore tablet: select_recipe → task_list → task_execute → task_complete, con barcode scanner e numpad touch (input USB calibro con burst detection).
+- Calcolo pass/fail con limiti UTL/UWL/LWL/LTL.
+- Dashboard SPC: capability (Cp/Cpk/Pp/Ppk), control chart (UCL/LCL = mean ± 3σ), istogramma con curva normale, calcoli puro stdlib (no numpy).
+- Report PDF (WeasyPrint + Kaleido SVG).
+- Setup page protetta da `SETUP_PASSWORD` per inizializzazione DB e seed.
+- i18n IT/EN (Flask-Babel + Alpine.js JSON).
+- Tema light/dark via `Alpine.store('theme')` + localStorage.
+
+### Aggiunte V2.0.0 (rev04 Fase 1 — Stazioni per-tablet)
+- Tabelle `stations` + `station_recipe_assignments` (Alembic migration `002_add_stations.py`).
+- Modelli ORM: `Station`, `StationRecipeAssignment` con vincolo unique `(station_id, recipe_id)`.
+- Schemas Pydantic: `StationCreate/Update/Response`, `StationRecipeAssignmentCreate/Response`, `RecipeSummary`.
+- Service `station_service` (CRUD + assegnazioni + cascade delete).
+- Router `/api/stations` con CRUD admin + endpoint operatore `GET /api/stations/by-code/{code}/recipes`.
+- Seed automatico `ST-DEFAULT` con tutte le ricette esistenti (idempotente).
+- Variabile env client `STATION_CODE` letta da `Config`, helper `APIClient.get_station_recipes()`.
+- Filtro `select_recipe`: il client mostra solo le ricette assegnate alla propria stazione, errore se `STATION_CODE` non configurato.
+- **GUI admin completa** in `/admin/stations`: tabella con search, modal create/edit, modal gestione assegnazioni ricette, conferma eliminazione, link in navbar (desktop + mobile).
+- 47 nuovi test (32 server + 15 client) tutti pass.
+
+### Aggiunte V2.0.0 (performance + multi-utente)
+- Gunicorn 5 workers × 4 thread (gthread) — capacità ~20 richieste concorrenti Flask, regge 20+ tablet.
+- Uvicorn 4 workers + `--proxy-headers --forwarded-allow-ips='*'`.
+- Rate limit middleware: identificazione IP reale via `X-Forwarded-For` → `X-Real-IP` → `request.client.host`.
+- Rate limit general 100 → 300 req/min/IP (per-tablet ora, non più condiviso).
+- Flask `ProxyFix(x_for=1, x_proto=1, x_host=1)` per IP reale dietro Nginx.
+- `APIClient` propaga `X-Forwarded-For` + `X-Real-IP` (sia JSON che multipart).
+- 12 test aggiuntivi (7 server + 5 client).
+
+### Aggiunte V2.0.0 (struttura monorepo)
+- `pyproject.toml` unico con extra `server`/`client`/`dev`. Niente più `requirements.txt`.
+- `uv.lock` (77 pacchetti) + `.python-version` (3.11) committati per build riproducibili.
+- Layout `src/backend/` + `src/frontend/flask_app/` (vedi sotto).
+- `Dockerfile` (root) + `Dockerfile.frontend` riscritti con `uv sync --frozen --no-dev --extra server|client`.
+- `docker-compose.{dev,}.yml` con build context `.`.
+- Alembic env.py aggiunge project root a `sys.path`; `script_location = %(here)s` resta valido.
+- `.dockerignore` aggiornato.
+
+### Hardening post-restructure (smoke test 2026-04-26)
+
+Sequenza di smoke test in locale (uvicorn + gunicorn + MySQL Docker) ha fatto emergere quattro regressioni che sarebbero rimaste invisibili al test suite:
+
+- **`src/backend/config.py`**: `env_file` era cwd-relative (`../../.env`). Rotto fuori da `src/backend/`. Risolto con percorso assoluto `Path(__file__).resolve().parents[2] / ".env"`.
+- **`src/backend/models/orm/__init__.py`**: `Station` e `StationRecipeAssignment` non erano esportati, quindi `Base.metadata.create_all` non creava le tabelle stations. Aggiunti agli import.
+- **`.env.example`**: `UPLOAD_DIR=server/uploads` era residuo della vecchia struttura → file landavano fuori dall'albero di progetto. Aggiornato a `UPLOAD_DIR=uploads`.
+- **Apostrofi italiani in template Alpine** (`l'utente`, `nell'eliminazione`, `nell'assegnazione`): chiudevano prematuramente JS string literals dentro `x-text` e blocchi `<script>`. Riscritti con delimitatori `&quot;...&quot;` o riformulazione testuale.
+
+Inoltre **UX rework** della modale assegnazione ricette su `/admin/stations`: dropdown sostituita da layout a 2 colonne (disponibili / assegnate) con bottone inline `+ Assegna`, search filter, empty state esplicativo (mostrava silenziosamente lista vuota se tutte le ricette erano già assegnate).
+
+Test guard aggiunto: `test_template_js_syntax.py` valida ogni inline `<script>` E ogni espressione Alpine (`x-*`, `@*`, `:*`) della pagina con `node --check`. Cattura automaticamente il bug-class apostrofo. Skip se Node non è installato.
+
+## Layout repository (V2.0.0)
+
+```
+TieMeasureFlow/
+├── pyproject.toml + uv.lock + .python-version
+├── Dockerfile (backend) + Dockerfile.frontend
+├── docker-compose.dev.yml + docker-compose.yml
+├── nginx/
+├── uploads/                      # volume Docker
+├── docs/                         # raggruppata e indicizzata
+│   ├── README.md (indice)
+│   ├── API.md / DEPLOYMENT.md / USER_GUIDE.md / I18N_SETUP.md
+│   ├── architecture/             # questo file + ROADMAP.md
+│   ├── archive/                  # piani storici
+│   ├── specs/                    # spec esterne (.docx)
+│   └── superpowers/plans/        # piani TDD dettagliati
+└── src/
+    ├── backend/
+    │   ├── main.py / config.py / database.py
+    │   ├── api/{routers,middleware}/
+    │   ├── models/{orm,api}/
+    │   ├── services/
+    │   ├── migrations/
+    │   ├── templates/
+    │   └── tests/
+    └── frontend/
+        └── flask_app/
+            ├── app.py / config.py / compile_translations.py
+            ├── blueprints/ (auth, maker, measure, statistics, admin)
+            ├── services/ (api_client.py)
+            ├── templates/ + static/ + translations/
+            └── tests/
+```
+
+## Smoke test status
+
+Validazione end-to-end in locale (2026-04-26):
+
+- ✅ MySQL container Docker up, schema creato, alembic stamp head OK
+- ✅ uvicorn `--reload` su :8000, `/api/health` risponde
+- ✅ Seed `/api/setup/seed` con `SETUP_PASSWORD=adriano77` → admin + 4 utenti demo + DEMO-001 + ST-DEFAULT con assegnazione automatica
+- ✅ Login `admin/admin123` via web, sessione persistente
+- ✅ `/admin/stations`: tabella, modal create/edit, modal gestione assegnazioni a 2 colonne con search, eliminazione con cascade
+- ✅ `/admin/users`, `/maker/recipes`, `/measure/select` (filtrato per stazione), `/statistics/dashboard`
+- ✅ Workflow MeasurementTec end-to-end: select_recipe → task_list → task_execute → task_complete (riepilogo con misure)
+- ✅ Hot reload Flask + uvicorn `--reload` + Tailwind watch attivi durante lo sviluppo
+
+## Test status
+| Backend (`src/backend/tests/`) | 127 | 3 | Fail pre-esistenti: `test_recipes` (2) + `test_tasks` (1). Nessuno introdotto dalla V2.0.0. |
+| Frontend (`src/frontend/flask_app/tests/`) | 46 | 1 | +2 test post-restructure (`test_template_js_syntax.py`). Fail pre-esistente: `test_save_measurement_proxy`. |
+| **Totale** | **173** | **4** | Tutti i fallimenti tracciati come tech debt da risolvere. |
+
+## Stack confermato
+
+- **Backend:** FastAPI + SQLAlchemy 2.0 async + MySQL 8 + Alembic + Pydantic v2 + WeasyPrint + Plotly/Kaleido.
+- **Frontend:** Flask + Jinja2 + Alpine.js + TailwindCSS + Fabric.js 5.3.1 + html5-qrcode + Plotly.js + Flask-Babel.
+- **Deploy:** Docker Compose. Dev = Nginx; Prod = Traefik + Let's Encrypt SSL.
+- **Tooling:** uv (package mgmt), pytest + pytest-asyncio + httpx + aiosqlite (test).
+
+## Decisioni architetturali rilevanti
+
+| Decisione | Stato | Note |
+|---|---|---|
+| Frontend Flask invece di React (deroga vs spec §8) | **Confermata** | Tablet UX server-side, USB calipers/barcode, Fabric.js editor, i18n Babel collaudato. Vedi conversazione 2026-04-25. |
+| NATS messaging (spec §7) | **Skippato** | Monorepo single-host, no microservizi. Nessuno stub `nats_client/` creato. |
+| Envelope risposta `{success,data,error}` (spec §6) | **Rimandato** | Costo 4-5gg refactor + rotture client. Eventuale v2 API in M2. |
+| Header `X-API-Key` vs spec `X-Api-Key` | **Mantenuto attuale** | Rinominare costa 50+ punti di codice + breaking per deploy. Rivedere in M2. |
+| Variabili `.env` (DB_HOST, SERVER_PORT, ...) | **Mantenute attuali** | Rename a SERVICE_NAME/SERVICE_DOMAIN/API_KEY rinviato (impatta deploy esistenti). |
+
+## Branch git
+
+- **Default:** `V2.0.0` (lavoro corrente)
+- **Mantenuti:** `V1.0.0` … `V1.0.7` (release branches storiche)
+- **Mergiato e chiuso:** `feature/rev04-phase1-stations` (in `V2.0.0` con commit `ea8e468`)

docs/superpowers/specs/2026-08-16-visionsuite-integration-design.md

@@ -0,0 +1,474 @@
+# Integrazione di VisionSuite in TieMeasureFlow — design
+
+> Progetto: TieMeasureFlow V3.0.0 → V3.1.0
+> Data: 2026-08-16
+> Stato: design approvato, piano di implementazione da scrivere
+
+## Perché
+
+TieMeasureFlow oggi registra misure prese a mano, con il calibro USB o con il
+tastierino. Il passo successivo che il cliente chiede è misurare con una camera:
+acquisire l'immagine di un pezzo e ricavarne le quote, oppure confrontarne il
+profilo con il disegno DXF di riferimento.
+
+La visione artificiale non va scritta qui. Tielogic ha già
+[VisionSuite](ssh://git@git.tielogic.xyz:222/Adriano/visionsuite.git), un
+monorepo di moduli riusabili il cui README indica esplicitamente TieMeasureFlow
+fra i consumatori previsti. Questo documento descrive come agganciarlo.
+
+Il terreno è in parte già preparato. La migrazione `007_task_type` ha introdotto
+il tipo di task dichiarato e l'enum contiene fin da allora due valori che nessuno
+produce né consuma ancora:
+
+```python
+TASK_TYPES = ("note", "measure", "drawing", "xf_compare", "camera_measure")
+MEASURING_TASK_TYPES = ("measure", "camera_measure")
+```
+
+Questo lavoro li riempie.
+
+## Che cosa entra e che cosa resta fuori
+
+**Entra:** l'acquisizione da camera, l'esecuzione di un programma di visione su
+quell'immagine — pattern matching più misure, oppure pattern matching più
+confronto DXF — e l'ingresso dei risultati nel modello di misura che già esiste.
+
+**Resta fuori, per ora:** i semafori USB. Sono previsti dall'architettura — è
+l'agente di stazione a possederli, insieme alle camere — ma non fanno parte di
+questa consegna. Il confine è disegnato perché ci entrino senza ristrutturazioni.
+
+## Le otto decisioni
+
+Sono le scelte prese in fase di brainstorming, con il motivo per cui sono state
+prese. Chi legge fra sei mesi ha bisogno del motivo più della scelta.
+
+### 1. La visione gira sia sul server sia sulla stazione
+
+Il motore è un componente unico distribuito in due posti: accanto al server e
+dentro l'agente di stazione. Questo dà la resilienza — se cade il collegamento,
+la stazione continua a misurare — e permette al server di elaborare in fase di
+produzione, di comporre ricette per conto del Maker e di rifare analisi su
+immagini archiviate.
+
+È coerente con la decisione **D-3**, chiusa il 28/07: dati sul server, app di
+stazione su ogni PC.
+
+### 2. Server e stazione montano la stessa identica versione di VisionSuite
+
+Non è una raccomandazione di build: è un vincolo metrologico. Se il server monta
+un commit e una stazione ne monta un altro, lo stesso pezzo con lo stesso grafo
+può dare due numeri diversi, e nessuno se ne accorge finché non arriva una
+contestazione.
+
+Da qui tre obblighi che il design deve garantire:
+
+- VisionSuite entra come **sottomodulo git su un commit preciso**, lo stesso per
+  entrambi — la forma che il README di VisionSuite prescrive
+- ogni misura porta scritto **con quale versione del motore** è stata prodotta
+- una stazione che monta una versione diversa da quella del server **si rifiuta
+  di misurare** e lo dichiara, invece di produrre numeri plausibili e sbagliati
+
+### 3. L'agente di stazione è sottile
+
+L'agente parla ai device, consegna immagini e — più avanti — piloterà i semafori.
+Non contiene decisioni di visione.
+
+La conseguenza pratica è l'impacchettamento: un agente che non contiene
+VisionSuite non contiene nemmeno PyTorch. Distribuirlo come eseguibile su Windows
+o come container su Linux diventa realistico, e la scelta su driver e formato può
+restare tardiva — che è bene, perché dipende dalla **D-4**, ancora aperta.
+
+L'agente incorpora il motore solo quando la stazione è configurata per elaborare
+in locale, e in quel caso resta comunque un guscio: riceve un grafo, lo esegue,
+restituisce le uscite.
+
+### 4. Il server non importa mai VisionSuite
+
+La visione sul lato server vive in un **servizio separato** nello stesso compose,
+il *vision worker*. Il server FastAPI continua a fare quello che fa —
+autenticazione, ricette, misure, statistica — e resta leggero.
+
+Quattro ragioni concrete: l'immagine dell'API non passa da poche centinaia di
+megabyte a cinque gigabyte; un aggiornamento di VisionSuite non richiede di
+riavviare l'API in produzione; un'elaborazione che va in crash non porta giù le
+richieste degli altri tablet; il giorno in cui la visione chiede una GPU o una
+macchina propria si sposta senza toccare il resto.
+
+Il confine va difeso: se un domani il server importasse VisionSuite per una
+scorciatoia, il vantaggio si perde in silenzio.
+
+### 5. La configurazione dei device sta sul PC
+
+Le camere — e domani i semafori — appartengono alla macchina, non alla ricetta.
+Il server ne tiene uno **specchio** in sola lettura, popolato dall'agente quando
+si connette, perché il Maker deve poter scegliere un device da un elenco stando
+alla scrivania.
+
+Un device che l'agente non dichiara più è **offline**, non cancellato: una
+ricetta che lo referenzia deve continuare a esistere e a spiegare perché non può
+girare.
+
+### 6. Dove si esegue lo decide la configurazione della stazione
+
+Ogni PC dichiara se elabora in locale o delega al server, nella stessa
+configurazione dove già stanno i device. Il default è il server.
+
+Le macchine capaci si prendono il carico e non mandano immagini in rete; quelle
+deboli delegano. Se il server non risponde e la stazione ha il motore, la
+stazione ripiega in locale e lo segna nel risultato. Un impianto misto si governa
+senza toccare le ricette.
+
+La ricetta non dice mai dove gira: legare una scelta di impianto a un documento
+di produzione significa dover rimettere mano alle ricette per spostare il carico.
+
+### 7. Il verdetto lo dà TieMeasureFlow, non il grafo
+
+Gli strumenti di quota di `vs-task` hanno anche loro `nominal`, `plus`, `minus` e
+un `outcome`. Quell'esito viene ignorato e quei limiti restano vuoti: il grafo
+produce numeri, `measurement_service` decide se sono buoni.
+
+Due giudici sullo stesso valore è il modo più rapido per ottenere un CSV che si
+contraddice e una contestazione che non si sa come chiudere.
+
+La conseguenza è che tutto il resto continua a funzionare senza modifiche:
+pass/warning/fail, gate del fuori tolleranza con autorizzazione del capoturno,
+Cp/Cpk, carta di controllo, colonne del CSV di produzione. Una quota da camera è
+una misura come le altre.
+
+### 8. La sorgente immagine è un'astrazione a tre facce
+
+Il motore riceve un'immagine e non sa da dove venga. Le tre sorgenti sono: dal
+vivo da un device tramite l'agente, un'immagine di riferimento salvata sul
+server, un file caricato dal Maker.
+
+Non è un'astrazione nostra: è quella che VisionSuite si è già data — *«espongono
+una sorgente, non una telecamera: è questo che permette di collaudare un impianto
+a cinque telecamere senza avere le cinque telecamere sul tavolo, e di scrivere
+prove che danno sempre lo stesso risultato»*.
+
+Sblocca tre cose che altrimenti non ci sarebbero:
+
+- il Maker compone alla scrivania, senza tenere occupata una stazione né avere il
+  pezzo in mano
+- una ricetta si prova prima di andare in produzione, invece di scoprirne i
+  difetti misurando
+- diventa possibile una **prova di non regressione**: salvando accanto
+  all'immagine il risultato atteso, si riesegue l'intero parco ricette dopo un
+  aggiornamento di VisionSuite e si vede quali quote si sono mosse
+
+Quest'ultima è la rete di sicurezza che rende sostenibile aggiornare il
+sottomodulo, con un motore che sta su due host.
+
+> **Non ancora consegnata.** La colonna `expected_json` esiste, ma nessun
+> percorso la scrive — nemmeno `preview`, che è l'unico posto dove un grafo gira
+> davvero su un'immagine. La prova di non regressione è quindi **promessa da
+> questo documento e non mantenuta da nessuna riga di codice**. Diventa una
+> consegna nominata del piano 1b: «`preview` può promuovere le proprie uscite a
+> `expected_json`». Finché non è fatta, questa decisione non va dichiarata
+> completa.
+
+## Architettura
+
+### I componenti
+
+| Componente | Nuovo | Dove vive | Responsabilità |
+|---|---|---|---|
+| `vendor/visionsuite` | sì | sottomodulo git | i pacchetti `vs-core`, `vs-camera`, `vs-pm2d`, `vs-measure`, `vs-dxf`, su un commit fisso |
+| Vision runner | sì | libreria condivisa | immagine + grafo + calibrazione → quote e artefatti. Unica implementazione, due host |
+| Vision worker | sì | container nel compose | espone il runner dietro una API interna. Il container pesante |
+| Agente di stazione | sì | PC di reparto | possiede i device, acquisisce, apre la connessione verso il server, incorpora il runner se configurato |
+| Backend FastAPI | modificato | container esistente | tipi di task, grafo sul task, registro device, inoltro, ingresso risultati. **Nessun import di VisionSuite** |
+| Frontend Flask | modificato | container esistente | Maker: device, fotogramma, aggancio quote. Measure: schermata del task con camera |
+
+L'aggancio del sottomodulo segue la forma prescritta dal README di VisionSuite:
+dipendenza dai soli pacchetti che servono, dichiarati in path editable.
+
+```toml
+[tool.uv.sources]
+visionsuite = { path = "vendor/visionsuite/packages/vs-core", editable = true }
+vs-camera   = { path = "vendor/visionsuite/packages/vs-camera", editable = true }
+```
+
+I due host non montano lo stesso insieme, ed è il punto dell'intera struttura di
+VisionSuite:
+
+| Pacchetto | Worker | Agente | Perché |
+|---|---|---|---|
+| `vs-core` | sì | sì | strutture di base e calibrazione, servono a entrambi |
+| `vs-pm2d` | sì | solo se elabora in locale | pattern matching |
+| `vs-measure` | sì | solo se elabora in locale | primitive e quote |
+| `vs-dxf` | sì | solo se elabora in locale | confronto col disegno. È quello che porta PyTorch |
+| `vs-camera` | **no** | sì | il worker non tocca hardware: riceve immagini già acquisite |
+
+Un agente su una stazione che delega al server monta quindi due soli pacchetti,
+`vs-core` e `vs-camera`, e resta leggero. È la differenza fra un eseguibile
+distribuibile e un'installazione da cinque gigabyte su ogni PC di reparto.
+
+I nomi di importazione non sono uniformi — `vs-core` si importa come
+`visionsuite`, `vs-pm2d` come `pm2d`, `vs-dxf` come `dxf_compare` — ed è una
+stortura nota e dichiarata a monte, non un errore da correggere qui.
+
+### Il flusso di una misura, esecuzione sul server
+
+```
+operatore apre il task
+  → client chiede al server di eseguire
+  → server chiede all'agente un fotogramma sul device dichiarato
+  → agente acquisisce e carica l'immagine
+  → server passa immagine e grafo al worker
+  → worker restituisce le uscite
+  → server applica pass/fail, gate del fuori tolleranza, salva in measurements
+  → client mostra esito e immagine con le zone dove ha guardato
+```
+
+### Il flusso con esecuzione in locale
+
+Il giro si accorcia: il server manda il grafo all'agente, l'agente acquisisce ed
+esegue, e rimanda uscite più immagine. **Da lì in poi il server fa le stesse
+identiche cose.** Il verdetto, il gate e il salvataggio non cambiano mai posto:
+cambia chi calcola i numeri, non chi decide se sono buoni.
+
+È questo che rende le due strade equivalenti a valle, e che permette di cambiare
+la configurazione di una stazione senza toccare né il frontend né le ricette.
+
+## Modello dati
+
+### Il grafo di visione
+
+`recipe_tasks` prende **`vision_json`** (JSON, nullable): il grafo di `vs-task`
+serializzato, opaco al server.
+
+Il versionamento arriva gratis. Il copy-on-write delle ricette copia già i task
+in profondità, quindi una modifica al grafo su una ricetta con misure produce una
+versione nuova come qualunque altra modifica; in `recipe_service` è un campo in
+più nella copia, non una logica nuova.
+
+Non si riusa `annotations_json`, che sta lì accanto: le annotazioni sono un
+disegno per l'operatore, il grafo è un programma da eseguire. Hanno cicli di vita
+e regole di validazione diversi, e mescolarli significa non poter validare né
+l'uno né l'altro.
+
+### Il legame fra quota del grafo e quota della ricetta
+
+`recipe_subtasks` prende **`vision_output`** (String, nullable): la coppia
+«identificativo dello strumento, nome della sua uscita».
+
+È la stessa forma di legame che `vs-task` usa internamente, e per la stessa
+ragione dichiarata nel suo README: *«Il legame è un ID, non un puntatore»* — un
+riferimento in memoria non si scrive su disco.
+
+Tutto il resto della subtask — nominale, UTL, UWL, LWL, LTL, unità — resta com'è
+e continua a significare quello che significa oggi.
+
+### Le misure
+
+`input_method` passa da `("usb_caliper", "manual")` a
+`("usb_caliper", "manual", "camera")`.
+
+Il contorno della visione **non** va su `measurements`. Quella tabella la leggono
+la statistica e l'export a ogni giro, e appesantirla la rovina. Ma soprattutto
+sarebbe sbagliato di modello: una sola acquisizione produce N quote, e immagine,
+overlay, device e versione del motore sono gli stessi per tutte.
+
+Nasce quindi **`vision_results`**, una riga per **esecuzione**:
+
+| Campo | Perché c'è |
+|---|---|
+| `image_path`, `overlay_path` | l'immagine, e quella con sopra le zone dove ha guardato |
+| `engine_version` | il commit di VisionSuite che ha prodotto i numeri |
+| `executed_on` | `server` o `station` |
+| `station_id`, `device_id` | da quale macchina e quale camera |
+| `calibration_snapshot` | la taratura in vigore in quel momento, **copiata** |
+| `graph_snapshot` | il grafo effettivamente eseguito, **copiato** |
+| `duration_ms`, `executed_at` | quanto è costata, quando |
+
+Su `measurements` una sola colonna nuova: `vision_result_id`, FK nullable.
+
+Calibrazione e grafo si copiano invece di essere referenziati per la stessa
+ragione: una taratura rifatta sei mesi dopo non deve riscrivere il significato di
+una misura già presa.
+
+### Il registro dei device
+
+**`station_devices`**: `station_id`, `code` (l'identificativo locale), `kind`
+(`camera` oggi, `light` domani), `label`, `capabilities` (JSON), `calibration`
+(JSON) con `calibration_taken_at`, `last_seen_at`.
+
+Si popola dall'agente. Il server lo riceve e non lo modifica.
+
+### Le immagini di riferimento
+
+**`vision_reference_images`**: `task_id`, `path`, e la provenienza per intero —
+`station_id`, `device_id`, snapshot della calibrazione, `engine_version`,
+`acquired_at`, una nota — più `expected_json` per l'atteso della prova di non
+regressione.
+
+La provenienza non è rimandabile. Senza, si riesegue il grafo mesi dopo, escono
+numeri diversi, e non c'è modo di distinguere un motore aggiornato da un'ottica
+spostata.
+
+### Migrazione 011
+
+Cinque cose: `vision_json` su `recipe_tasks`, `vision_output` su
+`recipe_subtasks`, `camera` in `input_method_enum`, `vision_result_id` su
+`measurements`, e le tre tabelle nuove.
+
+Più una sesta, **da fare adesso o mai**: `task_type_enum` contiene `xf_compare`,
+refuso per `dxf_compare`. Oggi nessuna riga usa quel valore e correggerlo costa
+una riga. Dal primo task salvato in poi diventa una migrazione di dati, e il
+refuso finisce nelle API pubbliche.
+
+Come le migrazioni 008 e 010, va scritta in *batch mode* per la compatibilità
+SQLite dei test e verificata in resa MySQL con `alembic upgrade X:Y --sql` prima
+del deploy: è la procedura che il progetto si è già dato.
+
+## API
+
+L'agente apre lui la connessione. **Il server non bussa mai a un PC**, e questo
+evita di dover aprire porte sulle macchine di reparto — che è materia della D-4.
+
+| Verso | Endpoint | Cosa fa |
+|---|---|---|
+| agente → server | `WS /api/stations/agent` | si autentica con la chiave di stazione, dichiara i device, resta in ascolto |
+| server → agente | comando `acquire` | «scatta sul device X»; l'agente carica con `POST /api/vision/frames` |
+| server → agente | comando `execute` | esecuzione in locale: manda il grafo, riceve le uscite |
+| client → server | `POST /api/vision/execute` | esegui il task per la produzione in corso; **il server sceglie dove** |
+| client → server | `POST /api/vision/preview` | authoring: esegui su una sorgente qualsiasi, non salvare nulla |
+| client → server | `GET /api/stations/{code}/devices` | l'elenco per il Maker |
+| client → server | `POST /api/vision/reference-images` | salva un fotogramma come riferimento |
+| server → worker | `POST /run` | interno, mai esposto: immagine e grafo → uscite e overlay |
+
+Un solo endpoint decide dove si esegue, ed è `execute`. Il client non sa e non
+deve sapere se ha misurato il server o la stazione: chiede una misura e riceve un
+esito.
+
+## Errori e casi limite
+
+| Caso | Comportamento |
+|---|---|
+| Versione del motore diversa fra stazione e server | la stazione si rifiuta di misurare e lo dichiara. Mai produrre numeri con un motore non concorde |
+| Agente non raggiungibile | il task di visione non è eseguibile e lo dice. Da decidere con il cliente se la stazione debba consentire il ripiego sul calibro |
+| Server non raggiungibile, stazione con motore | la stazione esegue in locale e lo segna in `executed_on` |
+| Device dichiarato dalla ricetta ma offline | la ricetta resta valida, il task spiega perché non può girare |
+| Il grafo non produce un'uscita attesa da una subtask | la misura non si salva, l'errore nomina l'uscita mancante |
+| Elaborazione in crash o oltre il tempo | il worker isola il guasto, l'API resta in piedi, il task riporta l'errore |
+| Device senza calibrazione | le quote in millimetri non si producono. Mai convertire con un fattore implicito |
+
+## Prove
+
+Ai tre file di test che il progetto ha già e che leggono i sorgenti invece di
+renderizzare — `test_offline.py`, `test_layout_shell.py`,
+`test_template_js_syntax.py` — se ne aggiunge la stessa specie:
+
+- **una prova che il server non importi VisionSuite.** È il confine su cui poggia
+  la decisione 4, e senza qualcuno che lo controlli si perde in silenzio
+- **una prova che il commit del sottomodulo dichiarato coincida** con quello che
+  server e agente montano
+
+Per il resto, il runner si prova su immagini di riferimento salvate — che è
+precisamente ciò per cui l'astrazione della sorgente esiste. Nessuna prova
+richiede una camera collegata.
+
+## Quello che questo design non risolve
+
+- **Il dimensionamento del server.** Il pattern matching è pesante e, con
+  l'esecuzione lato server come default, la CPU è contesa fra tutte le stazioni.
+  Va aggiunto alla **D-4**, che finora chiedeva solo quale macchina e quanto
+  disco: ora chiede anche quanti core e quanta RAM.
+- **La conservazione delle immagini.** Due politiche distinte, entrambe da
+  portare in D-4: le immagini di riferimento sono poche e vivono quanto la
+  ricetta; quelle di produzione sono una per pezzo e crescono senza limite, e
+  vogliono una scadenza o una regola del tipo «si conservano solo quelle dei
+  pezzi fuori tolleranza».
+- **Che cosa fa la stazione se la rete cade** e non è configurata per elaborare
+  in locale: si blocca, o consente il calibro e segna il task come non
+  eseguibile? È una domanda per il cliente.
+- **L'impacchettamento dell'agente.** Le macchine possono essere Linux o Windows
+  e i driver possono stare sulla macchina o nel container. La decisione è
+  volutamente tardiva; il design la rende tale tenendo l'agente sottile e neutro.
+
+## Correzioni imposte dall'esecuzione
+
+Il piano derivato da questo documento è stato eseguito il 16/08/2026. Sei cose
+che qui erano scritte male o taciute, corrette dai fatti.
+
+### Il gate del fuori tolleranza vale per l'acquisizione, non per la quota
+
+La decisione 7 e il testo del punto 5 parlano di «la quota successiva». Quella
+regola era stata pensata per la misura **manuale e sequenziale**, dove esiste
+davvero un momento successivo in cui l'operatore potrebbe tirare avanti con una
+lettura cattiva irrisolta.
+
+Una camera produce tutte le N quote **nello stesso istante**. Dentro
+un'acquisizione non esiste un «tirare avanti», quindi lì non c'è nulla da
+presidiare — e applicare la regola sequenziale a un evento simultaneo produceva
+un blocco permanente: la lettura fuori tolleranza non veniva mai registrata,
+quindi nessuno poteva autorizzarla, e rifare la stessa foto ripeteva l'errore.
+
+Il gate si valuta **una volta per acquisizione**: se la misura pendente è fra le
+quote che questa acquisizione riprodurrà non blocca — riacquisire *è* rimisurare
+quelle quote, che questo documento già dichiara la via d'uscita che non richiede
+il capoturno. Fra due acquisizioni, e alla chiusura del ciclo, blocca come prima.
+
+### Il repository è deliberatamente bi-ambiente
+
+`vs-task`, `vs-measure` e `vs-pm2d` dichiarano `requires-python >= 3.13`;
+`vs-core` si ferma a 3.10. Backend e frontend restano su **3.11**, l'albero della
+visione gira su **3.13**, e l'extra `vision` porta i marker d'ambiente che
+tengono separate le due risoluzioni.
+
+Questo rafforza la decisione 4 più di come era argomentata: con VisionSuite dentro
+il server FastAPI, questa scoperta avrebbe costretto l'intero backend a 3.13.
+
+### La versione del motore si timbra al build
+
+`vendor/visionsuite/.git` è un file che rimanda a `../../.git/modules/...`, che
+nel contesto di build non esiste: interrogare git dentro il container non
+funziona. `engine_version()` legge `VISION_ENGINE_VERSION` dall'ambiente, ripiega
+su git nei checkout di sviluppo, e **alza un errore** se non risolve né l'uno né
+l'altro. Il build fallisce se la variabile manca.
+
+### La divisione degli extra compra esplicitezza, non peso
+
+Gli extra sono due — `vision` per il runner, `vision-worker` per il servizio web
+— ed è la forma giusta. Ma la motivazione che sembrava ovvia è falsa: `vs-pm2d`
+dichiara **di suo** `fastapi`, `uvicorn`, `python-multipart` e `pillow`, quindi
+un agente di stazione che fa pattern matching se li porterà dietro comunque. La
+correzione vera è a monte — una libreria non dovrebbe dipendere da un server web
+— e va aggiunta alle verifiche su VisionSuite qui sotto.
+
+### L'immagine di produzione non viene ancora conservata
+
+`vision_results.image_path` e `overlay_path` migrano e restano `NULL`: nulla le
+scrive. Il flusso descritto in questo documento si chiude con «il client mostra
+esito e immagine», e finché quelle colonne sono vuote una misura contestata non
+ha prova materiale. Consegna nominata del piano 1b, insieme all'overlay.
+
+### Prima di applicare le migrazioni in produzione
+
+La 011 rinomina un valore di enum su una tabella in esercizio. Va contato prima
+che nessuna riga usi `xf_compare`: in strict mode una riga sopravvissuta fa
+fallire l'`ALTER` a metà deploy, fuori strict mode diventa `''` in silenzio.
+Attenzione anche alla finestra: il `MODIFY` su `recipe_tasks` è una rinomina di
+valore e userà `ALGORITHM=COPY`, mentre quello su `measurements.input_method`
+aggiunge in coda e dovrebbe restare `INPLACE` — da confermare sulle dimensioni
+reali, perché `measurements` è la tabella grossa.
+
+## Da verificare prima di cominciare
+
+VisionSuite è stato lavorato fino al 15/08/2026 e il suo ultimo commit è
+*«Rassegna delle firme pubbliche prima che si congelino»*: le API si stanno
+stabilizzando proprio ora. Prima di fissare il commit del sottomodulo:
+
+- **il suo README è indietro rispetto al codice.** Dichiara `vs-measure` «da
+  scrivere» quando invece esiste, con circa 4.865 righe e tredici file di test, e
+  dichiara `vs-dxf` non avviabile quando l'import di `common.calibration` non c'è
+  più. Va riletto lo stato reale, non quello documentato.
+- **il test di calibrazione dato per rosso** — la focale recuperata a 373 contro
+  456 attesi, con tolleranza al 5% — va verificato sullo stato attuale. Il README
+  di VisionSuite avverte di guardarlo *prima* di costruirci sopra misure in
+  millimetri, che è esattamente ciò che questo lavoro fa. Le calibrazioni sono
+  state toccate il 13/08 e lo stato potrebbe essere cambiato.
+- **il peso dell'ambiente.** `vs-dxf` dipende da PyTorch e l'installazione
+  completa arriva a circa 5 GB. È la ragione per cui il worker è un container a
+  parte, e va confermato che il solo sottoinsieme necessario pesi meno.

pyproject.toml

@@ -30,6 +30,8 @@ server = [
     "plotly>=5.0.0",
     "kaleido>=0.2.0",
     "weasyprint>=62.0",
+    "pdfplumber>=0.10.0",
+    "httpx>=0.27.0",
 ]
 
 # Frontend (Flask tablet UI).
@@ -51,6 +53,32 @@ dev = [
     "coverage>=7.0.0",
 ]
 
+# Vision runner (VisionSuite submodule): what src/vision/runner.py needs to
+# execute a graph. vs-task/vs-measure/vs-pm2d require Python >=3.13 upstream;
+# marker keeps the base 3.11 resolution untouched. numpy is declared
+# explicitly - runner.py imports it directly, and it must not depend on
+# being pulled in as a side effect of vs-pm2d's own dependencies.
+vision = [
+    "visionsuite; python_version>='3.13'",
+    "vs-task; python_version>='3.13'",
+    "vs-measure; python_version>='3.13'",
+    "vs-pm2d; python_version>='3.13'",
+    "numpy; python_version>='3.13'",
+]
+
+# Vision worker (src/vision_worker/): the FastAPI shell around the runner,
+# on top of `vision`. Kept separate on purpose (Ruling R9) - VisionSuite's
+# whole structure exists so a consumer installs only what it needs, and a
+# later station agent embeds the same runner without being a web service.
+# It must not drag fastapi/uvicorn onto shop-floor PCs.
+vision-worker = [
+    "tiemeasureflow[vision]",
+    "fastapi>=0.110.0; python_version>='3.13'",
+    "uvicorn[standard]>=0.30.0; python_version>='3.13'",
+    "pillow>=10.0.0; python_version>='3.13'",
+    "python-multipart>=0.0.6; python_version>='3.13'",
+]
+
 [project.scripts]
 # Backend
 server = "uvicorn:run"  # placeholder, real CMD lives in Dockerfile
@@ -68,6 +96,12 @@ packages = ["src/backend", "src/frontend"]
 # Pin the resolver to the deps we declared; reproducible builds.
 package = false
 
+[tool.uv.sources]
+visionsuite = { path = "vendor/visionsuite/packages/vs-core", editable = true }
+vs-task     = { path = "vendor/visionsuite/packages/vs-task", editable = true }
+vs-measure  = { path = "vendor/visionsuite/packages/vs-measure", editable = true }
+vs-pm2d     = { path = "vendor/visionsuite/packages/vs-pm2d", editable = true }
+
 [tool.pytest.ini_options]
 asyncio_mode = "auto"
-testpaths = ["src/backend/tests", "src/frontend/flask_app/tests"]
+testpaths = ["src/backend/tests", "src/frontend/flask_app/tests", "src/vision/tests", "src/vision_worker/tests"]

scripts/seed_collaudo.py

@@ -0,0 +1,321 @@
+"""Seed the recipes used to accept points 2, 3, 5, 8, 9 and 11 on a real screen.
+
+The tests prove the rules hold; this puts them in front of an operator. Two
+recipes, because they need opposite settings:
+
+  COLLAUDO-A  typing allowed, lot compulsory, two measurement tasks with
+              documental tasks before, between and after them, interval of two
+              minutes. This is the one to drive: the measurement loop, the
+              bidirectional countdown, the out-of-tolerance gate, the cycle that
+              closes only on the last measurement task.
+
+  COLLAUDO-B  caliper only, lot and serial compulsory. This one is to look at:
+              the keypad is not drawn, and Avvia does not start until both
+              traceability fields are filled in.
+
+Run it inside the server container, where src.backend and the database settings
+are already in place:
+
+    docker compose cp scripts/seed_collaudo.py server:/tmp/seed_collaudo.py
+    docker compose exec server uv run python /tmp/seed_collaudo.py --station ST-DEFAULT
+
+Re-running finds the recipes already there and stops. --replace deletes them
+first, along with anything measured against them - which is what you want between
+one acceptance session and the next, and never what you want by accident.
+"""
+import argparse
+import asyncio
+import sys
+
+from sqlalchemy import delete, select
+
+sys.path.insert(0, "/app")
+
+from src.backend.database import async_session_factory  # noqa: E402
+from src.backend.models.orm.measurement import Measurement  # noqa: E402
+from src.backend.models.orm.production import (  # noqa: E402
+    ProductionEvent, ProductionRun,
+)
+from src.backend.models.orm.recipe import Recipe, RecipeVersion  # noqa: E402
+from src.backend.models.orm.station import (  # noqa: E402
+    Station, StationRecipeAssignment,
+)
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask  # noqa: E402
+from src.backend.models.orm.user import User  # noqa: E402
+from src.backend.services import auth_service  # noqa: E402
+
+CODES = ("COLLAUDO-A", "COLLAUDO-B")
+
+# Limits chosen so every outcome is one deliberate keystroke away:
+#   10.00 -> conforme      10.30 -> attenzione (dentro tolleranza)
+#   12.00 -> fuori tolleranza, serve il capoturno
+QUOTES_A1 = [
+    dict(marker_number=1, description="Diametro esterno",
+         nominal=10.0, ltl=9.5, lwl=9.8, uwl=10.2, utl=10.5, unit="mm"),
+    dict(marker_number=2, description="Spessore parete",
+         nominal=25.0, ltl=24.5, lwl=24.8, uwl=25.2, utl=25.5, unit="mm"),
+]
+QUOTES_A2 = [
+    dict(marker_number=3, description="Lunghezza totale",
+         nominal=100.0, ltl=99.0, lwl=99.5, uwl=100.5, utl=101.0, unit="mm"),
+]
+QUOTES_B = [
+    dict(marker_number=1, description="Diametro con calibro",
+         nominal=10.0, ltl=9.5, lwl=9.8, uwl=10.2, utl=10.5, unit="mm"),
+]
+
+# A description with a blank line and a bold phrase: point 11 is visible or it is
+# not, and this is where you look.
+NOTA_INIZIALE = (
+    "Prendere il pezzo dalla cassetta **a sinistra** della postazione.\n"
+    "Pulire la superficie prima di misurare.\n\n"
+    "**Attenzione**: non misurare pezzi ancora caldi."
+)
+NOTA_INTERMEDIA = (
+    "Girare il pezzo di **90 gradi** prima della seconda misura.\n"
+    "Se il pezzo non appoggia in piano, ripetere il posizionamento."
+)
+NOTA_FINALE = (
+    "Riporre il pezzo nella cassetta **a destra**.\n"
+    "Segnalare al capoturno qualsiasi quota fuori tolleranza."
+)
+
+
+async def _find_recipes(session, codes):
+    result = await session.execute(select(Recipe).where(Recipe.code.in_(codes)))
+    return list(result.scalars().all())
+
+
+async def _delete_recipes(session, recipes):
+    """Remove the collaudo recipes and everything measured against them."""
+    for recipe in recipes:
+        versions = (await session.execute(
+            select(RecipeVersion.id).where(RecipeVersion.recipe_id == recipe.id)
+        )).scalars().all()
+        runs = (await session.execute(
+            select(ProductionRun.id).where(ProductionRun.recipe_id == recipe.id)
+        )).scalars().all()
+
+        if versions:
+            await session.execute(
+                delete(Measurement).where(Measurement.version_id.in_(versions))
+            )
+        if runs:
+            await session.execute(
+                delete(ProductionEvent).where(ProductionEvent.run_id.in_(runs))
+            )
+            await session.execute(
+                delete(ProductionRun).where(ProductionRun.id.in_(runs))
+            )
+        await session.execute(
+            delete(StationRecipeAssignment).where(
+                StationRecipeAssignment.recipe_id == recipe.id
+            )
+        )
+        # Versions, tasks and subtasks go with the recipe through the ORM cascade.
+        await session.delete(recipe)
+    await session.flush()
+
+
+async def _add_task(session, version_id, order_index, title, task_type,
+                    directive=None, description=None, quotes=()):
+    task = RecipeTask(
+        version_id=version_id,
+        order_index=order_index,
+        title=title,
+        task_type=task_type,
+        directive=directive,
+        description=description,
+    )
+    session.add(task)
+    await session.flush()
+    for quote in quotes:
+        session.add(RecipeSubtask(task_id=task.id, **quote))
+    await session.flush()
+    return task
+
+
+async def _create_recipe(session, user_id, *, code, name, description,
+                         interval, requires_lot, requires_serial,
+                         allow_manual_input):
+    recipe = Recipe(
+        code=code,
+        name=name,
+        description=description,
+        created_by=user_id,
+        active=True,
+        measurement_interval_minutes=interval,
+        requires_lot=requires_lot,
+        requires_serial=requires_serial,
+        allow_manual_input=allow_manual_input,
+    )
+    session.add(recipe)
+    await session.flush()
+
+    version = RecipeVersion(
+        recipe_id=recipe.id,
+        version_number=1,
+        is_current=True,
+        created_by=user_id,
+        change_notes="Ricetta di collaudo",
+    )
+    session.add(version)
+    await session.flush()
+    return recipe, version
+
+
+async def _assign(session, station, recipe, user_id):
+    session.add(StationRecipeAssignment(
+        station_id=station.id, recipe_id=recipe.id, assigned_by=user_id,
+    ))
+    await session.flush()
+
+
+async def _ensure_supervisor(session, username, password):
+    """A capoturno with the role, not an admin standing in for one.
+
+    An admin is accepted by the authorisation check, but the acceptance session
+    should exercise the role the shop floor will actually use.
+    """
+    existing = (await session.execute(
+        select(User).where(User.username == username)
+    )).scalar_one_or_none()
+    if existing is not None:
+        return existing, False
+
+    user = await auth_service.create_user(
+        session,
+        username=username,
+        password=password,
+        display_name="Capo Turno (collaudo)",
+        roles=["Supervisor", "MeasurementTec"],
+    )
+    await session.flush()
+    return user, True
+
+
+async def seed(station_code: str, replace: bool, supervisor_password: str) -> int:
+    async with async_session_factory() as session:
+        owner = (await session.execute(
+            select(User).where(User.is_admin == True)  # noqa: E712
+        )).scalars().first()
+        if owner is None:
+            print("Nessun utente amministratore: impossibile intestare le ricette.")
+            return 2
+
+        station = (await session.execute(
+            select(Station).where(Station.code == station_code)
+        )).scalar_one_or_none()
+        if station is None:
+            print(f"Stazione '{station_code}' inesistente.")
+            return 2
+
+        existing = await _find_recipes(session, CODES)
+        if existing and not replace:
+            found = ", ".join(r.code for r in existing)
+            print(f"Ricette di collaudo gia' presenti ({found}).")
+            print("Usa --replace per rifarle da zero (cancella anche le misure).")
+            return 1
+        if existing:
+            await _delete_recipes(session, existing)
+            print(f"Rimosse: {', '.join(r.code for r in existing)}")
+
+        # ---- COLLAUDO-A: the one to drive ----
+        recipe_a, version_a = await _create_recipe(
+            session, owner.id,
+            code="COLLAUDO-A",
+            name="Collaudo — ciclo di misura",
+            description=(
+                "Due task di misura con task documentali intorno. "
+                "Lotto obbligatorio, inserimento manuale consentito, "
+                "intervallo di 2 minuti."
+            ),
+            interval=2,
+            requires_lot=True,
+            requires_serial=False,
+            allow_manual_input=True,
+        )
+        await _add_task(
+            session, version_a.id, 0, "Preparazione del pezzo", "note",
+            directive="Leggere prima di iniziare",
+            description=NOTA_INIZIALE,
+        )
+        await _add_task(
+            session, version_a.id, 1, "Misura 1 — diametro e spessore", "measure",
+            directive="Misurare le due quote in sequenza",
+            description="Quota **1** e quota **2**.\nUsare il calibro o il tastierino.",
+            quotes=QUOTES_A1,
+        )
+        await _add_task(
+            session, version_a.id, 2, "Riposizionamento", "note",
+            directive="Fra la prima e la seconda misura",
+            description=NOTA_INTERMEDIA,
+        )
+        await _add_task(
+            session, version_a.id, 3, "Misura 2 — lunghezza", "measure",
+            directive="Ultimo task di misura: chiude il ciclo",
+            description="Alla conferma di questa quota **riparte l'intervallo**.",
+            quotes=QUOTES_A2,
+        )
+        await _add_task(
+            session, version_a.id, 4, "Chiusura", "note",
+            directive="Dopo le misure",
+            description=NOTA_FINALE,
+        )
+        await _assign(session, station, recipe_a, owner.id)
+
+        # ---- COLLAUDO-B: the one to look at ----
+        recipe_b, version_b = await _create_recipe(
+            session, owner.id,
+            code="COLLAUDO-B",
+            name="Collaudo — solo calibro",
+            description=(
+                "Lotto e seriale obbligatori, inserimento manuale vietato: "
+                "il tastierino non compare e Avvia non parte finche' mancano i dati."
+            ),
+            interval=5,
+            requires_lot=True,
+            requires_serial=True,
+            allow_manual_input=False,
+        )
+        await _add_task(
+            session, version_b.id, 0, "Misura con calibro", "measure",
+            directive="Il valore deve arrivare dal calibro",
+            description="Nessun tastierino: la ricetta **non ammette valori digitati**.",
+            quotes=QUOTES_B,
+        )
+        await _assign(session, station, recipe_b, owner.id)
+
+        supervisor, created = await _ensure_supervisor(
+            session, "capoturno", supervisor_password,
+        )
+
+        await session.commit()
+
+    print(f"Ricette create e assegnate alla stazione {station_code}:")
+    print("  COLLAUDO-A  ciclo di misura, lotto obbligatorio, intervallo 2 min")
+    print("  COLLAUDO-B  solo calibro, lotto e seriale obbligatori")
+    if created:
+        print(f"Capoturno creato: {supervisor.username} / {supervisor_password}")
+        print("  cambiare la password prima di qualunque uso reale.")
+    else:
+        print(f"Capoturno gia' presente: {supervisor.username} (password invariata)")
+    return 0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--station", default="ST-DEFAULT",
+                        help="codice della stazione a cui assegnare le ricette")
+    parser.add_argument("--replace", action="store_true",
+                        help="cancella le ricette di collaudo esistenti e le misure fatte su di esse")
+    parser.add_argument("--supervisor-password", default="Collaudo2026!",
+                        help="password del capoturno, se va creato")
+    args = parser.parse_args()
+    return asyncio.run(
+        seed(args.station, args.replace, args.supervisor_password)
+    )
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/backend/api/middleware/__init__.py

@@ -6,6 +6,7 @@ from src.backend.api.middleware.api_key import (
     require_maker,
     require_measurement_tec,
     require_metrologist,
+    require_supervisor,
     require_admin_user,
 )
 from src.backend.api.middleware.logging import AccessLogMiddleware
@@ -19,6 +20,7 @@ __all__ = [
     "require_maker",
     "require_measurement_tec",
     "require_metrologist",
+    "require_supervisor",
     "require_admin_user",
     "AccessLogMiddleware",
     "RateLimitMiddleware",

src/backend/api/middleware/api_key.py

@@ -68,4 +68,5 @@ def require_admin():
 require_maker = require_role("Maker")
 require_measurement_tec = require_role("MeasurementTec")
 require_metrologist = require_role("Metrologist")
+require_supervisor = require_role("Supervisor")
 require_admin_user = require_admin()

src/backend/api/middleware/security_headers.py

@@ -10,16 +10,18 @@ from starlette.middleware.base import BaseHTTPMiddleware
 
 from src.backend.config import settings
 
-# Content Security Policy - allows CDN resources used by the client
+# Content Security Policy - same-origin only: every third-party library now ships with
+# the app (see frontend static/vendor/), so no CDN host needs allowing. The install runs
+# on an isolated shop-floor network where an outbound fetch would simply fail.
 # Note: 'unsafe-eval' required for Plotly.js runtime evaluation in SPC charts
 CSP = (
     "default-src 'self'; "
-    "script-src 'self' 'unsafe-inline' 'unsafe-eval' "
-    "https://cdn.tailwindcss.com https://cdn.jsdelivr.net https://cdn.plot.ly; "
-    "style-src 'self' 'unsafe-inline' https://fonts.googleapis.com; "
-    "font-src 'self' https://fonts.gstatic.com; "
+    "script-src 'self' 'unsafe-inline' 'unsafe-eval'; "
+    "style-src 'self' 'unsafe-inline'; "
+    "font-src 'self'; "
     "img-src 'self' data: blob:; "
-    "connect-src 'self'"
+    "connect-src 'self'; "
+    "worker-src 'self' blob:"
 )
 
 

src/backend/api/routers/files.py

@@ -1,5 +1,5 @@
 """File upload/download/delete router for images and PDFs."""
-import os
+import logging
 import re
 from pathlib import Path
 
@@ -11,6 +11,8 @@ from src.backend.config import settings
 from src.backend.api.middleware.api_key import get_current_user, require_maker
 from src.backend.models.orm.user import User
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter(prefix="/api/files", tags=["files"])
 
 # Allowed MIME types
@@ -65,6 +67,37 @@ def sanitize_filename(filename: str) -> str:
     return filename
 
 
+def resolve_upload_path(file_path: str) -> Path:
+    """Resolve a user-supplied relative path inside the uploads directory.
+
+    Raises:
+        HTTPException: 404 if the path escapes the uploads directory,
+            cannot be resolved, or does not point to an existing file.
+    """
+    try:
+        full_path = (settings.upload_path / file_path).resolve()
+        if not full_path.is_relative_to(settings.upload_path.resolve()):
+            raise HTTPException(
+                status_code=status.HTTP_403_FORBIDDEN,
+                detail="Access denied",
+            )
+    except HTTPException:
+        raise
+    except (OSError, RuntimeError, ValueError):
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail="File not found",
+        )
+
+    if not full_path.exists() or not full_path.is_file():
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail="File not found",
+        )
+
+    return full_path
+
+
 def generate_thumbnail(image_path: Path, thumbnail_path: Path, max_size: tuple[int, int] = (200, 200)):
     """Generate a thumbnail for an image."""
     try:
@@ -73,7 +106,7 @@ def generate_thumbnail(image_path: Path, thumbnail_path: Path, max_size: tuple[i
             img.save(thumbnail_path, quality=85)
     except Exception as e:
         # If thumbnail generation fails, we continue without it
-        print(f"Thumbnail generation failed: {e}")
+        logger.warning("Thumbnail generation failed for %s: %s", image_path, e)
 
 
 @router.post("/upload")
@@ -164,30 +197,7 @@ async def get_file(
 
     Requires authentication but no specific role.
     """
-    # Construct full path
-    full_path = settings.upload_path / file_path
-
-    # Security: ensure path is within uploads directory
-    try:
-        full_path = full_path.resolve()
-        if not str(full_path).startswith(str(settings.upload_path.resolve())):
-            raise HTTPException(
-                status_code=status.HTTP_403_FORBIDDEN,
-                detail="Access denied",
-            )
-    except Exception:
-        raise HTTPException(
-            status_code=status.HTTP_404_NOT_FOUND,
-            detail="File not found",
-        )
-
-    # Check if file exists
-    if not full_path.exists() or not full_path.is_file():
-        raise HTTPException(
-            status_code=status.HTTP_404_NOT_FOUND,
-            detail="File not found",
-        )
-
+    full_path = resolve_upload_path(file_path)
     return FileResponse(full_path)
 
 
@@ -200,29 +210,7 @@ async def delete_file(
 
     Also deletes associated thumbnail if it exists.
     """
-    # Construct full path
-    full_path = settings.upload_path / file_path
-
-    # Security: ensure path is within uploads directory
-    try:
-        full_path = full_path.resolve()
-        if not str(full_path).startswith(str(settings.upload_path.resolve())):
-            raise HTTPException(
-                status_code=status.HTTP_403_FORBIDDEN,
-                detail="Access denied",
-            )
-    except Exception:
-        raise HTTPException(
-            status_code=status.HTTP_404_NOT_FOUND,
-            detail="File not found",
-        )
-
-    # Check if file exists
-    if not full_path.exists() or not full_path.is_file():
-        raise HTTPException(
-            status_code=status.HTTP_404_NOT_FOUND,
-            detail="File not found",
-        )
+    full_path = resolve_upload_path(file_path)
 
     # Delete thumbnail if it exists
     if full_path.parent.name != "thumbnails":

src/backend/api/routers/measurements.py

@@ -19,11 +19,14 @@ from src.backend.models.orm.recipe import RecipeVersion
 from src.backend.models.orm.setting import SystemSetting
 from src.backend.models.orm.user import User
 from src.backend.models.api.measurement import (
+    MeasurementAuthorisation,
     MeasurementBatchCreate,
     MeasurementCreate,
     MeasurementListResponse,
     MeasurementResponse,
+    TaskProgressListResponse,
 )
+from src.backend.services import auth_service, measurement_service
 from src.backend.services.measurement_service import save_measurement
 
 router = APIRouter(prefix="/api/measurements", tags=["measurements"])
@@ -46,6 +49,8 @@ async def create_measurement(
             lot_number=data.lot_number,
             serial_number=data.serial_number,
             input_method=data.input_method,
+            input_duration_ms=data.input_duration_ms,
+            production_run_id=data.production_run_id,
         )
         return MeasurementResponse.model_validate(measurement)
     except ValueError as e:
@@ -55,6 +60,78 @@ async def create_measurement(
         )
 
 
+@router.get("/pending-authorisation", response_model=MeasurementResponse | None)
+async def get_pending_authorisation(
+    version_id: int = Query(..., gt=0),
+    production_run_id: int | None = Query(None, gt=0),
+    user: User = Depends(require_measurement_tec),
+    db: AsyncSession = Depends(get_db),
+):
+    """The out-of-tolerance measurement holding this operator up, or null.
+
+    Asked on load, so that reloading the page is not a way past the gate: the
+    screen finds the same block waiting for it that the server enforces.
+    """
+    blocking = await measurement_service.pending_authorisation(
+        db, version_id, user.id, production_run_id,
+    )
+    return MeasurementResponse.model_validate(blocking) if blocking else None
+
+
+@router.get("/task-progress", response_model=TaskProgressListResponse)
+async def get_task_progress(
+    version_id: int = Query(..., gt=0),
+    production_run_id: int | None = Query(None, gt=0),
+    user: User = Depends(require_measurement_tec),
+    db: AsyncSession = Depends(get_db),
+):
+    """How many quotes of each measurement task are already taken.
+
+    The task list is where an operator picks up work they left behind, and until
+    now it showed a task begun and a task never touched identically.
+    """
+    tasks = await measurement_service.task_progress(
+        db, version_id, user.id, production_run_id,
+    )
+    return TaskProgressListResponse(tasks=tasks)
+
+
+@router.post("/{measurement_id}/authorise", response_model=MeasurementResponse)
+async def authorise_measurement(
+    measurement_id: int,
+    action: MeasurementAuthorisation,
+    user: User = Depends(require_measurement_tec),
+    db: AsyncSession = Depends(get_db),
+):
+    """Let an out-of-tolerance value stand, on the capoturno's name.
+
+    Credentials go to this endpoint rather than to a separate check because the
+    approval has to end up attached to the measurement. A check whose answer is
+    thrown away is what this point had before: a modal that closed.
+    """
+    result = await db.execute(
+        select(Measurement).where(Measurement.id == measurement_id)
+    )
+    measurement = result.scalar_one_or_none()
+    if measurement is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND, detail="Measurement not found",
+        )
+    if measurement.pass_fail != "fail":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="This measurement is within tolerance: there is nothing to authorise",
+        )
+
+    supervisor = await auth_service.authorise_supervisor(
+        db, action.supervisor_username, action.supervisor_password,
+    )
+    measurement = await measurement_service.authorise_measurement(
+        db, measurement, supervisor,
+    )
+    return MeasurementResponse.model_validate(measurement)
+
+
 @router.post("/batch", response_model=list[MeasurementResponse])
 async def create_measurement_batch(
     data: MeasurementBatchCreate,
@@ -74,6 +151,8 @@ async def create_measurement_batch(
                 lot_number=measurement_data.lot_number,
                 serial_number=measurement_data.serial_number,
                 input_method=measurement_data.input_method,
+                input_duration_ms=measurement_data.input_duration_ms,
+                production_run_id=measurement_data.production_run_id,
             )
             measurements.append(measurement)
         return [MeasurementResponse.model_validate(m) for m in measurements]
@@ -84,6 +163,44 @@ async def create_measurement_batch(
         )
 
 
+def _build_measurement_filters(
+    recipe_id: int | None,
+    version_id: int | None,
+    subtask_id: int | None,
+    measured_by: int | None,
+    lot_number: str | None,
+    serial_number: str | None,
+    date_from: datetime | None,
+    date_to: datetime | None,
+    pass_fail: str | None,
+) -> list:
+    """Build SQLAlchemy filter conditions shared by list and CSV export."""
+    filters = []
+    if recipe_id is not None:
+        # Filter by recipe via subquery on RecipeVersion
+        version_ids = select(RecipeVersion.id).where(
+            RecipeVersion.recipe_id == recipe_id
+        )
+        filters.append(Measurement.version_id.in_(version_ids))
+    if version_id is not None:
+        filters.append(Measurement.version_id == version_id)
+    if subtask_id is not None:
+        filters.append(Measurement.subtask_id == subtask_id)
+    if measured_by is not None:
+        filters.append(Measurement.measured_by == measured_by)
+    if lot_number is not None:
+        filters.append(Measurement.lot_number == lot_number)
+    if serial_number is not None:
+        filters.append(Measurement.serial_number == serial_number)
+    if date_from is not None:
+        filters.append(Measurement.measured_at >= date_from)
+    if date_to is not None:
+        filters.append(Measurement.measured_at <= date_to)
+    if pass_fail is not None:
+        filters.append(Measurement.pass_fail == pass_fail)
+    return filters
+
+
 @router.get("/", response_model=MeasurementListResponse)
 async def get_measurements(
     recipe_id: int | None = Query(None),
@@ -108,30 +225,10 @@ async def get_measurements(
     filters. Measurements carry no PII beyond numeric values + a recipe
     reference, so role-gating beyond authentication isn't justified.
     """
-    # Build filter conditions
-    filters = []
-    if recipe_id is not None:
-        # Filter by recipe via subquery on RecipeVersion
-        version_ids = select(RecipeVersion.id).where(
-            RecipeVersion.recipe_id == recipe_id
+    filters = _build_measurement_filters(
+        recipe_id, version_id, subtask_id, measured_by, lot_number,
+        serial_number, date_from, date_to, pass_fail,
     )
-        filters.append(Measurement.version_id.in_(version_ids))
-    if version_id is not None:
-        filters.append(Measurement.version_id == version_id)
-    if subtask_id is not None:
-        filters.append(Measurement.subtask_id == subtask_id)
-    if measured_by is not None:
-        filters.append(Measurement.measured_by == measured_by)
-    if lot_number is not None:
-        filters.append(Measurement.lot_number == lot_number)
-    if serial_number is not None:
-        filters.append(Measurement.serial_number == serial_number)
-    if date_from is not None:
-        filters.append(Measurement.measured_at >= date_from)
-    if date_to is not None:
-        filters.append(Measurement.measured_at <= date_to)
-    if pass_fail is not None:
-        filters.append(Measurement.pass_fail == pass_fail)
 
     # Build query
     query = select(Measurement).where(and_(*filters) if filters else True)
@@ -196,29 +293,10 @@ async def export_measurements_csv(
     decimal_setting = decimal_result.scalar_one_or_none()
     decimal_separator = decimal_setting.setting_value if decimal_setting else "."
 
-    # Build filter conditions (same as get_measurements)
-    filters = []
-    if recipe_id is not None:
-        version_ids = select(RecipeVersion.id).where(
-            RecipeVersion.recipe_id == recipe_id
+    filters = _build_measurement_filters(
+        recipe_id, version_id, subtask_id, measured_by, lot_number,
+        serial_number, date_from, date_to, pass_fail,
     )
-        filters.append(Measurement.version_id.in_(version_ids))
-    if version_id is not None:
-        filters.append(Measurement.version_id == version_id)
-    if subtask_id is not None:
-        filters.append(Measurement.subtask_id == subtask_id)
-    if measured_by is not None:
-        filters.append(Measurement.measured_by == measured_by)
-    if lot_number is not None:
-        filters.append(Measurement.lot_number == lot_number)
-    if serial_number is not None:
-        filters.append(Measurement.serial_number == serial_number)
-    if date_from is not None:
-        filters.append(Measurement.measured_at >= date_from)
-    if date_to is not None:
-        filters.append(Measurement.measured_at <= date_to)
-    if pass_fail is not None:
-        filters.append(Measurement.pass_fail == pass_fail)
 
     # Query all matching measurements
     query = select(Measurement).where(and_(*filters) if filters else True)

src/backend/api/routers/production.py

@@ -0,0 +1,163 @@
+"""Production runs router - open, read, and drive the life of a production.
+
+Deliberately stateless: nothing lives in process memory. With one station app
+installed per PC the database is the only shared place, and any station app must be
+able to ask "what is going on here?" and get the same answer.
+"""
+from fastapi import APIRouter, Depends, Query, status
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.api.middleware.api_key import get_current_user
+from src.backend.database import get_db
+from src.backend.models.api.production import (
+    CycleCompletePayload,
+    ProductionEventResponse,
+    ProductionRunCreate,
+    ProductionRunResponse,
+    ProductionRunWithEventsResponse,
+    RemeasurePayload,
+    SupervisorAction,
+)
+from src.backend.models.orm.production import ProductionRun
+from src.backend.models.orm.user import User
+from src.backend.services import auth_service, production_service
+
+router = APIRouter(prefix="/api/production-runs", tags=["production"])
+
+_DERIVED = {
+    "seconds_to_next_measurement", "overdue", "server_time", "measurement_task_ids",
+}
+
+
+async def _as_response(db: AsyncSession, run: ProductionRun) -> ProductionRunResponse:
+    """The stored run plus everything the client cannot work out on its own."""
+    return ProductionRunResponse(
+        **ProductionRunResponse.model_validate(run).model_dump(exclude=_DERIVED),
+        **production_service.describe(run),
+        measurement_task_ids=await production_service.measurement_task_ids(
+            db, run.version_id,
+        ),
+    )
+
+
+@router.post("", response_model=ProductionRunResponse, status_code=status.HTTP_201_CREATED)
+async def open_production_run(
+    data: ProductionRunCreate,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Start a production at a station."""
+    run = await production_service.open_run(db, data, user)
+    return await _as_response(db, run)
+
+
+@router.get("/current", response_model=ProductionRunResponse | None)
+async def get_current_production_run(
+    station_code: str = Query(..., min_length=1),
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """The run open at this station, or null.
+
+    This is what every page asks on load instead of keeping the timer in memory.
+    """
+    station = await production_service.get_station_by_code(db, station_code)
+    run = await production_service.get_open_run_for_station(db, station.id)
+    return await _as_response(db, run) if run is not None else None
+
+
+@router.get("/{run_id}", response_model=ProductionRunWithEventsResponse)
+async def get_production_run(
+    run_id: int,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """A run with its full trace - the history a production did not have before."""
+    run = await production_service.get_run(db, run_id)
+    events = await production_service.list_run_events(db, run_id)
+    return ProductionRunWithEventsResponse(
+        **(await _as_response(db, run)).model_dump(),
+        events=[ProductionEventResponse.model_validate(e) for e in events],
+    )
+
+
+@router.post("/{run_id}/cycle", response_model=ProductionRunResponse)
+async def complete_measurement_cycle(
+    run_id: int,
+    payload: CycleCompletePayload | None = None,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Record a finished measurement task; the last one restarts the interval."""
+    run = await production_service.get_run(db, run_id)
+    run = await production_service.complete_cycle(
+        db, run, user,
+        task_id=payload.task_id if payload else None,
+        note=payload.note if payload else None,
+    )
+    return await _as_response(db, run)
+
+
+@router.post("/{run_id}/remeasure", response_model=ProductionRunResponse)
+async def remeasure_in_cycle(
+    run_id: int,
+    payload: RemeasurePayload | None = None,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Turn the piece over and measure again, without closing the cycle."""
+    run = await production_service.get_run(db, run_id)
+    run = await production_service.remeasure(
+        db, run, user,
+        task_id=payload.task_id if payload else None,
+        note=payload.note if payload else None,
+    )
+    return await _as_response(db, run)
+
+
+@router.post("/{run_id}/pause", response_model=ProductionRunResponse)
+async def pause_production_run(
+    run_id: int,
+    action: SupervisorAction,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Fermo linea - requires a supervisor."""
+    run = await production_service.get_run(db, run_id)
+    supervisor = await auth_service.authorise_supervisor(
+        db, action.supervisor_username, action.supervisor_password,
+    )
+    run = await production_service.pause_run(db, run, user, supervisor, note=action.note)
+    return await _as_response(db, run)
+
+
+@router.post("/{run_id}/resume", response_model=ProductionRunResponse)
+async def resume_production_run(
+    run_id: int,
+    action: SupervisorAction,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Restart a stopped line - requires a supervisor."""
+    run = await production_service.get_run(db, run_id)
+    supervisor = await auth_service.authorise_supervisor(
+        db, action.supervisor_username, action.supervisor_password,
+    )
+    run = await production_service.resume_run(db, run, user, supervisor, note=action.note)
+    return await _as_response(db, run)
+
+
+@router.post("/{run_id}/close", response_model=ProductionRunResponse)
+async def close_production_run(
+    run_id: int,
+    action: SupervisorAction,
+    user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Fine produzione - requires a supervisor. Stops the timer for good."""
+    run = await production_service.get_run(db, run_id)
+    supervisor = await auth_service.authorise_supervisor(
+        db, action.supervisor_username, action.supervisor_password,
+    )
+    run = await production_service.close_run(db, run, user, supervisor, note=action.note)
+    return await _as_response(db, run)

src/backend/api/routers/recipes.py

@@ -1,5 +1,7 @@
-"""Recipe router - CRUD, versioning, barcode lookup."""
-from fastapi import APIRouter, Depends, Query
+"""Recipe router - CRUD, versioning, barcode lookup, AI parsing."""
+import logging
+
+from fastapi import APIRouter, Depends, File, HTTPException, Query, UploadFile, status
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from src.backend.database import get_db
@@ -164,3 +166,31 @@ async def get_measurement_count(
     """
     count = await recipe_service.get_measurement_count(db, recipe_id, version_number)
     return {"recipe_id": recipe_id, "version_number": version_number, "measurement_count": count}
+
+
+@router.post("/{recipe_id}/parse-technical-sheet")
+async def parse_technical_sheet(
+    recipe_id: int,
+    file: UploadFile = File(...),
+    _user: User = Depends(require_maker),
+    db: AsyncSession = Depends(get_db),
+):
+    """Parse a PDF technical sheet using AI and return suggested tasks."""
+    from src.backend.services import ai_service
+
+    if not file.content_type or "pdf" not in file.content_type:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Il file deve essere un PDF")
+
+    pdf_bytes = await file.read()
+    if len(pdf_bytes) > 20 * 1024 * 1024:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="File troppo grande (max 20MB)")
+
+    try:
+        tasks = await ai_service.parse_technical_sheet(pdf_bytes)
+    except ValueError as e:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(e))
+    except Exception as e:
+        logging.getLogger(__name__).error("AI parsing error: %s", e)
+        raise HTTPException(status_code=status.HTTP_502_BAD_GATEWAY, detail="Errore nel servizio AI")
+
+    return {"recipe_id": recipe_id, "suggested_tasks": tasks}

src/backend/api/routers/setup.py

@@ -15,6 +15,7 @@ from pydantic import BaseModel
 from sqlalchemy import inspect as sa_inspect, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from src.backend.api.routers.users import VALID_ROLES
 from src.backend.config import settings
 from src.backend.database import Base, engine, async_session_factory
 from src.backend.models.orm import (
@@ -26,7 +27,7 @@ from src.backend.services.measurement_service import calculate_pass_fail
 
 router = APIRouter(prefix="/api/setup", tags=["setup"])
 
-_templates_dir = Path(__file__).resolve().parent.parent / "templates"
+_templates_dir = Path(__file__).resolve().parent.parent.parent / "templates"
 templates = Jinja2Templates(directory=str(_templates_dir))
 
 
@@ -215,7 +216,7 @@ async def _seed_default_station(session: AsyncSession, admin_user: User) -> None
 async def setup_page(request: Request):
     """Serve the setup page HTML."""
     _check_setup_enabled()
-    return templates.TemplateResponse("setup/setup.html", {"request": request})
+    return templates.TemplateResponse(request, "setup/setup.html")
 
 
 @router.get("/status")
@@ -544,7 +545,6 @@ async def manage_user(body: ManageUserBody):
     """Create or update a user. user_id=null creates new, user_id=int updates."""
     _check_password(body.password)
 
-    VALID_ROLES = {"Maker", "MeasurementTec", "Metrologist"}
     invalid = set(body.roles) - VALID_ROLES
     if invalid:
         raise HTTPException(

src/backend/api/routers/stations.py

@@ -11,6 +11,8 @@ from src.backend.models.api.station import (
     StationResponse,
     StationRecipeAssignmentCreate,
     StationRecipeAssignmentResponse,
+    StationResetResponse,
+    StationWithRecipesResponse,
     RecipeSummary,
 )
 from src.backend.services import station_service
@@ -18,15 +20,36 @@ from src.backend.services import station_service
 router = APIRouter(prefix="/api/stations", tags=["stations"])
 
 
-@router.get("", response_model=list[StationResponse])
+@router.get("", response_model=list[StationWithRecipesResponse])
 async def list_stations(
     active_only: bool = False,
     admin: User = Depends(require_admin_user),
     db: AsyncSession = Depends(get_db),
 ):
-    """List all stations (admin only)."""
+    """List all stations with the recipes assigned to each (admin only).
+
+    The admin station list shows which products a station handles, so the
+    assignments travel with the station instead of needing a call per row.
+    Station.assignments is selectin-loaded, so this costs no extra query.
+
+    Only active recipes are listed, matching what an operator actually sees
+    at that station via /by-code/{code}/recipes.
+    """
     stations = await station_service.list_stations(db, active_only=active_only)
-    return [StationResponse.model_validate(s) for s in stations]
+    return [
+        StationWithRecipesResponse(
+            **StationResponse.model_validate(s).model_dump(),
+            recipes=sorted(
+                (
+                    RecipeSummary.model_validate(a.recipe)
+                    for a in s.assignments
+                    if a.recipe is not None and a.recipe.active
+                ),
+                key=lambda r: r.code,
+            ),
+        )
+        for s in stations
+    ]
 
 
 @router.post("", response_model=StationResponse, status_code=status.HTTP_201_CREATED)
@@ -128,6 +151,22 @@ async def assign_recipe_to_station(
     return StationRecipeAssignmentResponse.model_validate(assignment)
 
 
+@router.delete("/{station_id}/recipes", response_model=StationResetResponse)
+async def reset_station_recipes(
+    station_id: int,
+    admin: User = Depends(require_admin_user),
+    db: AsyncSession = Depends(get_db),
+):
+    """Clear every recipe assignment of a station (admin only).
+
+    Backs the per-row reset in the admin list: the station is left with no
+    recipes and can be reassigned from scratch, without deleting the station
+    itself. Idempotent - resetting an already empty station reports removed=0.
+    """
+    removed = await station_service.unassign_all_recipes(db, station_id)
+    return StationResetResponse(station_id=station_id, removed=removed)
+
+
 @router.delete(
     "/{station_id}/recipes/{recipe_id}",
     status_code=status.HTTP_204_NO_CONTENT,

src/backend/api/routers/tasks.py

@@ -172,12 +172,14 @@ async def create_task(
     new_task = RecipeTask(
         version_id=new_version.id,
         order_index=max_order + 1,
+        task_type=data.task_type,
         title=data.title,
         directive=data.directive,
         description=data.description,
         file_path=data.file_path,
         file_type=data.file_type,
         annotations_json=data.annotations_json,
+        vision_json=data.vision_json,
     )
     db.add(new_task)
     await db.flush()
@@ -196,6 +198,7 @@ async def create_task(
             ltl=sub_data.ltl,
             unit=sub_data.unit,
             image_path=sub_data.image_path,
+            vision_output=sub_data.vision_output,
         )
         db.add(sub)
 
@@ -223,7 +226,10 @@ async def reorder_tasks(
     result = await db.execute(
         select(RecipeTask)
         .where(RecipeTask.id.in_(data.task_ids))
-        .options(selectinload(RecipeTask.subtasks))
+        .options(
+            selectinload(RecipeTask.subtasks),
+            selectinload(RecipeTask.version),
+        )
     )
     tasks_map = {t.id: t for t in result.scalars().all()}
 
@@ -336,6 +342,7 @@ async def create_subtask(
         ltl=data.ltl,
         unit=data.unit,
         image_path=data.image_path,
+        vision_output=data.vision_output,
     )
     db.add(subtask)
     await db.flush()

src/backend/api/routers/users.py

@@ -11,6 +11,10 @@ from src.backend.services.auth_service import create_user, hash_password, regene
 
 router = APIRouter(prefix="/api/users", tags=["users"])
 
+# Combinable user roles. Supervisor (capoturno) authorizes out-of-tolerance
+# overrides during measurement without needing full admin privileges.
+VALID_ROLES = {"Maker", "MeasurementTec", "Metrologist", "Supervisor"}
+
 
 @router.get("", response_model=list[UserResponse])
 async def list_users(
@@ -40,12 +44,11 @@ async def create_new_user(
             detail=f"Username '{data.username}' already exists",
         )
     # Validate roles
-    valid_roles = {"Maker", "MeasurementTec", "Metrologist"}
     for role in data.roles:
-        if role not in valid_roles:
+        if role not in VALID_ROLES:
             raise HTTPException(
                 status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
-                detail=f"Invalid role '{role}'. Valid roles: {valid_roles}",
+                detail=f"Invalid role '{role}'. Valid roles: {VALID_ROLES}",
             )
     user = await create_user(
         db,
@@ -77,12 +80,11 @@ async def update_user(
     update_data = data.model_dump(exclude_unset=True)
     # Validate roles if provided
     if "roles" in update_data:
-        valid_roles = {"Maker", "MeasurementTec", "Metrologist"}
         for role in update_data["roles"]:
-            if role not in valid_roles:
+            if role not in VALID_ROLES:
                 raise HTTPException(
                     status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
-                    detail=f"Invalid role '{role}'. Valid roles: {valid_roles}",
+                    detail=f"Invalid role '{role}'. Valid roles: {VALID_ROLES}",
                 )
     for field, value in update_data.items():
         setattr(user, field, value)

src/backend/api/routers/vision.py

@@ -0,0 +1,133 @@
+"""The client asks for a measurement and gets an outcome.
+
+It does not know, and must not know, whether the server or the station computed
+the numbers: that is what lets a station's configuration change without touching
+the frontend or the recipes.
+"""
+from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile, status
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.api.middleware.api_key import require_maker, require_measurement_tec
+from src.backend.api.routers.files import ALLOWED_IMAGE_TYPES, validate_file_size
+from src.backend.config import settings
+from src.backend.database import get_db
+from src.backend.models.api.vision import VisionExecuteResponse, VisionMeasurementResult
+from src.backend.models.orm.task import RecipeTask
+from src.backend.models.orm.user import User
+from src.backend.services import vision_service
+
+router = APIRouter(prefix="/api/vision", tags=["vision"])
+
+# A vision graph measures pixels, not a PDF: stricter than files.py's
+# ALLOWED_TYPES, which also accepts documents for technical drawings.
+_IMAGE_EXTENSIONS = {
+    "image/jpeg": "jpg",
+    "image/png": "png",
+    "image/gif": "gif",
+    "image/webp": "webp",
+}
+
+
+async def _task_or_404(db: AsyncSession, task_id: int) -> RecipeTask:
+    """Shared by the endpoints of this router."""
+    task = (await db.execute(
+        select(RecipeTask).where(RecipeTask.id == task_id)
+    )).scalar_one_or_none()
+    if task is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND, detail="Task not found"
+        )
+    return task
+
+
+async def _read_image(image: UploadFile) -> bytes:
+    """Type checked from the header before reading, size checked once the body
+    is in hand - the same two-step convention `files.py::upload_file` already
+    uses, reused rather than reinvented here.
+    """
+    if image.content_type not in ALLOWED_IMAGE_TYPES:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail=f"File type {image.content_type} not allowed. Must be an image.",
+        )
+    content = await image.read()
+    if not validate_file_size(len(content)):
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail=(
+                f"File size {len(content)} bytes exceeds maximum "
+                f"{settings.max_upload_size_mb}MB"
+            ),
+        )
+    return content
+
+
+@router.post("/execute", response_model=VisionExecuteResponse)
+async def execute(
+    task_id: int = Form(...),
+    image: UploadFile = File(...),
+    lot_number: str | None = Form(None),
+    serial_number: str | None = Form(None),
+    production_run_id: int | None = Form(None),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_measurement_tec),
+) -> VisionExecuteResponse:
+    task = await _task_or_404(db, task_id)
+
+    result, saved = await vision_service.execute_task(
+        db,
+        task,
+        await _read_image(image),
+        current_user.id,
+        version_id=task.version_id,
+        lot_number=lot_number,
+        serial_number=serial_number,
+        production_run_id=production_run_id,
+    )
+    await db.commit()
+
+    return VisionExecuteResponse(
+        vision_result_id=result.id,
+        engine_version=result.engine_version,
+        measurements=[
+            VisionMeasurementResult(
+                subtask_id=m.subtask_id, value=float(m.value), pass_fail=m.pass_fail,
+            )
+            for m in saved
+        ],
+    )
+
+
+# Composing a recipe and keeping a reference image are Maker operations, not
+# measuring: both endpoints below gate on `require_maker`, not the
+# `require_measurement_tec` that `execute` above uses.
+
+
+@router.post("/reference-images")
+async def add_reference_image(
+    task_id: int = Form(...),
+    image: UploadFile = File(...),
+    note: str | None = Form(None),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_maker),
+) -> dict:
+    task = await _task_or_404(db, task_id)
+    content = await _read_image(image)
+    reference = await vision_service.save_reference_image(
+        db, task.id, content, note=note,
+        extension=_IMAGE_EXTENSIONS[image.content_type],
+    )
+    await db.commit()
+    return {"id": reference.id, "path": reference.path, "note": reference.note}
+
+
+@router.post("/preview")
+async def preview(
+    task_id: int = Form(...),
+    image: UploadFile = File(...),
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(require_maker),
+) -> dict:
+    task = await _task_or_404(db, task_id)
+    return await vision_service.preview(db, task, await _read_image(image))

src/backend/config.py

@@ -34,6 +34,13 @@ class Settings(BaseSettings):
     # Setup page (empty = disabled)
     setup_password: str | None = None
 
+    # AI / OpenRouter (for technical sheet parsing)
+    openrouter_api_key: str | None = None
+    openrouter_model: str = "anthropic/claude-sonnet-4"
+
+    # Vision worker (internal network only, reachable as `vision` in Compose)
+    vision_worker_url: str = "http://vision:8100"
+
     @property
     def database_url(self) -> str:
         """Async MySQL connection string."""

src/backend/main.py

@@ -21,6 +21,8 @@ from src.backend.api.routers.reports import router as reports_router
 from src.backend.api.routers.statistics import router as statistics_router
 from src.backend.api.routers.setup import router as setup_router
 from src.backend.api.routers.stations import router as stations_router
+from src.backend.api.routers.production import router as production_router
+from src.backend.api.routers.vision import router as vision_router
 
 
 @asynccontextmanager
@@ -73,6 +75,8 @@ app.include_router(statistics_router)
 app.include_router(reports_router)
 app.include_router(setup_router)
 app.include_router(stations_router)
+app.include_router(production_router)
+app.include_router(vision_router)
 
 
 @app.get("/api/health")

src/backend/migrations/env.py

@@ -36,6 +36,8 @@ from src.backend.models.orm.task import RecipeTask, RecipeSubtask  # noqa: F401
 from src.backend.models.orm.measurement import Measurement  # noqa: F401
 from src.backend.models.orm.access_log import AccessLog  # noqa: F401
 from src.backend.models.orm.setting import SystemSetting, RecipeVersionAudit  # noqa: F401
+from src.backend.models.orm.station import Station, StationRecipeAssignment  # noqa: F401
+from src.backend.models.orm.production import ProductionRun, ProductionEvent  # noqa: F401
 
 target_metadata = Base.metadata
 

src/backend/migrations/versions/003_add_measurement_interval_and_auto_logout.py

@@ -0,0 +1,24 @@
+"""add measurement_interval_minutes to recipes
+
+Revision ID: 003_measurement_interval
+Revises: 002_add_stations
+Create Date: 2026-05-23
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '003_measurement_interval'
+down_revision: Union[str, None] = '002_add_stations'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column('recipes', sa.Column('measurement_interval_minutes', sa.SmallInteger(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column('recipes', 'measurement_interval_minutes')

src/backend/migrations/versions/004_add_input_duration_ms.py

@@ -0,0 +1,27 @@
+"""add input_duration_ms to measurements
+
+Revision ID: 004_input_duration
+Revises: 003_measurement_interval
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '004_input_duration'
+down_revision: Union[str, None] = '003_measurement_interval'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        'measurements',
+        sa.Column('input_duration_ms', sa.Integer, nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column('measurements', 'input_duration_ms')

src/backend/migrations/versions/005_add_production_runs.py

@@ -0,0 +1,100 @@
+"""add production_runs and production_events
+
+Gives a production a life of its own: before this its state lived in the Alpine
+component of task_execute.html, so changing task - a full page load - lost the timer,
+the cycle count and the "production started" flag.
+
+Revision ID: 005_production_runs
+Revises: 004_input_duration
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '005_production_runs'
+down_revision: Union[str, None] = '004_input_duration'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        'production_runs',
+        sa.Column('id', sa.Integer, primary_key=True, autoincrement=True),
+        sa.Column('station_id', sa.Integer, sa.ForeignKey('stations.id'), nullable=False),
+        sa.Column('recipe_id', sa.Integer, sa.ForeignKey('recipes.id'), nullable=False),
+        sa.Column('version_id', sa.Integer, sa.ForeignKey('recipe_versions.id'), nullable=False),
+        sa.Column('operator_id', sa.Integer, sa.ForeignKey('users.id'), nullable=False),
+        sa.Column('lot_number', sa.String(100), nullable=True),
+        sa.Column('serial_number', sa.String(100), nullable=True),
+        sa.Column(
+            'status',
+            sa.Enum('running', 'paused', 'closed', name='production_run_status_enum'),
+            nullable=False,
+            server_default='running',
+        ),
+        sa.Column('measurement_interval_minutes', sa.SmallInteger, nullable=True),
+        sa.Column('next_measurement_at', sa.DateTime, nullable=True),
+        sa.Column('cycle_count', sa.Integer, nullable=False, server_default='0'),
+        sa.Column('started_at', sa.DateTime, nullable=False, server_default=sa.func.now()),
+        sa.Column('paused_at', sa.DateTime, nullable=True),
+        sa.Column('closed_at', sa.DateTime, nullable=True),
+        sa.Column('closed_by', sa.Integer, sa.ForeignKey('users.id'), nullable=True),
+        # Mirrors station_id while open, NULL once closed. The unique constraint makes
+        # "one open run per station" a database guarantee rather than a race; repeated
+        # NULLs do not collide, so closed runs are free to pile up.
+        sa.Column('active_station_id', sa.Integer, nullable=True),
+        # Declared inline rather than added afterwards: adding a constraint is an ALTER,
+        # which SQLite cannot do, and the test databases are SQLite.
+        sa.UniqueConstraint('active_station_id', name='uq_production_runs_active_station'),
+        mysql_engine='InnoDB',
+        mysql_charset='utf8mb4',
+    )
+    op.create_index('ix_production_runs_station_id', 'production_runs', ['station_id'])
+    op.create_index('ix_production_runs_recipe_id', 'production_runs', ['recipe_id'])
+    op.create_index('ix_production_runs_version_id', 'production_runs', ['version_id'])
+    op.create_index('ix_production_runs_operator_id', 'production_runs', ['operator_id'])
+    op.create_index('ix_production_runs_lot_number', 'production_runs', ['lot_number'])
+    op.create_index('ix_production_runs_serial_number', 'production_runs', ['serial_number'])
+    op.create_index('ix_production_runs_status', 'production_runs', ['status'])
+    op.create_index(
+        'ix_production_runs_next_measurement_at', 'production_runs', ['next_measurement_at'],
+    )
+    op.create_index(
+        'ix_production_runs_station_status', 'production_runs', ['station_id', 'status'],
+    )
+
+    op.create_table(
+        'production_events',
+        sa.Column('id', sa.Integer, primary_key=True, autoincrement=True),
+        sa.Column(
+            'run_id', sa.Integer,
+            sa.ForeignKey('production_runs.id', ondelete='CASCADE'), nullable=False,
+        ),
+        sa.Column(
+            'event_type',
+            sa.Enum(
+                'start', 'cycle_completed', 'line_stop', 'resume', 'close',
+                name='production_event_type_enum',
+            ),
+            nullable=False,
+        ),
+        sa.Column('user_id', sa.Integer, sa.ForeignKey('users.id'), nullable=False),
+        sa.Column('supervisor_id', sa.Integer, sa.ForeignKey('users.id'), nullable=True),
+        sa.Column('note', sa.Text, nullable=True),
+        sa.Column('created_at', sa.DateTime, nullable=False, server_default=sa.func.now()),
+        mysql_engine='InnoDB',
+        mysql_charset='utf8mb4',
+    )
+    op.create_index('ix_production_events_run_id', 'production_events', ['run_id'])
+    op.create_index('ix_production_events_event_type', 'production_events', ['event_type'])
+    op.create_index('ix_production_events_user_id', 'production_events', ['user_id'])
+    op.create_index('ix_production_events_created_at', 'production_events', ['created_at'])
+
+
+def downgrade() -> None:
+    op.drop_table('production_events')
+    op.drop_table('production_runs')

src/backend/migrations/versions/006_link_measurements_to_production.py

@@ -0,0 +1,63 @@
+"""link measurements to their production run
+
+Fine produzione has to hand the measurements of the whole production to the
+statistics file. Without this column "the measurements of this production" is not a
+query: measurements only knew their recipe version, lot and serial, none of which
+delimits one run from the next on the same recipe and lot.
+
+Nullable on purpose: measurements taken before this - and any taken outside a
+production - simply have no run.
+
+Revision ID: 006_measurement_run
+Revises: 005_production_runs
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '006_measurement_run'
+down_revision: Union[str, None] = '005_production_runs'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Batch mode, because the column carries a foreign key: adding a constraint is
+    # an ALTER that SQLite cannot do, and alembic raises rather than silently
+    # dropping it. On MySQL this is a plain ALTER; on SQLite it rebuilds the table.
+    # Keeping the key is worth the ceremony - ON DELETE SET NULL means a measurement,
+    # the record that matters in an audit, survives its run being deleted.
+    with op.batch_alter_table('measurements') as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                'production_run_id', sa.Integer,
+                sa.ForeignKey(
+                    'production_runs.id',
+                    ondelete='SET NULL',
+                    # Named because batch mode requires it, and because an anonymous
+                    # constraint cannot be referred to later.
+                    name='fk_measurements_production_run',
+                ),
+                nullable=True,
+            ),
+        )
+    op.create_index(
+        'ix_measurements_production_run_id', 'measurements', ['production_run_id'],
+    )
+
+    # Where the statistics file for a closed run was written. No constraint, so a
+    # plain ALTER is enough.
+    op.add_column(
+        'production_runs',
+        sa.Column('statistics_path', sa.String(500), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column('production_runs', 'statistics_path')
+    op.drop_index('ix_measurements_production_run_id', table_name='measurements')
+    with op.batch_alter_table('measurements') as batch_op:
+        batch_op.drop_column('production_run_id')

src/backend/migrations/versions/007_add_task_type.py

@@ -0,0 +1,64 @@
+"""add an explicit type to recipe tasks
+
+Until now the kind of a task was deduced: quotes present meant a measurement,
+otherwise a note. A measurement task whose quotes had not been entered yet was
+therefore treated as a note - the system behaved differently depending on how
+complete the recipe happened to be.
+
+Existing rows are classified by the rule that was in force, so nothing changes
+behaviour on upgrade: tasks with quotes become 'measure'. Tasks with no quotes but
+a drawing attached become 'drawing' rather than 'note' - that is what they already
+displayed as, and calling them notes would be the one place this migration did
+change behaviour.
+
+Revision ID: 007_task_type
+Revises: 006_measurement_run
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '007_task_type'
+down_revision: Union[str, None] = '006_measurement_run'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+TASK_TYPES = ('note', 'measure', 'drawing', 'xf_compare', 'camera_measure')
+
+
+def upgrade() -> None:
+    op.add_column(
+        'recipe_tasks',
+        sa.Column(
+            'task_type',
+            sa.Enum(*TASK_TYPES, name='task_type_enum'),
+            nullable=False,
+            server_default='note',
+        ),
+    )
+    op.create_index('ix_recipe_tasks_task_type', 'recipe_tasks', ['task_type'])
+
+    # Backfill by the rule the frontend used, so no recipe changes behaviour.
+    op.execute(
+        """
+        UPDATE recipe_tasks
+           SET task_type = 'measure'
+         WHERE id IN (SELECT DISTINCT task_id FROM recipe_subtasks)
+        """
+    )
+    op.execute(
+        """
+        UPDATE recipe_tasks
+           SET task_type = 'drawing'
+         WHERE task_type = 'note'
+           AND file_path IS NOT NULL
+        """
+    )
+
+
+def downgrade() -> None:
+    op.drop_index('ix_recipe_tasks_task_type', table_name='recipe_tasks')
+    op.drop_column('recipe_tasks', 'task_type')

src/backend/migrations/versions/008_measurement_loop_events.py

@@ -0,0 +1,74 @@
+"""widen the production event trace for the measurement loop
+
+The loop needs to distinguish two things the trace could not say before: a
+measurement task finished while the cycle is still running (task_measured), and the
+piece turned over and measured again without closing the cycle (remeasure). Only
+the last measurement task of a recipe still produces cycle_completed, because only
+that one restarts the interval.
+
+Both values are added in one go. Widening a MySQL enum rewrites the table, so the
+values that are already planned go in now rather than one migration at a time -
+the same reasoning as the task types in 007.
+
+The events that happen inside a cycle also need to say which task they are about,
+so the column comes with them: a task_measured that does not name its task records
+nothing usable.
+
+Revision ID: 008_loop_events
+Revises: 007_task_type
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '008_loop_events'
+down_revision: Union[str, None] = '007_task_type'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+OLD_TYPES = ('start', 'cycle_completed', 'line_stop', 'resume', 'close')
+NEW_TYPES = (
+    'start', 'cycle_completed', 'task_measured', 'remeasure',
+    'line_stop', 'resume', 'close',
+)
+
+
+def _alter(values: Sequence[str]) -> None:
+    """Restate the column with a new set of allowed values.
+
+    On MySQL this is the ALTER that rewrites the enum. On SQLite, where an enum is
+    stored as a plain VARCHAR, batch mode rebuilds the table instead of emitting an
+    ALTER the engine does not support - the tests run there.
+    """
+    with op.batch_alter_table('production_events') as batch:
+        batch.alter_column(
+            'event_type',
+            existing_type=sa.Enum(*OLD_TYPES, name='production_event_type_enum'),
+            type_=sa.Enum(*values, name='production_event_type_enum'),
+            existing_nullable=False,
+        )
+
+
+def upgrade() -> None:
+    _alter(NEW_TYPES)
+    with op.batch_alter_table('production_events') as batch:
+        batch.add_column(sa.Column('task_id', sa.Integer(), nullable=True))
+        batch.create_foreign_key(
+            'fk_production_events_task_id', 'recipe_tasks', ['task_id'], ['id'],
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('production_events') as batch:
+        batch.drop_constraint('fk_production_events_task_id', type_='foreignkey')
+        batch.drop_column('task_id')
+    # The new kinds of event have no equivalent in the old vocabulary, and a row
+    # left with a value the column no longer accepts would fail the rebuild.
+    op.execute(
+        "DELETE FROM production_events "
+        "WHERE event_type IN ('task_measured', 'remeasure')"
+    )
+    _alter(OLD_TYPES)

src/backend/migrations/versions/009_recipe_measurement_rules.py

@@ -0,0 +1,57 @@
+"""traceability and manual input become rules of the recipe
+
+Three settings that were nowhere: whether the lot and the serial are compulsory,
+and whether a value may be typed instead of read from the caliper. Until now lot
+and serial were optional everywhere and the keypad was always available, so a
+value that happens to be in tolerance could simply be entered by hand.
+
+Backfill: the two traceability flags start false, which is exactly today's
+behaviour. Manual input starts *true* on the recipes that already exist - the
+column default is false, so recipes written from now on are caliper-only, but
+flipping the ones already in use would stop a running line at the next
+measurement. Turning them off is a decision for whoever owns the recipe, taken in
+the editor, not a side effect of an upgrade.
+
+Revision ID: 009_recipe_rules
+Revises: 008_loop_events
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '009_recipe_rules'
+down_revision: Union[str, None] = '008_loop_events'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        'recipes',
+        sa.Column(
+            'requires_lot', sa.Boolean(), nullable=False, server_default='0',
+        ),
+    )
+    op.add_column(
+        'recipes',
+        sa.Column(
+            'requires_serial', sa.Boolean(), nullable=False, server_default='0',
+        ),
+    )
+    op.add_column(
+        'recipes',
+        sa.Column(
+            'allow_manual_input', sa.Boolean(), nullable=False, server_default='0',
+        ),
+    )
+    # Recipes already in production keep the behaviour they were written under.
+    op.execute("UPDATE recipes SET allow_manual_input = 1")
+
+
+def downgrade() -> None:
+    op.drop_column('recipes', 'allow_manual_input')
+    op.drop_column('recipes', 'requires_serial')
+    op.drop_column('recipes', 'requires_lot')

src/backend/migrations/versions/010_measurement_authorisation.py

@@ -0,0 +1,49 @@
+"""record who authorised an out-of-tolerance measurement
+
+The supervisor gate existed on screen and nowhere else: the modal opened, the
+credentials were checked, the modal closed, and nothing was written down. Nothing
+depended on the answer either, so a value outside tolerance could be left behind
+by dismissing the modal.
+
+These two columns are what the rule now stands on. A failed measurement with no
+supervisor on it is a measurement waiting for one, and while one is waiting the
+operator cannot move to the next quote.
+
+Existing rows are left null. Backfilling an authorisation that never happened
+would be inventing an audit record; the fails already in the database belong to
+production runs that are over, and nothing is waiting on them.
+
+Revision ID: 010_meas_authorisation
+Revises: 009_recipe_rules
+Create Date: 2026-07-28
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '010_meas_authorisation'
+down_revision: Union[str, None] = '009_recipe_rules'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    with op.batch_alter_table('measurements') as batch:
+        batch.add_column(sa.Column('supervisor_id', sa.Integer(), nullable=True))
+        batch.add_column(sa.Column('authorised_at', sa.DateTime(), nullable=True))
+        batch.create_foreign_key(
+            'fk_measurements_supervisor_id', 'users', ['supervisor_id'], ['id'],
+        )
+        batch.create_index(
+            'ix_measurements_supervisor_id', ['supervisor_id'],
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('measurements') as batch:
+        batch.drop_index('ix_measurements_supervisor_id')
+        batch.drop_constraint('fk_measurements_supervisor_id', type_='foreignkey')
+        batch.drop_column('authorised_at')
+        batch.drop_column('supervisor_id')

src/backend/migrations/versions/011_vision_graph.py

@@ -0,0 +1,55 @@
+"""the vision graph lives on the task, and the quote names its output
+
+`xf_compare` was a typo for `dxf_compare`. No row uses either value yet, so
+renaming costs one statement here; from the first saved task onwards it would
+be a data migration and the typo would already be in the public API.
+
+Revision ID: 011_vision_graph
+Revises: 010_meas_authorisation
+Create Date: 2026-08-16
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '011_vision_graph'
+down_revision: Union[str, None] = '010_meas_authorisation'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+OLD_TYPES = ('note', 'measure', 'drawing', 'xf_compare', 'camera_measure')
+NEW_TYPES = ('note', 'measure', 'drawing', 'dxf_compare', 'camera_measure')
+
+
+def upgrade() -> None:
+    with op.batch_alter_table('recipe_tasks') as batch:
+        batch.add_column(sa.Column('vision_json', sa.JSON(), nullable=True))
+        batch.alter_column(
+            'task_type',
+            existing_type=sa.Enum(*OLD_TYPES, name='task_type_enum'),
+            type_=sa.Enum(*NEW_TYPES, name='task_type_enum'),
+            existing_nullable=False,
+            existing_server_default='note',
+        )
+
+    with op.batch_alter_table('recipe_subtasks') as batch:
+        batch.add_column(
+            sa.Column('vision_output', sa.String(length=120), nullable=True)
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('recipe_subtasks') as batch:
+        batch.drop_column('vision_output')
+
+    with op.batch_alter_table('recipe_tasks') as batch:
+        batch.alter_column(
+            'task_type',
+            existing_type=sa.Enum(*NEW_TYPES, name='task_type_enum'),
+            type_=sa.Enum(*OLD_TYPES, name='task_type_enum'),
+            existing_nullable=False,
+            existing_server_default='note',
+        )
+        batch.drop_column('vision_json')

src/backend/migrations/versions/012_vision_results.py

@@ -0,0 +1,79 @@
+"""one row per vision execution, and camera as an input method
+
+The vision surroundings do not go on `measurements`: statistics and the export
+read that table on every pass. And it would be wrong modelling anyway - one
+acquisition produces N quotes, and image, overlay, device and engine version are
+the same for all of them.
+
+Revision ID: 012_vision_results
+Revises: 011_vision_graph
+Create Date: 2026-08-16
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '012_vision_results'
+down_revision: Union[str, None] = '011_vision_graph'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+OLD_METHODS = ('usb_caliper', 'manual')
+NEW_METHODS = ('usb_caliper', 'manual', 'camera')
+
+
+def upgrade() -> None:
+    op.create_table(
+        'vision_results',
+        sa.Column('id', sa.Integer(), primary_key=True, autoincrement=True),
+        sa.Column('task_id', sa.Integer(),
+                  sa.ForeignKey('recipe_tasks.id'), nullable=False, index=True),
+        sa.Column('image_path', sa.String(length=500), nullable=True),
+        sa.Column('overlay_path', sa.String(length=500), nullable=True),
+        sa.Column('engine_version', sa.String(length=64), nullable=False),
+        sa.Column('executed_on',
+                  sa.Enum('server', 'station', name='vision_executed_on_enum'),
+                  nullable=False),
+        sa.Column('station_id', sa.Integer(),
+                  sa.ForeignKey('stations.id'), nullable=True),
+        sa.Column('device_code', sa.String(length=100), nullable=True),
+        sa.Column('calibration_snapshot', sa.JSON(), nullable=True),
+        sa.Column('graph_snapshot', sa.JSON(), nullable=False),
+        sa.Column('duration_ms', sa.Integer(), nullable=True),
+        sa.Column('executed_at', sa.DateTime(), nullable=False,
+                  server_default=sa.func.now()),
+        mysql_engine='InnoDB',
+        mysql_charset='utf8mb4',
+    )
+
+    with op.batch_alter_table('measurements') as batch:
+        batch.alter_column(
+            'input_method',
+            existing_type=sa.Enum(*OLD_METHODS, name='input_method_enum'),
+            type_=sa.Enum(*NEW_METHODS, name='input_method_enum'),
+            existing_nullable=False,
+        )
+        batch.add_column(
+            sa.Column('vision_result_id', sa.Integer(), nullable=True)
+        )
+        batch.create_foreign_key(
+            'fk_measurements_vision_result',
+            'vision_results', ['vision_result_id'], ['id'],
+        )
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('measurements') as batch:
+        batch.drop_constraint('fk_measurements_vision_result',
+                              type_='foreignkey')
+        batch.drop_column('vision_result_id')
+        batch.alter_column(
+            'input_method',
+            existing_type=sa.Enum(*NEW_METHODS, name='input_method_enum'),
+            type_=sa.Enum(*OLD_METHODS, name='input_method_enum'),
+            existing_nullable=False,
+        )
+
+    op.drop_table('vision_results')

src/backend/migrations/versions/013_reference_images.py

@@ -0,0 +1,49 @@
+"""reference images, with their provenance
+
+A reference image without provenance is a trap: the graph is re-run months later,
+different numbers come out, and there is no way to tell an updated engine from a
+moved lens.
+
+`expected_json` holds what the graph is supposed to produce on this image. It is
+what makes a regression run possible after a VisionSuite upgrade - the safety net
+that makes bumping the submodule sustainable with an engine on two hosts.
+
+Revision ID: 013_reference_images
+Revises: 012_vision_results
+Create Date: 2026-08-16
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+revision: str = '013_reference_images'
+down_revision: Union[str, None] = '012_vision_results'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        'vision_reference_images',
+        sa.Column('id', sa.Integer(), primary_key=True, autoincrement=True),
+        sa.Column('task_id', sa.Integer(),
+                  sa.ForeignKey('recipe_tasks.id'), nullable=False, index=True),
+        sa.Column('path', sa.String(length=500), nullable=False),
+        sa.Column('station_id', sa.Integer(),
+                  sa.ForeignKey('stations.id'), nullable=True),
+        sa.Column('device_code', sa.String(length=100), nullable=True),
+        sa.Column('calibration_snapshot', sa.JSON(), nullable=True),
+        sa.Column('engine_version', sa.String(length=64), nullable=True),
+        sa.Column('note', sa.String(length=500), nullable=True),
+        sa.Column('expected_json', sa.JSON(), nullable=True),
+        sa.Column('acquired_at', sa.DateTime(), nullable=False,
+                  server_default=sa.func.now()),
+        mysql_engine='InnoDB',
+        mysql_charset='utf8mb4',
+    )
+
+
+def downgrade() -> None:
+    op.drop_table('vision_reference_images')

src/backend/models/api/measurement.py

@@ -12,7 +12,15 @@ class MeasurementCreate(BaseModel):
     value: float
     lot_number: Optional[str] = Field(None, max_length=100)
     serial_number: Optional[str] = Field(None, max_length=100)
+    # "camera" is deliberately absent from this pattern, even though the ORM
+    # enum and the database allow it. Do not add it here: a camera
+    # measurement must be produced by vision_service.execute_task, never by a
+    # client posting JSON straight to this endpoint - there is no image, no
+    # graph and no engine_version behind a value that arrived this way.
     input_method: str = Field("manual", pattern="^(usb_caliper|manual)$")
+    input_duration_ms: Optional[int] = Field(None, ge=0)
+    # The production this belongs to, when one is open at the station.
+    production_run_id: Optional[int] = Field(None, gt=0)
 
 
 class MeasurementBatchCreate(BaseModel):
@@ -34,6 +42,12 @@ class MeasurementResponse(BaseModel):
     lot_number: Optional[str] = None
     serial_number: Optional[str] = None
     input_method: str
+    input_duration_ms: Optional[int] = None
+    production_run_id: Optional[int] = None
+    # Who let this value stand, when it was out of tolerance. Null on a value in
+    # tolerance - and on one that is not and is still waiting for an answer.
+    supervisor_id: Optional[int] = None
+    authorised_at: Optional[datetime] = None
     measured_at: datetime
     synced_to_csv: bool
 
@@ -47,6 +61,29 @@ class MeasurementListResponse(BaseModel):
     pages: int
 
 
+class MeasurementAuthorisation(BaseModel):
+    """Credentials of the supervisor allowing an out-of-tolerance value to stand."""
+
+    supervisor_username: str = Field(..., min_length=1)
+    supervisor_password: str = Field(..., min_length=1)
+
+
+class TaskProgressResponse(BaseModel):
+    """How far one measurement task has got: quotes taken out of quotes expected."""
+
+    task_id: int
+    quotes: int
+    measured: int
+    # empty (declared as measurement, no quotes yet) | none | partial | complete
+    state: str
+
+
+class TaskProgressListResponse(BaseModel):
+    """Progress of every measurement task of a version."""
+
+    tasks: list[TaskProgressResponse]
+
+
 class MeasurementQuery(BaseModel):
     """Schema for measurement query filters."""
     recipe_id: Optional[int] = None

src/backend/models/api/production.py

@@ -0,0 +1,89 @@
+"""Pydantic schemas for production runs and their events."""
+from datetime import datetime
+from typing import Literal, Optional
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class ProductionRunCreate(BaseModel):
+    station_code: str = Field(..., min_length=1, max_length=100)
+    recipe_id: int = Field(..., gt=0)
+    version_id: Optional[int] = Field(default=None, gt=0)
+    lot_number: Optional[str] = Field(default=None, max_length=100)
+    serial_number: Optional[str] = Field(default=None, max_length=100)
+
+
+class ProductionEventResponse(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+    id: int
+    event_type: str
+    user_id: int
+    supervisor_id: Optional[int]
+    task_id: Optional[int] = None
+    note: Optional[str]
+    created_at: datetime
+
+
+class ProductionRunResponse(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
+    id: int
+    station_id: int
+    recipe_id: int
+    version_id: int
+    operator_id: int
+    lot_number: Optional[str]
+    serial_number: Optional[str]
+    status: str
+    measurement_interval_minutes: Optional[int]
+    next_measurement_at: Optional[datetime]
+    cycle_count: int
+    started_at: datetime
+    paused_at: Optional[datetime]
+    closed_at: Optional[datetime]
+    closed_by: Optional[int]
+    statistics_path: Optional[str] = None
+
+    # Derived server-side so every client agrees on the countdown regardless of
+    # clock skew. Negative once the interval has elapsed: how long the run has been
+    # overdue is a fact the operator must see, not deduce.
+    seconds_to_next_measurement: Optional[int] = None
+    overdue: bool = False
+    server_time: Optional[datetime] = None
+
+    # The measurement tasks of this run's version, in execution order. The client
+    # needs them to bring the operator back to the measurement when the interval
+    # expires, whichever screen they are on, and to know which task closes a cycle.
+    measurement_task_ids: list[int] = Field(default_factory=list)
+
+
+class ProductionRunWithEventsResponse(ProductionRunResponse):
+    events: list[ProductionEventResponse] = Field(default_factory=list)
+
+
+class SupervisorAction(BaseModel):
+    """Credentials of the supervisor authorising a line stop or a close."""
+
+    supervisor_username: str = Field(..., min_length=1)
+    supervisor_password: str = Field(..., min_length=1)
+    note: Optional[str] = None
+
+
+class CycleCompletePayload(BaseModel):
+    # Which measurement task was just finished. The interval restarts only on the
+    # last one of the recipe; omitting it closes the cycle outright.
+    task_id: Optional[int] = Field(default=None, gt=0)
+    note: Optional[str] = None
+
+
+class RemeasurePayload(BaseModel):
+    """The piece was turned over: measure it again inside the same cycle."""
+
+    task_id: Optional[int] = Field(default=None, gt=0)
+    note: Optional[str] = None
+
+
+ProductionEventType = Literal[
+    "start", "cycle_completed", "task_measured", "remeasure",
+    "line_stop", "resume", "close",
+]

src/backend/models/api/recipe.py

@@ -14,6 +14,12 @@ class RecipeCreate(BaseModel):
     name: str = Field(..., min_length=1, max_length=255)
     description: Optional[str] = None
     image_path: Optional[str] = Field(None, max_length=500)
+    measurement_interval_minutes: Optional[int] = Field(None, ge=1, le=1440)
+    # Rules of the recipe: what the operator must supply, and how a value may be
+    # entered. Manual input defaults to forbidden - the caliper is the instrument.
+    requires_lot: bool = False
+    requires_serial: bool = False
+    allow_manual_input: bool = False
     # Optional task-level fields for the initial technical drawing
     file_path: Optional[str] = Field(None, max_length=500)
     file_type: Optional[str] = Field(None, pattern="^(image|pdf)$")
@@ -25,6 +31,10 @@ class RecipeUpdate(BaseModel):
     name: Optional[str] = Field(None, min_length=1, max_length=255)
     description: Optional[str] = None
     image_path: Optional[str] = Field(None, max_length=500)
+    measurement_interval_minutes: Optional[int] = Field(None, ge=1, le=1440)
+    requires_lot: Optional[bool] = None
+    requires_serial: Optional[bool] = None
+    allow_manual_input: Optional[bool] = None
     change_notes: Optional[str] = None
     # Task-level fields: saved to the first task of the new version
     file_path: Optional[str] = Field(None, max_length=500)
@@ -55,6 +65,12 @@ class RecipeResponse(BaseModel):
     name: str
     description: Optional[str] = None
     image_path: Optional[str] = None
+    measurement_interval_minutes: Optional[int] = None
+    # The operator's screen reads these to know whether to ask for lot and serial
+    # before starting, and whether to offer the keypad at all.
+    requires_lot: bool = False
+    requires_serial: bool = False
+    allow_manual_input: bool = False
     created_by: int
     created_at: datetime
     active: bool

src/backend/models/api/station.py

@@ -51,7 +51,20 @@ class RecipeSummary(BaseModel):
     code: str
     name: str
     active: bool
+    image_path: Optional[str] = None
+    description: Optional[str] = None
+    # The selection screen has to know before the operator starts: what a recipe
+    # demands is the difference between an enabled Avvia and a disabled one.
+    requires_lot: bool = False
+    requires_serial: bool = False
 
 
 class StationWithRecipesResponse(StationResponse):
     recipes: list[RecipeSummary] = Field(default_factory=list)
+
+
+class StationResetResponse(BaseModel):
+    """Outcome of clearing every recipe assignment of a station."""
+
+    station_id: int
+    removed: int

src/backend/models/api/task.py

@@ -1,8 +1,13 @@
 """Pydantic schemas for RecipeTask and RecipeSubtask operations."""
-from typing import Any, Optional
+from typing import Any, Literal, Optional
 
 from pydantic import BaseModel, ConfigDict, Field
 
+from src.backend.models.orm.task import TASK_TYPES
+
+# Kept in step with the ORM enum so the two cannot drift apart.
+TaskType = Literal[TASK_TYPES]
+
 
 class SubtaskCreate(BaseModel):
     """Schema for creating a subtask."""
@@ -16,6 +21,7 @@ class SubtaskCreate(BaseModel):
     ltl: Optional[float] = None
     unit: str = Field("mm", max_length=20)
     image_path: Optional[str] = Field(None, max_length=500)
+    vision_output: Optional[str] = None
 
 
 class SubtaskUpdate(BaseModel):
@@ -30,6 +36,7 @@ class SubtaskUpdate(BaseModel):
     ltl: Optional[float] = None
     unit: Optional[str] = Field(None, max_length=20)
     image_path: Optional[str] = Field(None, max_length=500)
+    vision_output: Optional[str] = None
 
 
 class SubtaskResponse(BaseModel):
@@ -48,27 +55,32 @@ class SubtaskResponse(BaseModel):
     ltl: Optional[float] = None
     unit: str
     image_path: Optional[str] = None
+    vision_output: Optional[str] = None
 
 
 class TaskCreate(BaseModel):
     """Schema for creating a task."""
     title: str = Field(..., min_length=1, max_length=255)
+    task_type: TaskType = "note"
     directive: Optional[str] = None
     description: Optional[str] = None
     file_path: Optional[str] = Field(None, max_length=500)
     file_type: Optional[str] = Field(None, pattern="^(image|pdf)$")
     annotations_json: Optional[dict[str, Any]] = None
+    vision_json: Optional[dict] = None
     subtasks: list[SubtaskCreate] = []
 
 
 class TaskUpdate(BaseModel):
     """Schema for updating a task."""
     title: Optional[str] = Field(None, min_length=1, max_length=255)
+    task_type: Optional[TaskType] = None
     directive: Optional[str] = None
     description: Optional[str] = None
     file_path: Optional[str] = Field(None, max_length=500)
     file_type: Optional[str] = Field(None, pattern="^(image|pdf)$")
     annotations_json: Optional[dict[str, Any]] = None
+    vision_json: Optional[dict] = None
 
 
 class TaskResponse(BaseModel):
@@ -79,12 +91,14 @@ class TaskResponse(BaseModel):
     version_id: int
     recipe_id: Optional[int] = None
     order_index: int
+    task_type: str
     title: str
     directive: Optional[str] = None
     description: Optional[str] = None
     file_path: Optional[str] = None
     file_type: Optional[str] = None
     annotations_json: Optional[dict[str, Any]] = None
+    vision_json: Optional[dict] = None
     subtasks: list[SubtaskResponse] = []
 
 

src/backend/models/api/vision.py

@@ -0,0 +1,16 @@
+"""Pydantic schemas for the vision execution response."""
+from pydantic import BaseModel
+
+
+class VisionMeasurementResult(BaseModel):
+    """One quote's outcome, as much of it as the client needs to show."""
+    subtask_id: int
+    value: float
+    pass_fail: str
+
+
+class VisionExecuteResponse(BaseModel):
+    """What `POST /api/vision/execute` returns: the run, and every quote it filled."""
+    vision_result_id: int
+    engine_version: str
+    measurements: list[VisionMeasurementResult]

src/backend/models/orm/__init__.py

@@ -6,6 +6,8 @@ from src.backend.models.orm.measurement import Measurement
 from src.backend.models.orm.access_log import AccessLog
 from src.backend.models.orm.setting import SystemSetting, RecipeVersionAudit
 from src.backend.models.orm.station import Station, StationRecipeAssignment
+from src.backend.models.orm.production import ProductionRun, ProductionEvent
+from src.backend.models.orm.vision import VisionReferenceImage, VisionResult
 
 __all__ = [
     "User",
@@ -19,4 +21,8 @@ __all__ = [
     "RecipeVersionAudit",
     "Station",
     "StationRecipeAssignment",
+    "ProductionRun",
+    "ProductionEvent",
+    "VisionResult",
+    "VisionReferenceImage",
 ]

src/backend/models/orm/measurement.py

@@ -38,11 +38,43 @@ class Measurement(Base):
 
     # Input method
     input_method: Mapped[str] = mapped_column(
-        Enum("usb_caliper", "manual", name="input_method_enum"),
+        Enum("usb_caliper", "manual", "camera", name="input_method_enum"),
         nullable=False,
         default="manual",
     )
 
+    # The vision execution that produced this value, when there was one. Null for
+    # everything taken by caliper or typed by hand.
+    vision_result_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("vision_results.id"), nullable=True
+    )
+
+    # Time spent by the operator entering this value (client-reported)
+    input_duration_ms: Mapped[Optional[int]] = mapped_column(
+        Integer, nullable=True
+    )
+
+    # The production this was taken during, when there was one. Nullable: rows
+    # predating production runs, and measurements taken outside a production, have
+    # none. This is what makes "the measurements of this production" a query, which
+    # fine produzione needs to emit the statistics file.
+    production_run_id: Mapped[Optional[int]] = mapped_column(
+        Integer,
+        ForeignKey("production_runs.id", ondelete="SET NULL"),
+        nullable=True,
+        index=True,
+    )
+
+    # Who let an out-of-tolerance value stand, and when. Null on a value that is
+    # within tolerance, and null on one that is not and has not been authorised -
+    # which is exactly the state that stops the operator moving on to the next
+    # quote. Before this the supervisor's approval was a modal that closed: nothing
+    # recorded it, and nothing depended on it.
+    supervisor_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("users.id"), nullable=True, index=True
+    )
+    authorised_at: Mapped[Optional[datetime]] = mapped_column(DateTime, nullable=True)
+
     # Timestamp
     measured_at: Mapped[datetime] = mapped_column(
         DateTime, nullable=False, server_default=func.now(), index=True

src/backend/models/orm/production.py

@@ -0,0 +1,166 @@
+"""ProductionRun and ProductionEvent models.
+
+A production run is what an operator opens when production starts and a supervisor
+closes when it ends. It is the thing that outlives a page: before this, the timer,
+the cycle count and the "production started" flag were Alpine variables in
+task_execute.html, and navigating between tasks - a full page load - destroyed them.
+
+The measurement deadline is stored as an absolute timestamp (next_measurement_at)
+rather than a remaining count. Any client can then derive the countdown from it, get
+the same answer, survive a reload, and tell how far *past* the interval it is by
+letting the difference go negative.
+
+With one station app installed per PC (see the install architecture of 28/07) the
+database is the only place this state can live: the API keeps nothing in process
+memory.
+"""
+from datetime import datetime
+from typing import TYPE_CHECKING, Optional
+
+from sqlalchemy import (
+    DateTime, Enum, ForeignKey, Index, Integer, SmallInteger, String, Text,
+    UniqueConstraint, func,
+)
+from sqlalchemy.orm import Mapped, mapped_column, relationship
+
+from src.backend.database import Base
+
+if TYPE_CHECKING:
+    from src.backend.models.orm.recipe import Recipe, RecipeVersion
+    from src.backend.models.orm.station import Station
+
+
+RUN_STATUSES = ("running", "paused", "closed")
+# task_measured marks a measurement task finished inside a cycle that is not over
+# yet; only the last one produces cycle_completed and restarts the interval.
+# remeasure is the piece being turned and measured again without closing the cycle.
+EVENT_TYPES = (
+    "start", "cycle_completed", "task_measured", "remeasure",
+    "line_stop", "resume", "close",
+)
+
+
+class ProductionRun(Base):
+    __tablename__ = "production_runs"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+
+    station_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("stations.id"), nullable=False, index=True
+    )
+    recipe_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("recipes.id"), nullable=False, index=True
+    )
+    version_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("recipe_versions.id"), nullable=False, index=True
+    )
+    operator_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("users.id"), nullable=False, index=True
+    )
+
+    # Traceability travels with the run instead of the Flask session, so it stays
+    # attached to every measurement taken during it.
+    lot_number: Mapped[Optional[str]] = mapped_column(String(100), nullable=True, index=True)
+    serial_number: Mapped[Optional[str]] = mapped_column(String(100), nullable=True, index=True)
+
+    status: Mapped[str] = mapped_column(
+        Enum(*RUN_STATUSES, name="production_run_status_enum"),
+        nullable=False,
+        default="running",
+        index=True,
+    )
+
+    # Snapshot of the recipe interval at start: editing the recipe mid-run must not
+    # move the deadline of a production already under way.
+    measurement_interval_minutes: Mapped[Optional[int]] = mapped_column(
+        SmallInteger, nullable=True
+    )
+    # When the next measurement falls due. Null when the recipe has no interval.
+    # Past this instant the run is overdue, and by how much is simply now - this.
+    next_measurement_at: Mapped[Optional[datetime]] = mapped_column(
+        DateTime, nullable=True, index=True
+    )
+    cycle_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+
+    started_at: Mapped[datetime] = mapped_column(
+        DateTime, nullable=False, server_default=func.now()
+    )
+    paused_at: Mapped[Optional[datetime]] = mapped_column(DateTime, nullable=True)
+    closed_at: Mapped[Optional[datetime]] = mapped_column(DateTime, nullable=True)
+    closed_by: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("users.id"), nullable=True
+    )
+    # Statistics file emitted when the run was closed, relative to the upload dir.
+    statistics_path: Mapped[Optional[str]] = mapped_column(String(500), nullable=True)
+
+    # Mirrors station_id while the run is open and goes NULL when it closes. A unique
+    # index on it lets the database - not a check-then-insert race - guarantee that a
+    # station never has two open runs. Repeated NULLs do not collide in a unique index.
+    active_station_id: Mapped[Optional[int]] = mapped_column(Integer, nullable=True)
+
+    station: Mapped["Station"] = relationship(lazy="selectin")
+    recipe: Mapped["Recipe"] = relationship(lazy="selectin")
+    version: Mapped["RecipeVersion"] = relationship(lazy="selectin")
+    events: Mapped[list["ProductionEvent"]] = relationship(
+        back_populates="run",
+        cascade="all, delete-orphan",
+        order_by="ProductionEvent.created_at",
+    )
+
+    __table_args__ = (
+        UniqueConstraint("active_station_id", name="uq_production_runs_active_station"),
+        Index("ix_production_runs_station_status", "station_id", "status"),
+        {"mysql_engine": "InnoDB", "mysql_charset": "utf8mb4"},
+    )
+
+    @property
+    def is_open(self) -> bool:
+        return self.status in ("running", "paused")
+
+    def __repr__(self) -> str:
+        return f"<ProductionRun {self.id} station={self.station_id} {self.status}>"
+
+
+class ProductionEvent(Base):
+    """Append-only trace of what happened during a run.
+
+    This is the record that did not exist before: without it there is no history of
+    a production, and nothing for the statistics file or the ERP hand-off to stand on.
+    """
+
+    __tablename__ = "production_events"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+    run_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("production_runs.id", ondelete="CASCADE"),
+        nullable=False, index=True,
+    )
+    event_type: Mapped[str] = mapped_column(
+        Enum(*EVENT_TYPES, name="production_event_type_enum"), nullable=False, index=True
+    )
+    user_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("users.id"), nullable=False, index=True
+    )
+    # Who authorised it, when the action needed a supervisor (fermo linea, chiusura).
+    supervisor_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("users.id"), nullable=True
+    )
+    # Which task the event is about, for the ones that happen inside a cycle. A
+    # task_measured or a remeasure that does not say which task is not a record of
+    # anything; null for the events that concern the run as a whole.
+    task_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("recipe_tasks.id"), nullable=True
+    )
+    note: Mapped[Optional[str]] = mapped_column(Text, nullable=True)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime, nullable=False, server_default=func.now(), index=True
+    )
+
+    run: Mapped["ProductionRun"] = relationship(back_populates="events")
+
+    __table_args__ = (
+        {"mysql_engine": "InnoDB", "mysql_charset": "utf8mb4"},
+    )
+
+    def __repr__(self) -> str:
+        return f"<ProductionEvent run={self.run_id} {self.event_type}>"

src/backend/models/orm/recipe.py

@@ -2,7 +2,7 @@
 from datetime import datetime
 from typing import TYPE_CHECKING, Optional
 
-from sqlalchemy import Boolean, DateTime, ForeignKey, Index, Integer, String, Text, UniqueConstraint, func
+from sqlalchemy import Boolean, DateTime, ForeignKey, Index, Integer, SmallInteger, String, Text, UniqueConstraint, func
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from src.backend.database import Base
@@ -24,6 +24,23 @@ class Recipe(Base):
         DateTime, nullable=False, server_default=func.now()
     )
     active: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True, index=True)
+    measurement_interval_minutes: Mapped[Optional[int]] = mapped_column(SmallInteger, nullable=True)
+
+    # Traceability, decided when the recipe is written rather than left to whoever
+    # is at the machine: a measurement without the lot it belongs to cannot be
+    # traced back afterwards, and tracing back is half of what the system is for.
+    requires_lot: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=False, server_default="0"
+    )
+    requires_serial: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=False, server_default="0"
+    )
+    # Whether a value may be typed instead of read from the caliper. Forbidden by
+    # default: without the rule, a value that happens to be in tolerance can simply
+    # be entered by hand, which is the case this setting exists to prevent.
+    allow_manual_input: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=False, server_default="0"
+    )
 
     # Relationships
     versions: Mapped[list["RecipeVersion"]] = relationship(

src/backend/models/orm/task.py

@@ -12,6 +12,15 @@ if TYPE_CHECKING:
     from src.backend.models.orm.recipe import RecipeVersion
 
 
+# What a task is, declared by whoever writes the recipe instead of guessed from its
+# contents. The last two are not built yet; they are in the enum from the start
+# because widening a MySQL enum later is an ALTER on a live table.
+TASK_TYPES = ("note", "measure", "drawing", "dxf_compare", "camera_measure")
+
+# Types the operator screen treats as "there are quotes to take here".
+MEASURING_TASK_TYPES = ("measure", "camera_measure")
+
+
 class RecipeTask(Base):
     __tablename__ = "recipe_tasks"
 
@@ -20,6 +29,16 @@ class RecipeTask(Base):
         Integer, ForeignKey("recipe_versions.id"), nullable=False
     )
     order_index: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    # Declared, not deduced. The old rule - "has subtasks, therefore a measurement" -
+    # meant a measurement task whose quotes had not been entered yet behaved as a
+    # note, so the system acted differently depending on how finished the recipe was.
+    task_type: Mapped[str] = mapped_column(
+        Enum(*TASK_TYPES, name="task_type_enum"),
+        nullable=False,
+        default="note",
+        server_default="note",
+        index=True,
+    )
     title: Mapped[str] = mapped_column(String(255), nullable=False)
     directive: Mapped[Optional[str]] = mapped_column(Text, nullable=True)
     description: Mapped[Optional[str]] = mapped_column(Text, nullable=True)
@@ -28,6 +47,10 @@ class RecipeTask(Base):
         Enum("image", "pdf", name="file_type_enum"), nullable=True
     )
     annotations_json: Mapped[Optional[dict]] = mapped_column(JSON, nullable=True)
+    # The vision graph, serialised by vs-task. Opaque here on purpose: the
+    # server routes it and never interprets it, which is what keeps VisionSuite
+    # out of this container.
+    vision_json: Mapped[Optional[dict]] = mapped_column(JSON, nullable=True)
 
     # Relationships
     version: Mapped["RecipeVersion"] = relationship(back_populates="tasks")
@@ -70,6 +93,12 @@ class RecipeSubtask(Base):
 
     unit: Mapped[str] = mapped_column(String(20), nullable=False, default="mm")
     image_path: Mapped[Optional[str]] = mapped_column(String(500), nullable=True)
+    # Which output of the vision graph fills this quote: "<tool_id>.<output>".
+    # A link is an id, not a pointer - the same rule vs-task follows, and for
+    # the same reason: a pointer does not survive being written to disk.
+    vision_output: Mapped[Optional[str]] = mapped_column(
+        String(120), nullable=True
+    )
 
     # Relationships
     task: Mapped["RecipeTask"] = relationship(back_populates="subtasks")

src/backend/models/orm/vision.py

@@ -0,0 +1,69 @@
+"""One row per vision execution - not per quote."""
+from datetime import datetime
+from typing import Optional
+
+from sqlalchemy import DateTime, Enum, ForeignKey, Integer, JSON, String, func
+from sqlalchemy.orm import Mapped, mapped_column
+
+from src.backend.database import Base
+
+
+class VisionResult(Base):
+    __tablename__ = "vision_results"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+    task_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("recipe_tasks.id"), nullable=False, index=True
+    )
+    image_path: Mapped[Optional[str]] = mapped_column(String(500), nullable=True)
+    overlay_path: Mapped[Optional[str]] = mapped_column(String(500), nullable=True)
+    engine_version: Mapped[str] = mapped_column(String(64), nullable=False)
+    executed_on: Mapped[str] = mapped_column(
+        Enum("server", "station", name="vision_executed_on_enum"), nullable=False
+    )
+    station_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("stations.id"), nullable=True
+    )
+    device_code: Mapped[Optional[str]] = mapped_column(String(100), nullable=True)
+    # Copied, not referenced: a calibration redone six months from now must not
+    # rewrite what a measurement already taken means. Same for the graph.
+    calibration_snapshot: Mapped[Optional[dict]] = mapped_column(JSON, nullable=True)
+    graph_snapshot: Mapped[dict] = mapped_column(JSON, nullable=False)
+    duration_ms: Mapped[Optional[int]] = mapped_column(Integer, nullable=True)
+    executed_at: Mapped[datetime] = mapped_column(
+        DateTime, nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = ({"mysql_engine": "InnoDB", "mysql_charset": "utf8mb4"},)
+
+    def __repr__(self) -> str:
+        return f"<VisionResult task={self.task_id} engine={self.engine_version}>"
+
+
+class VisionReferenceImage(Base):
+    """An image kept so a graph can be composed, tried and re-tried without a
+    camera - which is exactly what VisionSuite's source abstraction is for."""
+
+    __tablename__ = "vision_reference_images"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+    task_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("recipe_tasks.id"), nullable=False, index=True
+    )
+    path: Mapped[str] = mapped_column(String(500), nullable=False)
+    station_id: Mapped[Optional[int]] = mapped_column(
+        Integer, ForeignKey("stations.id"), nullable=True
+    )
+    device_code: Mapped[Optional[str]] = mapped_column(String(100), nullable=True)
+    calibration_snapshot: Mapped[Optional[dict]] = mapped_column(JSON, nullable=True)
+    engine_version: Mapped[Optional[str]] = mapped_column(String(64), nullable=True)
+    note: Mapped[Optional[str]] = mapped_column(String(500), nullable=True)
+    expected_json: Mapped[Optional[dict]] = mapped_column(JSON, nullable=True)
+    acquired_at: Mapped[datetime] = mapped_column(
+        DateTime, nullable=False, server_default=func.now()
+    )
+
+    __table_args__ = ({"mysql_engine": "InnoDB", "mysql_charset": "utf8mb4"},)
+
+    def __repr__(self) -> str:
+        return f"<VisionReferenceImage task={self.task_id} path={self.path}>"

src/backend/services/ai_service.py

@@ -0,0 +1,74 @@
+"""AI service for parsing technical sheets via OpenRouter."""
+import json
+import logging
+from io import BytesIO
+
+import httpx
+import pdfplumber
+
+from src.backend.config import settings
+
+logger = logging.getLogger(__name__)
+
+SYSTEM_PROMPT = """Sei un assistente specializzato nell'analisi di schede tecniche industriali.
+Analizza il testo estratto da una scheda tecnica e identifica i blocchi operativi distinti.
+Per ogni blocco, estrai:
+- title: titolo breve del blocco (es. "MATERIALI", "TEMPERATURA", "MISURA", "IMBALLO")
+- directive: istruzione operativa principale (es. "VERIFICARE ATTREZZATURE")
+- description: dettagli completi del blocco, preservando gli "a capo" originali
+
+Rispondi SOLO con un array JSON valido, senza markdown o testo aggiuntivo.
+Esempio:
+[
+  {"title": "MATERIALI", "directive": "Verificare materiali", "description": "RIGIDO (36): AFT9/UV CRI 7\\nMORBIDO (E38): M70"},
+  {"title": "TEMPERATURA", "directive": "Verificare impostazioni", "description": "RIGIDO: TEMP: 170/170/170/170/170\\nMORBIDO: TEMP: 130-135-140-145"}
+]"""
+
+
+def extract_text_from_pdf(pdf_bytes: bytes) -> str:
+    with pdfplumber.open(BytesIO(pdf_bytes)) as pdf:
+        pages = []
+        for page in pdf.pages:
+            text = page.extract_text()
+            if text:
+                pages.append(text)
+        return "\n\n---\n\n".join(pages)
+
+
+async def parse_technical_sheet(pdf_bytes: bytes) -> list[dict]:
+    text = extract_text_from_pdf(pdf_bytes)
+    if not text.strip():
+        return []
+
+    if not settings.openrouter_api_key:
+        raise ValueError("OPENROUTER_API_KEY non configurata")
+
+    async with httpx.AsyncClient(timeout=60) as client:
+        resp = await client.post(
+            "https://openrouter.ai/api/v1/chat/completions",
+            headers={
+                "Authorization": f"Bearer {settings.openrouter_api_key}",
+                "Content-Type": "application/json",
+            },
+            json={
+                "model": settings.openrouter_model,
+                "messages": [
+                    {"role": "system", "content": SYSTEM_PROMPT},
+                    {"role": "user", "content": f"Analizza questa scheda tecnica ed estrai i task:\n\n{text}"},
+                ],
+                "temperature": 0.1,
+            },
+        )
+        resp.raise_for_status()
+
+    data = resp.json()
+    content = data["choices"][0]["message"]["content"].strip()
+
+    # Strip markdown fences if present
+    if content.startswith("```"):
+        content = content.split("\n", 1)[1]
+        if content.endswith("```"):
+            content = content[:-3]
+        content = content.strip()
+
+    return json.loads(content)

src/backend/services/auth_service.py

@@ -3,6 +3,7 @@ import secrets
 from datetime import datetime
 
 import bcrypt
+from fastapi import HTTPException, status
 from sqlalchemy import select, update
 from sqlalchemy.ext.asyncio import AsyncSession
 
@@ -39,6 +40,31 @@ async def authenticate_user(
     return user
 
 
+async def authorise_supervisor(db: AsyncSession, username: str, password: str) -> User:
+    """Check the credentials of the supervisor authorising something.
+
+    Used wherever the shop floor needs a capoturno: a line stop, a closure, an
+    out-of-tolerance value allowed to stand. Lives here rather than beside any one
+    of them because it is the same question every time - who is this, and are they
+    allowed to say yes.
+
+    Uses authenticate_user rather than a full login: logging in would rotate the
+    supervisor's API key and knock out whatever session they have open elsewhere.
+    """
+    supervisor = await authenticate_user(db, username, password)
+    if supervisor is None:
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid supervisor credentials",
+        )
+    if not supervisor.has_role("Supervisor") and not supervisor.is_admin:
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="User is not authorised as supervisor",
+        )
+    return supervisor
+
+
 async def login_user(db: AsyncSession, user: User) -> str:
     """Generate API key and update last_login for user."""
     api_key = generate_api_key()

src/backend/services/measurement_service.py

@@ -1,11 +1,17 @@
 """Measurement service - pass/fail calculation, data storage."""
+from datetime import datetime
 from decimal import Decimal
 
-from sqlalchemy import select
+from fastapi import HTTPException, status
+from sqlalchemy import func, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from src.backend.models.orm.measurement import Measurement
-from src.backend.models.orm.task import RecipeSubtask
+from src.backend.models.orm.recipe import Recipe, RecipeVersion
+from src.backend.models.orm.task import (
+    MEASURING_TASK_TYPES, RecipeSubtask, RecipeTask,
+)
+from src.backend.models.orm.user import User
 
 
 def calculate_pass_fail(
@@ -39,6 +45,174 @@ def calculate_pass_fail(
     return "pass", deviation
 
 
+async def recipe_of_version(db: AsyncSession, version_id: int) -> Recipe | None:
+    """The recipe a version belongs to - where the rules of measurement live."""
+    result = await db.execute(
+        select(Recipe)
+        .join(RecipeVersion, RecipeVersion.recipe_id == Recipe.id)
+        .where(RecipeVersion.id == version_id)
+    )
+    return result.scalar_one_or_none()
+
+
+def _enforce_recipe_rules(
+    recipe: Recipe | None,
+    lot_number: str | None,
+    serial_number: str | None,
+    input_method: str,
+) -> None:
+    """Refuse a measurement the recipe does not allow.
+
+    Checked here rather than only on the screen because a rule that lives in the
+    frontend is a suggestion: the keypad can be hidden and the same request still
+    sent. This is the one place every measurement passes through.
+    """
+    if recipe is None:
+        return
+
+    if recipe.requires_lot and not (lot_number or "").strip():
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="This recipe requires a lot number",
+        )
+    if recipe.requires_serial and not (serial_number or "").strip():
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="This recipe requires a serial number",
+        )
+    if input_method == "manual" and not recipe.allow_manual_input:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="This recipe does not allow values typed by hand: use the caliper",
+        )
+
+
+def _scope(query, version_id: int, measured_by: int, production_run_id: int | None):
+    """Narrow a measurement query to the run in progress, or to this operator's work.
+
+    Inside a production the run is the boundary. Outside one - a recipe executed on
+    its own - the boundary is the version and the person doing it, so two operators
+    on the same recipe do not block each other.
+    """
+    if production_run_id is not None:
+        return query.where(Measurement.production_run_id == production_run_id)
+    return query.where(
+        Measurement.version_id == version_id,
+        Measurement.measured_by == measured_by,
+        Measurement.production_run_id.is_(None),
+    )
+
+
+async def pending_authorisation(
+    db: AsyncSession,
+    version_id: int,
+    measured_by: int,
+    production_run_id: int | None = None,
+) -> Measurement | None:
+    """The out-of-tolerance measurement that is holding everything up, if any.
+
+    Only the *latest* reading of each quote counts. Measuring the same quote again
+    is allowed - the caliper slips, the piece is reseated - and a second reading
+    within tolerance releases the block. What is not allowed is moving on to the
+    next quote while the one in hand is out of tolerance and nobody has said so.
+
+    How many attempts that permits is point 4's business, not this one's. Every
+    reading stays in the record either way.
+    """
+    latest_per_subtask = _scope(
+        select(func.max(Measurement.id)),
+        version_id, measured_by, production_run_id,
+    ).group_by(Measurement.subtask_id)
+
+    result = await db.execute(
+        select(Measurement)
+        .where(
+            Measurement.id.in_(latest_per_subtask),
+            Measurement.pass_fail == "fail",
+            Measurement.supervisor_id.is_(None),
+        )
+        .order_by(Measurement.id)
+        .limit(1)
+    )
+    return result.scalar_one_or_none()
+
+
+async def task_progress(
+    db: AsyncSession,
+    version_id: int,
+    measured_by: int,
+    production_run_id: int | None = None,
+) -> list[dict]:
+    """How far each measurement task of a version has got.
+
+    A task left half done is the thing the operator cannot see today: they walk
+    away from it, come back to the list, and it looks exactly like one never
+    started. Counting quotes against measurements is what tells them apart.
+
+    Documental tasks are not in here. A note cannot be half read, and marking one
+    incomplete for ever would be noise on the only screen that has to stay legible.
+    Which tasks count is read from the declared type, not from whether quotes happen
+    to have been entered yet - the same rule point 2 settled everywhere else.
+    """
+    rows = (await db.execute(
+        select(RecipeTask.id, RecipeSubtask.id)
+        .join(RecipeSubtask, RecipeSubtask.task_id == RecipeTask.id, isouter=True)
+        .where(
+            RecipeTask.version_id == version_id,
+            RecipeTask.task_type.in_(MEASURING_TASK_TYPES),
+        )
+    )).all()
+    if not rows:
+        return []
+
+    by_task: dict[int, set[int]] = {}
+    for task_id, subtask_id in rows:
+        quotes = by_task.setdefault(task_id, set())
+        if subtask_id is not None:
+            quotes.add(subtask_id)
+
+    measured = set((await db.execute(
+        _scope(
+            select(Measurement.subtask_id).distinct(),
+            version_id, measured_by, production_run_id,
+        )
+    )).scalars().all())
+
+    progress = []
+    for task_id, subtask_ids in by_task.items():
+        total = len(subtask_ids)
+        done = len(subtask_ids & measured)
+        if total == 0:
+            # Declared as a measurement but with no quotes on it yet: nothing to
+            # take, so nothing to report as missing either.
+            state = "empty"
+        elif done == total:
+            state = "complete"
+        elif done:
+            state = "partial"
+        else:
+            state = "none"
+        progress.append({
+            "task_id": task_id,
+            "quotes": total,
+            "measured": done,
+            "state": state,
+        })
+    progress.sort(key=lambda row: row["task_id"])
+    return progress
+
+
+async def authorise_measurement(
+    db: AsyncSession, measurement: Measurement, supervisor: User,
+) -> Measurement:
+    """Let an out-of-tolerance value stand, on the supervisor's name."""
+    measurement.supervisor_id = supervisor.id
+    measurement.authorised_at = datetime.now()
+    await db.flush()
+    await db.refresh(measurement)
+    return measurement
+
+
 async def save_measurement(
     db: AsyncSession,
     subtask_id: int,
@@ -48,8 +222,20 @@ async def save_measurement(
     lot_number: str | None = None,
     serial_number: str | None = None,
     input_method: str = "manual",
+    input_duration_ms: int | None = None,
+    production_run_id: int | None = None,
+    vision_result_id: int | None = None,
+    enforce_tolerance_gate: bool = True,
 ) -> Measurement:
-    """Save a single measurement with auto-calculated pass/fail."""
+    """Save a single measurement with auto-calculated pass/fail.
+
+    ``enforce_tolerance_gate=False`` skips only the pending-authorisation check
+    below - everything else (verdict, recipe rules, deviation) still applies.
+    This is for a caller that has already evaluated the gate itself for a whole
+    batch of quotes taken at the same instant (see Ruling R10 / vision_service):
+    the gate models "moving on" between measurements taken one after another,
+    which does not exist inside a single camera acquisition.
+    """
     # Get subtask for tolerance values
     result = await db.execute(
         select(RecipeSubtask).where(RecipeSubtask.id == subtask_id)
@@ -58,6 +244,25 @@ async def save_measurement(
     if subtask is None:
         raise ValueError(f"Subtask {subtask_id} not found")
 
+    recipe = await recipe_of_version(db, version_id)
+    _enforce_recipe_rules(recipe, lot_number, serial_number, input_method)
+
+    # No moving on while a quote is out of tolerance and unauthorised. Measuring
+    # that same quote again is the way out that does not need the capoturno; the
+    # next quote is not.
+    if enforce_tolerance_gate:
+        blocking = await pending_authorisation(
+            db, version_id, measured_by, production_run_id,
+        )
+        if blocking is not None and blocking.subtask_id != subtask_id:
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail=(
+                    "A measurement is out of tolerance and awaiting the supervisor: "
+                    "authorise it or measure that quote again before going on"
+                ),
+            )
+
     pass_fail, deviation = calculate_pass_fail(value, subtask)
 
     measurement = Measurement(
@@ -70,6 +275,9 @@ async def save_measurement(
         lot_number=lot_number,
         serial_number=serial_number,
         input_method=input_method,
+        input_duration_ms=input_duration_ms,
+        production_run_id=production_run_id,
+        vision_result_id=vision_result_id,
     )
     db.add(measurement)
     await db.flush()

src/backend/services/production_export_service.py

@@ -0,0 +1,168 @@
+"""Statistics file emitted when a production is closed.
+
+Fine produzione has to hand the measurements of the whole production over. That part
+does not depend on the ERP: the file is produced now, and the hand-off to GAIA plugs
+in beside it once the protocol is agreed (questions D-1 and D-2). Keeping it in its
+own module makes that seam obvious rather than buried in the close path.
+
+The CSV uses the delimiters configured in system_settings, the same ones the manual
+export honours, so a shop floor set to the Italian convention (';' and ',') gets
+files that open correctly in their Excel.
+"""
+import csv
+import io
+from datetime import datetime
+from pathlib import Path
+
+from sqlalchemy import select, update
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.config import settings
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.production import ProductionRun
+from src.backend.models.orm.recipe import Recipe
+from src.backend.models.orm.setting import SystemSetting
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import RecipeSubtask
+
+# Written under the upload directory: it is the volume that is already mounted and
+# backed up, so the files survive a container being replaced.
+EXPORT_SUBDIR = "statistics"
+
+HEADER = [
+    "production_run_id",
+    "station_code",
+    "recipe_code",
+    "version_id",
+    "lot_number",
+    "serial_number",
+    "measurement_id",
+    "subtask_id",
+    "marker_number",
+    "subtask_description",
+    "nominal",
+    "value",
+    "deviation",
+    "pass_fail",
+    # The limits travel with the row: without them a pass/fail cannot be re-derived
+    # from the file years later, which is the whole point of an audit record.
+    "ltl",
+    "lwl",
+    "uwl",
+    "utl",
+    "unit",
+    "input_method",
+    "input_duration_ms",
+    "measured_by",
+    "measured_at",
+    # A value outside tolerance that was allowed to stand carries the name of who
+    # allowed it. Without these two columns the file would show the failure and
+    # not the decision, which is the half an auditor asks about.
+    "authorised_by",
+    "authorised_at",
+]
+
+
+async def _csv_format(db: AsyncSession) -> tuple[str, str]:
+    """Delimiter and decimal separator as configured for this installation."""
+    delimiter_row = await db.execute(
+        select(SystemSetting).where(SystemSetting.setting_key == "csv_delimiter")
+    )
+    delimiter = delimiter_row.scalar_one_or_none()
+    decimal_row = await db.execute(
+        select(SystemSetting).where(
+            SystemSetting.setting_key == "csv_decimal_separator"
+        )
+    )
+    decimal = decimal_row.scalar_one_or_none()
+    return (
+        delimiter.setting_value if delimiter else ",",
+        decimal.setting_value if decimal else ".",
+    )
+
+
+async def build_statistics_csv(db: AsyncSession, run: ProductionRun) -> str:
+    """Render every measurement of a run as CSV text."""
+    delimiter, decimal_separator = await _csv_format(db)
+
+    station = (await db.execute(
+        select(Station).where(Station.id == run.station_id)
+    )).scalar_one_or_none()
+    recipe = (await db.execute(
+        select(Recipe).where(Recipe.id == run.recipe_id)
+    )).scalar_one_or_none()
+
+    rows = (await db.execute(
+        select(Measurement, RecipeSubtask)
+        .join(RecipeSubtask, RecipeSubtask.id == Measurement.subtask_id, isouter=True)
+        .where(Measurement.production_run_id == run.id)
+        .order_by(Measurement.measured_at, Measurement.id)
+    )).all()
+
+    def number(value) -> str:
+        return "" if value is None else str(value).replace(".", decimal_separator)
+
+    output = io.StringIO()
+    writer = csv.writer(output, delimiter=delimiter, lineterminator="\n")
+    writer.writerow(HEADER)
+
+    for measurement, subtask in rows:
+        writer.writerow([
+            run.id,
+            station.code if station else "",
+            recipe.code if recipe else "",
+            measurement.version_id,
+            measurement.lot_number or "",
+            measurement.serial_number or "",
+            measurement.id,
+            measurement.subtask_id,
+            subtask.marker_number if subtask else "",
+            subtask.description if subtask else "",
+            number(subtask.nominal) if subtask else "",
+            number(measurement.value),
+            number(measurement.deviation),
+            measurement.pass_fail,
+            number(subtask.ltl) if subtask else "",
+            number(subtask.lwl) if subtask else "",
+            number(subtask.uwl) if subtask else "",
+            number(subtask.utl) if subtask else "",
+            (subtask.unit or "") if subtask else "",
+            measurement.input_method,
+            "" if measurement.input_duration_ms is None else measurement.input_duration_ms,
+            measurement.measured_by,
+            measurement.measured_at.isoformat() if measurement.measured_at else "",
+            measurement.supervisor_id or "",
+            measurement.authorised_at.isoformat() if measurement.authorised_at else "",
+        ])
+
+    return output.getvalue()
+
+
+async def export_run_statistics(db: AsyncSession, run: ProductionRun) -> str | None:
+    """Write the run's statistics file and mark its measurements as exported.
+
+    Returns the path relative to the upload directory, or None when the run has no
+    measurements - an empty file would only be noise in the folder.
+    """
+    content = await build_statistics_csv(db, run)
+    # Header only: nothing was measured during this production.
+    if len(content.strip().splitlines()) <= 1:
+        return None
+
+    directory = Path(settings.upload_path) / EXPORT_SUBDIR
+    directory.mkdir(parents=True, exist_ok=True)
+
+    stamp = (run.closed_at or datetime.now()).strftime("%Y%m%d_%H%M%S")
+    filename = f"production_{run.id}_{stamp}.csv"
+    (directory / filename).write_text(content, encoding="utf-8-sig")
+
+    # utf-8-sig: the BOM is what makes Excel read accented characters correctly,
+    # and these files are opened in Excel on the shop floor.
+
+    await db.execute(
+        update(Measurement)
+        .where(Measurement.production_run_id == run.id)
+        .values(synced_to_csv=True)
+    )
+
+    return f"{EXPORT_SUBDIR}/{filename}"

src/backend/services/production_service.py

@@ -0,0 +1,451 @@
+"""Business logic for production runs.
+
+Routers must call into these functions rather than manipulating models directly.
+All functions are async and accept an AsyncSession; they flush but do NOT commit
+(commit is handled by the FastAPI get_db dependency).
+
+Clock convention
+----------------
+Every timestamp this module writes comes from ``_now()``, so the deadline and the
+comparison against it are always on the same clock. Clients are handed
+``seconds_to_next_measurement`` already computed rather than the raw deadline: a
+naive datetime crossing the wire would be parsed in the browser's own timezone, and
+the countdown would be off by the UTC offset. The number is the contract; the
+timestamp is only there for the audit trail.
+"""
+from datetime import datetime, timedelta
+from typing import Optional
+
+from fastapi import HTTPException, status
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.models.api.production import ProductionRunCreate
+from src.backend.models.orm.production import ProductionEvent, ProductionRun
+from src.backend.models.orm.recipe import Recipe, RecipeVersion
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import MEASURING_TASK_TYPES, RecipeTask
+from src.backend.models.orm.user import User
+from src.backend.services import (
+    auth_service, measurement_service, production_export_service,
+)
+
+
+def _now() -> datetime:
+    """Single source of time for runs.
+
+    Naive local time, matching what the database writes for ``func.now()`` defaults,
+    so a run's started_at and its deadline are comparable.
+    """
+    return datetime.now()
+
+
+# ---------------------------------------------------------------------------
+# Derived view
+# ---------------------------------------------------------------------------
+
+
+def seconds_to_next_measurement(run: ProductionRun, now: Optional[datetime] = None) -> Optional[int]:
+    """Seconds until the next measurement is due; negative once it is overdue.
+
+    Letting the value go negative is deliberate: the operator has to see how long
+    the line has been past the interval, not merely that it elapsed.
+    While the run is paused the countdown is frozen at what was left when the line
+    stopped - a fermo linea must not eat into the measurement interval.
+    """
+    if run.next_measurement_at is None:
+        return None
+    reference = run.paused_at if (run.status == "paused" and run.paused_at) else (now or _now())
+    return int(round((run.next_measurement_at - reference).total_seconds()))
+
+
+def describe(run: ProductionRun) -> dict:
+    """Fields the API adds on top of the stored columns."""
+    now = _now()
+    remaining = seconds_to_next_measurement(run, now)
+    return {
+        "seconds_to_next_measurement": remaining,
+        "overdue": remaining is not None and remaining < 0,
+        "server_time": now,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Lookups
+# ---------------------------------------------------------------------------
+
+
+async def get_run(db: AsyncSession, run_id: int) -> ProductionRun:
+    result = await db.execute(select(ProductionRun).where(ProductionRun.id == run_id))
+    run = result.scalar_one_or_none()
+    if run is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND, detail="Production run not found",
+        )
+    return run
+
+
+async def get_open_run_for_station(
+    db: AsyncSession, station_id: int,
+) -> Optional[ProductionRun]:
+    """The run currently open at a station, running or paused."""
+    result = await db.execute(
+        select(ProductionRun).where(ProductionRun.active_station_id == station_id)
+    )
+    return result.scalar_one_or_none()
+
+
+async def get_station_by_code(db: AsyncSession, code: str) -> Station:
+    result = await db.execute(select(Station).where(Station.code == code))
+    station = result.scalar_one_or_none()
+    if station is None or not station.active:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=f"Station '{code}' not found or inactive",
+        )
+    return station
+
+
+async def measurement_task_ids(db: AsyncSession, version_id: int) -> list[int]:
+    """Ids of the measurement tasks of a version, in the order they are executed.
+
+    This is the loop the operator lives in once production starts: the client uses
+    the first of these to bring the measurement back on screen when the interval
+    expires, and the last to know which task closes a cycle.
+
+    The type is read from the declared task_type (point 2), not from whether quotes
+    happen to be present: a measurement task with no quotes yet is still where the
+    operator has to be taken.
+    """
+    result = await db.execute(
+        select(RecipeTask.id)
+        .where(
+            RecipeTask.version_id == version_id,
+            RecipeTask.task_type.in_(MEASURING_TASK_TYPES),
+        )
+        .order_by(RecipeTask.order_index, RecipeTask.id)
+    )
+    return list(result.scalars().all())
+
+
+async def list_run_events(db: AsyncSession, run_id: int) -> list[ProductionEvent]:
+    result = await db.execute(
+        select(ProductionEvent)
+        .where(ProductionEvent.run_id == run_id)
+        .order_by(ProductionEvent.created_at, ProductionEvent.id)
+    )
+    return list(result.scalars().all())
+
+
+# ---------------------------------------------------------------------------
+# Supervisor authorisation
+# The capoturno check lives in auth_service, next to the rest of the credential
+# handling: a line stop, a closure and an out-of-tolerance value all ask the same
+# question, and only one of the three is about production.
+
+
+# ---------------------------------------------------------------------------
+# Lifecycle
+# ---------------------------------------------------------------------------
+
+
+def _add_event(
+    db: AsyncSession,
+    run: ProductionRun,
+    event_type: str,
+    user: User,
+    supervisor: Optional[User] = None,
+    note: Optional[str] = None,
+    task_id: Optional[int] = None,
+) -> ProductionEvent:
+    """Append to the run's trace.
+
+    Added through the session rather than run.events: touching the collection would
+    trigger a lazy load, which raises under async SQLAlchemy.
+    """
+    event = ProductionEvent(
+        run_id=run.id,
+        event_type=event_type,
+        user_id=user.id,
+        supervisor_id=supervisor.id if supervisor else None,
+        note=note,
+        task_id=task_id,
+        created_at=_now(),
+    )
+    db.add(event)
+    return event
+
+
+async def open_run(
+    db: AsyncSession, data: ProductionRunCreate, operator: User,
+) -> ProductionRun:
+    """Start a production at a station.
+
+    Refuses if the station already has one open: two runs on the same station would
+    mean two timers and two histories for one physical line.
+    """
+    station = await get_station_by_code(db, data.station_code)
+
+    existing = await get_open_run_for_station(db, station.id)
+    if existing is not None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=f"Station '{station.code}' already has an open production run",
+        )
+
+    recipe_row = await db.execute(select(Recipe).where(Recipe.id == data.recipe_id))
+    recipe = recipe_row.scalar_one_or_none()
+    if recipe is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND, detail="Recipe not found",
+        )
+
+    if data.version_id is not None:
+        version_row = await db.execute(
+            select(RecipeVersion).where(
+                RecipeVersion.id == data.version_id,
+                RecipeVersion.recipe_id == recipe.id,
+            )
+        )
+        version = version_row.scalar_one_or_none()
+        if version is None:
+            raise HTTPException(
+                status_code=status.HTTP_404_NOT_FOUND,
+                detail="Recipe version not found for this recipe",
+            )
+    else:
+        version_row = await db.execute(
+            select(RecipeVersion).where(
+                RecipeVersion.recipe_id == recipe.id,
+                RecipeVersion.is_current == True,  # noqa: E712
+            )
+        )
+        version = version_row.scalar_one_or_none()
+        if version is None:
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail="Recipe has no current version",
+            )
+
+    # Traceability the recipe declares compulsory has to be there before the first
+    # measurement, not after it: a lot entered halfway through leaves the values
+    # taken up to that point unattributable.
+    if recipe.requires_lot and not (data.lot_number or "").strip():
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="This recipe requires a lot number to start a production",
+        )
+    if recipe.requires_serial and not (data.serial_number or "").strip():
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="This recipe requires a serial number to start a production",
+        )
+
+    now = _now()
+    # The interval is copied, not referenced: editing the recipe mid-production must
+    # not move the deadline of a run already under way.
+    interval = recipe.measurement_interval_minutes
+    run = ProductionRun(
+        station_id=station.id,
+        recipe_id=recipe.id,
+        version_id=version.id,
+        operator_id=operator.id,
+        lot_number=data.lot_number,
+        serial_number=data.serial_number,
+        status="running",
+        measurement_interval_minutes=interval,
+        next_measurement_at=(
+            now + timedelta(minutes=interval) if interval and interval > 0 else None
+        ),
+        cycle_count=0,
+        started_at=now,
+        active_station_id=station.id,
+    )
+    db.add(run)
+    await db.flush()
+    _add_event(db, run, "start", operator)
+    await db.flush()
+    await db.refresh(run)
+    return run
+
+
+def _require_open(run: ProductionRun) -> None:
+    if run.status == "closed":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="Production run is already closed",
+        )
+
+
+def _require_measurable(run: ProductionRun) -> None:
+    _require_open(run)
+    if run.status == "paused":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="Production run is paused: resume it before measuring",
+        )
+
+
+async def complete_cycle(
+    db: AsyncSession,
+    run: ProductionRun,
+    user: User,
+    task_id: Optional[int] = None,
+    note: Optional[str] = None,
+) -> ProductionRun:
+    """Record a finished measurement task, and restart the interval if it was the last.
+
+    A cycle spans every measurement task of the recipe, so only the last one closes
+    it. Restarting the interval at the first would make a recipe with three
+    measurement tasks measure three times as often as it was set up to.
+
+    An unrecognised task_id - or none at all - closes the cycle: a caller that does
+    not know about the sequence gets the behaviour that existed before it.
+    """
+    _require_measurable(run)
+
+    # A cycle that contains a quote out of tolerance is not a finished cycle. This
+    # is the same gate as on the next measurement: closing the cycle would be
+    # another way of walking past it.
+    blocking = await measurement_service.pending_authorisation(
+        db, run.version_id, user.id, run.id,
+    )
+    if blocking is not None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=(
+                "A measurement is out of tolerance and awaiting the supervisor: "
+                "the cycle cannot be closed until it is authorised"
+            ),
+        )
+
+    measure_ids = await measurement_task_ids(db, run.version_id)
+    closes_cycle = (
+        task_id is None or task_id not in measure_ids or task_id == measure_ids[-1]
+    )
+
+    if not closes_cycle:
+        _add_event(db, run, "task_measured", user, note=note, task_id=task_id)
+        await db.flush()
+        await db.refresh(run)
+        return run
+
+    now = _now()
+    run.cycle_count += 1
+    if run.measurement_interval_minutes and run.measurement_interval_minutes > 0:
+        # Measured from now, not from the previous deadline: a late measurement must
+        # not compress the following interval.
+        run.next_measurement_at = now + timedelta(
+            minutes=run.measurement_interval_minutes
+        )
+    _add_event(db, run, "cycle_completed", user, note=note, task_id=task_id)
+    await db.flush()
+    await db.refresh(run)
+    return run
+
+
+async def remeasure(
+    db: AsyncSession,
+    run: ProductionRun,
+    user: User,
+    task_id: Optional[int] = None,
+    note: Optional[str] = None,
+) -> ProductionRun:
+    """Measure the same task again without closing the cycle.
+
+    This is the piece being turned over. The deadline and the cycle count are left
+    exactly as they are - a second reading of the same part is not a new cycle, and
+    it must not buy the operator another interval's worth of time. The values
+    themselves are ordinary measurements: both readings stay in the statistics,
+    which is the point of taking them.
+    """
+    _require_measurable(run)
+    _add_event(db, run, "remeasure", user, note=note, task_id=task_id)
+    await db.flush()
+    await db.refresh(run)
+    return run
+
+
+async def pause_run(
+    db: AsyncSession,
+    run: ProductionRun,
+    user: User,
+    supervisor: User,
+    note: Optional[str] = None,
+) -> ProductionRun:
+    """Fermo linea: suspend the run and freeze the countdown."""
+    _require_open(run)
+    if run.status == "paused":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT, detail="Production run is already paused",
+        )
+    run.status = "paused"
+    run.paused_at = _now()
+    _add_event(db, run, "line_stop", user, supervisor=supervisor, note=note)
+    await db.flush()
+    await db.refresh(run)
+    return run
+
+
+async def resume_run(
+    db: AsyncSession,
+    run: ProductionRun,
+    user: User,
+    supervisor: User,
+    note: Optional[str] = None,
+) -> ProductionRun:
+    """Restart a paused run, pushing the deadline out by the length of the stop."""
+    _require_open(run)
+    if run.status != "paused":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT, detail="Production run is not paused",
+        )
+    now = _now()
+    if run.next_measurement_at is not None and run.paused_at is not None:
+        # Shift rather than recompute: whatever was left when the line stopped is
+        # what the operator gets back, so a stop neither grants nor costs time.
+        run.next_measurement_at = run.next_measurement_at + (now - run.paused_at)
+    run.status = "running"
+    run.paused_at = None
+    _add_event(db, run, "resume", user, supervisor=supervisor, note=note)
+    await db.flush()
+    await db.refresh(run)
+    return run
+
+
+async def close_run(
+    db: AsyncSession,
+    run: ProductionRun,
+    user: User,
+    supervisor: User,
+    note: Optional[str] = None,
+) -> ProductionRun:
+    """Fine produzione: close the run for good, stop the timer, emit the statistics.
+
+    Everything here happens without the ERP. Handing the production over to GAIA
+    goes at the marked seam below, once D-1 and D-2 are answered; until then the
+    file is produced and the run is properly closed, which is what the shop floor
+    needs either way.
+    """
+    _require_open(run)
+    run.status = "closed"
+    run.closed_at = _now()
+    run.closed_by = supervisor.id
+    run.next_measurement_at = None
+    run.paused_at = None
+    # Releasing the slot lets the station open a new run; the unique index on this
+    # column is what keeps exactly one open at a time.
+    run.active_station_id = None
+    _add_event(db, run, "close", user, supervisor=supervisor, note=note)
+    await db.flush()
+
+    # Statistics file for the whole production, and the measurements marked as sent.
+    run.statistics_path = await production_export_service.export_run_statistics(db, run)
+
+    # --- ERP hand-off (GAIA) plugs in here once the protocol is defined (D-1, D-2).
+    # Deliberately absent rather than stubbed: an empty call that looks wired is
+    # worse than none at all.
+
+    await db.flush()
+    await db.refresh(run)
+    return run

src/backend/services/recipe_service.py

@@ -19,6 +19,30 @@ from src.backend.models.api.recipe import RecipeCreate, RecipeUpdate
 # Helpers
 # ---------------------------------------------------------------------------
 
+# Columns of the recipe header itself, as opposed to the versioned content. Both
+# update paths - copy-on-write and in-place - write exactly these, so the list
+# lives once: a field added to one and forgotten in the other would appear to save
+# and then quietly not, depending on whether the version had measurements.
+HEADER_FIELDS = (
+    "name",
+    "description",
+    "image_path",
+    "measurement_interval_minutes",
+    "requires_lot",
+    "requires_serial",
+    "allow_manual_input",
+)
+
+
+def _header_fields(data: RecipeUpdate) -> dict:
+    """The header columns the caller actually asked to change."""
+    return {
+        field: getattr(data, field)
+        for field in HEADER_FIELDS
+        if getattr(data, field) is not None
+    }
+
+
 async def _get_recipe_or_404(db: AsyncSession, recipe_id: int) -> Recipe:
     """Return a recipe or raise 404."""
     result = await db.execute(select(Recipe).where(Recipe.id == recipe_id))
@@ -55,12 +79,20 @@ async def _copy_tasks_to_version(
         new_task = RecipeTask(
             version_id=target_version.id,
             order_index=task.order_index,
+            # Carried over explicitly: a copy-on-write version that dropped the type
+            # would silently reclassify every task the moment a recipe is edited.
+            task_type=task.task_type,
             title=task.title,
             directive=task.directive,
             description=task.description,
             file_path=task.file_path,
             file_type=task.file_type,
             annotations_json=task.annotations_json,
+            # Carried over explicitly, same reasoning as task_type above: a
+            # copy-on-write version that dropped the camera graph would leave
+            # a camera_measure task unable to execute from the moment a
+            # recipe with measurements is next edited.
+            vision_json=task.vision_json,
         )
         db.add(new_task)
         await db.flush()  # get new_task.id
@@ -78,6 +110,7 @@ async def _copy_tasks_to_version(
                 ltl=sub.ltl,
                 unit=sub.unit,
                 image_path=sub.image_path,
+                vision_output=sub.vision_output,
             )
             db.add(new_sub)
 
@@ -134,6 +167,10 @@ async def create_recipe(
         name=data.name,
         description=data.description,
         image_path=data.image_path,
+        measurement_interval_minutes=data.measurement_interval_minutes,
+        requires_lot=data.requires_lot,
+        requires_serial=data.requires_serial,
+        allow_manual_input=data.allow_manual_input,
         created_by=user.id,
     )
     db.add(recipe)
@@ -253,6 +290,7 @@ async def create_new_version(
             default_task = RecipeTask(
                 version_id=new_version.id,
                 order_index=0,
+                task_type="drawing",
                 title="Technical Drawing",
                 file_path=data.file_path,
                 file_type=data.file_type or (
@@ -266,13 +304,7 @@ async def create_new_version(
         await db.flush()
 
     # Apply header updates
-    update_fields: dict = {}
-    if data.name is not None:
-        update_fields["name"] = data.name
-    if data.description is not None:
-        update_fields["description"] = data.description
-    if data.image_path is not None:
-        update_fields["image_path"] = data.image_path
+    update_fields = _header_fields(data)
     if update_fields:
         await db.execute(
             update(Recipe).where(Recipe.id == recipe_id).values(**update_fields)
@@ -361,14 +393,7 @@ async def update_current_version(
     data: RecipeUpdate,
 ) -> RecipeVersion:
     """Update recipe header in-place on the current version (no copy-on-write)."""
-    # Apply header updates (name, description, image_path)
-    update_fields: dict = {}
-    if data.name is not None:
-        update_fields["name"] = data.name
-    if data.description is not None:
-        update_fields["description"] = data.description
-    if data.image_path is not None:
-        update_fields["image_path"] = data.image_path
+    update_fields = _header_fields(data)
     if update_fields:
         await db.execute(
             update(Recipe).where(Recipe.id == recipe_id).values(**update_fields)

src/backend/services/station_service.py

@@ -138,6 +138,26 @@ async def unassign_recipe(
     await db.flush()
 
 
+async def unassign_all_recipes(db: AsyncSession, station_id: int) -> int:
+    """Drop every recipe assignment of a station and return how many were removed.
+
+    Used by the per-row reset in the admin station list: the station goes back to
+    having no recipes so it can be reassigned from scratch. Unlike unassign_recipe
+    this is idempotent - resetting an already empty station returns 0 rather than 404.
+    """
+    await get_station(db, station_id)
+    result = await db.execute(
+        select(StationRecipeAssignment).where(
+            StationRecipeAssignment.station_id == station_id
+        )
+    )
+    assignments = list(result.scalars().all())
+    for assignment in assignments:
+        await db.delete(assignment)
+    await db.flush()
+    return len(assignments)
+
+
 async def list_station_recipes(
     db: AsyncSession, station_id: int,
 ) -> list[Recipe]:

src/backend/services/vision_service.py

@@ -0,0 +1,234 @@
+"""Route the image to the worker, map the outputs onto the quotes, save.
+
+This module never imports VisionSuite - see test_vision_boundary.py. It speaks
+to the worker over HTTP and stays ignorant of how a number was produced.
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from uuid import uuid4
+
+import httpx
+from fastapi import HTTPException, status
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.config import settings
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.models.orm.vision import VisionReferenceImage, VisionResult
+from src.backend.services import measurement_service
+
+
+async def _call_worker(image_bytes: bytes, graph: dict) -> dict:
+    """The one place that talks to the worker; the tests replace it.
+
+    The worker's fault must not become the API's fault (spec's edge-case
+    table: "il worker isola il guasto, l'API resta in piedi, il task
+    riporta l'errore"). A stopped worker raises `httpx.ConnectError`; a job
+    past the timeout raises `httpx.ReadTimeout` - both would otherwise
+    propagate past this function and surface to the operator as a naked
+    500. The response shape is validated here too: a proxy answering 200
+    with something that is not the worker's JSON contract must not turn
+    into a bare KeyError further down.
+    """
+    try:
+        async with httpx.AsyncClient(timeout=120.0) as client:
+            response = await client.post(
+                f"{settings.vision_worker_url}/run",
+                files={"image": ("frame.png", image_bytes, "image/png")},
+                data={"graph": json.dumps(graph)},
+            )
+    except httpx.ConnectError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=f"the vision worker is unreachable: {exc}",
+        ) from exc
+    except httpx.ReadTimeout as exc:
+        raise HTTPException(
+            status_code=status.HTTP_504_GATEWAY_TIMEOUT,
+            detail=f"the vision worker did not answer in time: {exc}",
+        ) from exc
+
+    if response.status_code != 200:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=f"the vision worker refused the job: {response.text}",
+        )
+
+    try:
+        report = response.json()
+    except ValueError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=f"the vision worker's answer was malformed: not valid JSON ({exc})",
+        ) from exc
+
+    if not isinstance(report, dict):
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail="the vision worker's answer was malformed: expected a JSON object",
+        )
+    missing = {"outputs", "failures", "engine_version"} - report.keys()
+    if missing:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=(
+                "the vision worker's answer was malformed: missing "
+                f"{', '.join(sorted(missing))}"
+            ),
+        )
+    return report
+
+
+async def execute_task(
+    db: AsyncSession,
+    task: RecipeTask,
+    image_bytes: bytes,
+    measured_by: int,
+    *,
+    version_id: int,
+    lot_number: str | None = None,
+    serial_number: str | None = None,
+    production_run_id: int | None = None,
+) -> tuple[VisionResult, list[Measurement]]:
+    if not task.vision_json:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="this task carries no vision graph",
+        )
+
+    # Ordered by marker_number - the number the operator sees on the annotated
+    # image and in reports (RecipeSubtask.__repr__, statistics.py, report_service.py
+    # all key off it). Which quote is saved first must never be an accident of
+    # how SQLite/MySQL happens to return rows.
+    quotes = (await db.execute(
+        select(RecipeSubtask)
+        .where(RecipeSubtask.task_id == task.id)
+        .where(RecipeSubtask.vision_output.is_not(None))
+        .order_by(RecipeSubtask.marker_number)
+    )).scalars().all()
+    if not quotes:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="no quote on this task names a vision output",
+        )
+
+    # Ruling R10: the gate is evaluated once per acquisition, not once per
+    # quote. `save_measurement`'s own gate models "moving on" between
+    # measurements taken one after another - a real later moment where the
+    # operator could leave a bad reading behind. A camera acquisition produces
+    # every quote of this task at the same instant: there is no "moving on"
+    # between them, so applying the sequential rule inside the batch would
+    # block on a fail this very request just saved. Check it once here,
+    # against the whole set this acquisition is about to produce, and pass
+    # enforce_tolerance_gate=False to every save_measurement call below.
+    #
+    # Checked before the worker runs (M3): a blocked operator finds out
+    # without burning a whole vision execution first - the worker is the
+    # expensive part, the gate check is not.
+    blocking = await measurement_service.pending_authorisation(
+        db, version_id, measured_by, production_run_id,
+    )
+    quote_ids = {q.id for q in quotes}
+    if blocking is not None and blocking.subtask_id not in quote_ids:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=(
+                "A measurement is out of tolerance and awaiting the supervisor: "
+                "authorise it or measure that quote again before going on"
+            ),
+        )
+
+    report = await _call_worker(image_bytes, task.vision_json)
+    outputs = report["outputs"]
+
+    # Every quote or none. A half-measured piece read as a whole one is worse
+    # than a piece not measured at all.
+    missing = [q.vision_output for q in quotes if q.vision_output not in outputs]
+    if missing:
+        failed = ", ".join(f["tool_name"] for f in report["failures"])
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail=(
+                f"the graph did not produce {', '.join(missing)}"
+                + (f" (tools that failed: {failed})" if failed else "")
+            ),
+        )
+
+    result = VisionResult(
+        task_id=task.id,
+        engine_version=report["engine_version"],
+        executed_on="server",
+        graph_snapshot=task.vision_json,
+        duration_ms=report.get("duration_ms"),
+    )
+    db.add(result)
+    await db.flush()
+
+    saved: list[Measurement] = []
+    for quote in quotes:
+        saved.append(await measurement_service.save_measurement(
+            db,
+            subtask_id=quote.id,
+            version_id=version_id,
+            measured_by=measured_by,
+            value=float(outputs[quote.vision_output]),
+            lot_number=lot_number,
+            serial_number=serial_number,
+            input_method="camera",
+            production_run_id=production_run_id,
+            vision_result_id=result.id,
+            enforce_tolerance_gate=False,
+        ))
+
+    return result, saved
+
+
+async def save_reference_image(
+    db: AsyncSession,
+    task_id: int,
+    image_bytes: bytes,
+    *,
+    note: str | None = None,
+    station_id: int | None = None,
+    device_code: str | None = None,
+    calibration_snapshot: dict | None = None,
+    extension: str = "png",
+) -> VisionReferenceImage:
+    """Keep an image so the graph can be composed and re-tried without a camera.
+
+    `extension` must match the upload's real content type (the router derives
+    it from `image.content_type`) - writing a jpeg to disk as `<uuid>.png`
+    would be a lie the filename tells forever.
+    """
+    # `upload_path` is the resolved Path, `upload_dir` the bare setting: the
+    # rest of the codebase uses the former (see production_export_service).
+    folder = settings.upload_path / "vision" / "reference" / str(task_id)
+    folder.mkdir(parents=True, exist_ok=True)
+    name = f"{uuid4().hex}.{extension}"
+    (folder / name).write_bytes(image_bytes)
+
+    reference = VisionReferenceImage(
+        task_id=task_id,
+        path=str(Path("vision") / "reference" / str(task_id) / name),
+        note=note,
+        station_id=station_id,
+        device_code=device_code,
+        calibration_snapshot=calibration_snapshot,
+    )
+    db.add(reference)
+    await db.flush()
+    await db.refresh(reference)
+    return reference
+
+
+async def preview(db: AsyncSession, task: RecipeTask, image_bytes: bytes) -> dict:
+    """Run the graph and save nothing. Composing is not measuring."""
+    if not task.vision_json:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            detail="this task carries no vision graph",
+        )
+    return await _call_worker(image_bytes, task.vision_json)

src/backend/templates/setup/setup.html

@@ -559,6 +559,19 @@
         return data;
     }
 
+    /* Setup operations rotate api_keys (DB init, admin creation, password
+     * change, user (de)activation). Any Flask session held in this browser
+     * still carries the old key, so we clear it to force a fresh /auth/login. */
+    async function invalidateClientSession() {
+        try {
+            await fetch('/auth/logout', {
+                method: 'POST',
+                credentials: 'same-origin',
+                redirect: 'manual',
+            });
+        } catch (_) { /* best-effort */ }
+    }
+
     /* ------------------------------------------------------------------ */
     /* Check DB status                                                     */
     /* ------------------------------------------------------------------ */
@@ -635,6 +648,7 @@
         try {
             const data = await apiPost('/init-db', { password: setupPassword });
             showToast(data.message, 'success');
+            await invalidateClientSession();
             await checkStatus();
         } catch (e) {
             showToast(e.message, 'error');
@@ -667,6 +681,7 @@
                 admin_display_name: display,
             });
             showToast(data.message, 'success');
+            await invalidateClientSession();
             // Clear form
             document.getElementById('admin-username').value = '';
             document.getElementById('admin-password').value = '';
@@ -690,6 +705,7 @@
             if (data.measurements_created) msg += ', measurements: ' + data.measurements_created;
             msg += ')';
             showToast(msg, 'success');
+            await invalidateClientSession();
             await loadUsers();
         } catch (e) {
             showToast(e.message, 'error');
@@ -876,6 +892,7 @@
             if (pwd) body.user_password = pwd;
             var data = await apiPost('/manage-user', body);
             showToast(data.message, 'success');
+            await invalidateClientSession();
             closeModal('modal-user');
             await loadUsers();
         } catch (e) {
@@ -908,6 +925,7 @@
                 new_password: newPwd,
             });
             showToast(data.message, 'success');
+            await invalidateClientSession();
             closeModal('modal-password');
         } catch (e) {
             showToast(e.message, 'error');
@@ -924,6 +942,7 @@
                 user_id: userId,
             });
             showToast(data.message, 'success');
+            await invalidateClientSession();
             await loadUsers();
         } catch (e) {
             showToast(e.message, 'error');

src/backend/tests/conftest.py

@@ -224,6 +224,10 @@ async def create_test_recipe(
         code=code,
         name=name,
         description="A recipe for testing",
+        # Typing is allowed here so that tests about something else can still save
+        # a measurement. The rule itself is exercised by test_recipe_rules.py, on
+        # recipes that declare it explicitly.
+        allow_manual_input=True,
         created_by=user_id,
     )
     session.add(recipe)
@@ -243,6 +247,10 @@ async def create_test_recipe(
         version_id=version.id,
         order_index=0,
         title="Test Task",
+        # Declared, like any task since the type stopped being deduced: this one has
+        # a quote under it, so a fixture that called it a note would be modelling
+        # exactly the inconsistency point 2 removed.
+        task_type="measure",
         directive="Measure the part",
         description="First measurement task",
     )

src/backend/tests/test_measurements.py

@@ -55,6 +55,59 @@ class TestCreateMeasurement:
         assert data["pass_fail"] in ("pass", "warning", "fail")
         assert data["measured_by"] == measurement_tec_user.id
 
+    async def test_create_measurement_with_input_duration(
+        self,
+        client: AsyncClient,
+        measurement_tec_user: User,
+        db_session: AsyncSession,
+    ):
+        """input_duration_ms is stored and returned when provided."""
+        recipe = await create_test_recipe(
+            db_session, measurement_tec_user.id
+        )
+        subtask_id, version_id = await _get_subtask_and_version(
+            client, measurement_tec_user, recipe.id
+        )
+
+        resp = await client.post(
+            "/api/measurements/",
+            headers=auth_headers(measurement_tec_user),
+            json={
+                "subtask_id": subtask_id,
+                "version_id": version_id,
+                "value": 10.0,
+                "input_duration_ms": 4250,
+            },
+        )
+        assert resp.status_code == 200
+        assert resp.json()["input_duration_ms"] == 4250
+
+    async def test_create_measurement_without_input_duration(
+        self,
+        client: AsyncClient,
+        measurement_tec_user: User,
+        db_session: AsyncSession,
+    ):
+        """input_duration_ms defaults to null when omitted."""
+        recipe = await create_test_recipe(
+            db_session, measurement_tec_user.id
+        )
+        subtask_id, version_id = await _get_subtask_and_version(
+            client, measurement_tec_user, recipe.id
+        )
+
+        resp = await client.post(
+            "/api/measurements/",
+            headers=auth_headers(measurement_tec_user),
+            json={
+                "subtask_id": subtask_id,
+                "version_id": version_id,
+                "value": 10.0,
+            },
+        )
+        assert resp.status_code == 200
+        assert resp.json()["input_duration_ms"] is None
+
     async def test_measurement_requires_auth(self, client: AsyncClient):
         """Creating measurement without auth returns 401."""
         resp = await client.post(

src/backend/tests/test_out_of_tolerance.py

@@ -0,0 +1,411 @@
+"""Point 5: a quote out of tolerance stops the line until someone says otherwise.
+
+The gate existed on screen and nowhere else - a modal that opened, checked a
+password and closed, while the measurement had already been saved. Dismissing it
+was enough to carry on. These tests are about the part that cannot be dismissed.
+"""
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.services import auth_service
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+GOOD = 10.0     # inside every limit of the fixture quote
+OUT = 99.0      # far outside
+
+
+async def _supervisor(db_session, username="capo-tol", password="TurnoPwd1"):
+    user = await auth_service.create_user(
+        db_session, username=username, password=password,
+        display_name="Capo Turno", roles=["Supervisor"],
+    )
+    await db_session.commit()
+    return user, password
+
+
+async def _two_quotes(db_session, recipe_id: int):
+    """The fixture recipe plus a second quote, so there is a 'next' one."""
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+
+    second = RecipeSubtask(
+        task_id=task.id, marker_number=2, description="Seconda quota",
+        nominal=10.0, utl=10.5, uwl=10.3, lwl=9.7, ltl=9.5, unit="mm",
+    )
+    db_session.add(second)
+    await db_session.commit()
+    await db_session.refresh(second)
+    return version.id, first, second
+
+
+async def _measure(client, user, subtask_id, version_id, value, run_id=None):
+    body = {"subtask_id": subtask_id, "version_id": version_id, "value": value}
+    if run_id is not None:
+        body["production_run_id"] = run_id
+    return await client.post(
+        "/api/measurements/", headers=auth_headers(user), json=body,
+    )
+
+
+async def _pending(client, user, version_id, run_id=None):
+    params = {"version_id": version_id}
+    if run_id is not None:
+        params["production_run_id"] = run_id
+    return await client.get(
+        "/api/measurements/pending-authorisation",
+        headers=auth_headers(user), params=params,
+    )
+
+
+# ---------------------------------------------------------------------------
+# The block
+# ---------------------------------------------------------------------------
+
+
+async def test_next_quote_is_refused_while_one_is_out_of_tolerance(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-NEXT")
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+
+    bad = await _measure(client, measurement_tec_user, first.id, version_id, OUT)
+    assert bad.status_code == 200
+    assert bad.json()["pass_fail"] == "fail"
+
+    refused = await _measure(client, measurement_tec_user, second.id, version_id, GOOD)
+    assert refused.status_code == 409
+    assert "supervisor" in refused.json()["detail"].lower()
+
+
+async def test_the_same_quote_can_be_measured_again(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """The caliper slips, the piece is reseated. Trying again is not going on."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-AGAIN")
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+
+    await _measure(client, measurement_tec_user, first.id, version_id, OUT)
+    again = await _measure(client, measurement_tec_user, first.id, version_id, OUT)
+    assert again.status_code == 200, "rimisurare la stessa quota deve restare possibile"
+
+
+async def test_a_good_reading_of_the_same_quote_releases_the_block(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """Only the latest reading of a quote counts: it is in tolerance now."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-FIXED")
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+
+    await _measure(client, measurement_tec_user, first.id, version_id, OUT)
+    await _measure(client, measurement_tec_user, first.id, version_id, GOOD)
+
+    ahead = await _measure(client, measurement_tec_user, second.id, version_id, GOOD)
+    assert ahead.status_code == 200
+
+    # Both readings stay on record - how many are allowed is point 4's business.
+    rows = (await db_session.execute(
+        select(Measurement).where(Measurement.subtask_id == first.id)
+    )).scalars().all()
+    assert len(rows) == 2
+
+
+async def test_a_warning_does_not_block(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """Outside the warning limits but inside tolerance is still inside tolerance."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-WARN")
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+
+    warned = await _measure(client, measurement_tec_user, first.id, version_id, 10.4)
+    assert warned.json()["pass_fail"] == "warning"
+
+    ahead = await _measure(client, measurement_tec_user, second.id, version_id, GOOD)
+    assert ahead.status_code == 200
+
+
+async def test_pending_authorisation_names_the_blocking_quote(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """What the screen asks on load, so a reload is not a way past the gate."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-PEND")
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+
+    assert (await _pending(client, measurement_tec_user, version_id)).json() is None
+
+    bad = (await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT,
+    )).json()
+
+    pending = (await _pending(client, measurement_tec_user, version_id)).json()
+    assert pending["id"] == bad["id"]
+    assert pending["subtask_id"] == first.id
+
+
+# ---------------------------------------------------------------------------
+# The authorisation
+# ---------------------------------------------------------------------------
+
+
+async def test_supervisor_authorisation_is_recorded_and_releases_the_block(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-AUTH")
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+    supervisor, password = await _supervisor(db_session)
+
+    bad = (await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT,
+    )).json()
+
+    authorised = await client.post(
+        f"/api/measurements/{bad['id']}/authorise",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": supervisor.username,
+            "supervisor_password": password,
+        },
+    )
+    assert authorised.status_code == 200, authorised.text
+    body = authorised.json()
+    # Who said yes is on the measurement itself, which is the whole point.
+    assert body["supervisor_id"] == supervisor.id
+    assert body["authorised_at"] is not None
+
+    ahead = await _measure(client, measurement_tec_user, second.id, version_id, GOOD)
+    assert ahead.status_code == 200
+
+
+async def test_authorisation_refuses_someone_who_is_not_a_supervisor(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-NOSUP")
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+
+    bad = (await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT,
+    )).json()
+
+    resp = await client.post(
+        f"/api/measurements/{bad['id']}/authorise",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": maker_user.username,
+            "supervisor_password": "testpassword123",
+        },
+    )
+    assert resp.status_code in (401, 403)
+
+    # And the block is still there.
+    refused = await _measure(client, measurement_tec_user, second.id, version_id, GOOD)
+    assert refused.status_code == 409
+
+
+async def test_authorisation_refuses_wrong_credentials(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-BADPW")
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+    supervisor, _password = await _supervisor(db_session, username="capo-badpw")
+
+    bad = (await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT,
+    )).json()
+    resp = await client.post(
+        f"/api/measurements/{bad['id']}/authorise",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": supervisor.username,
+            "supervisor_password": "sbagliata",
+        },
+    )
+    assert resp.status_code == 401
+
+
+async def test_there_is_nothing_to_authorise_on_a_good_measurement(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-GOOD")
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+    supervisor, password = await _supervisor(db_session, username="capo-good")
+
+    good = (await _measure(
+        client, measurement_tec_user, first.id, version_id, GOOD,
+    )).json()
+    resp = await client.post(
+        f"/api/measurements/{good['id']}/authorise",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": supervisor.username,
+            "supervisor_password": password,
+        },
+    )
+    assert resp.status_code == 409
+
+
+# ---------------------------------------------------------------------------
+# The production run
+# ---------------------------------------------------------------------------
+
+
+async def test_cycle_cannot_close_over_an_unauthorised_fail(
+    client: AsyncClient, maker_user, measurement_tec_user, admin_user, db_session,
+):
+    """Closing the cycle would be another way of walking past the quote."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-CYC")
+    recipe.measurement_interval_minutes = 10
+    station = Station(
+        code="ST-TOL", name="Stazione", active=True, created_by=admin_user.id,
+    )
+    db_session.add(station)
+    await db_session.commit()
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+
+    run = (await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )).json()
+
+    await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT, run_id=run["id"],
+    )
+
+    refused = await client.post(
+        f"/api/production-runs/{run['id']}/cycle",
+        headers=auth_headers(measurement_tec_user),
+    )
+    assert refused.status_code == 409
+
+    supervisor, password = await _supervisor(db_session, username="capo-cyc")
+    blocking = (await _pending(
+        client, measurement_tec_user, version_id, run_id=run["id"],
+    )).json()
+    await client.post(
+        f"/api/measurements/{blocking['id']}/authorise",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": supervisor.username,
+            "supervisor_password": password,
+        },
+    )
+
+    allowed = await client.post(
+        f"/api/production-runs/{run['id']}/cycle",
+        headers=auth_headers(measurement_tec_user),
+    )
+    assert allowed.status_code == 200
+    assert allowed.json()["cycle_count"] == 1
+
+
+async def test_the_statistics_file_carries_the_decision(
+    client: AsyncClient, maker_user, measurement_tec_user, admin_user, db_session,
+    tmp_path, monkeypatch,
+):
+    """The file has to show the failure and who let it stand, not just the failure."""
+    from src.backend.config import settings
+
+    monkeypatch.setattr(
+        type(settings), "upload_path", property(lambda self: tmp_path), raising=False,
+    )
+
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-CSV")
+    station = Station(
+        code="ST-CSV", name="Stazione", active=True, created_by=admin_user.id,
+    )
+    db_session.add(station)
+    await db_session.commit()
+    version_id, first, _second = await _two_quotes(db_session, recipe.id)
+    supervisor, password = await _supervisor(db_session, username="capo-csv")
+    creds = {
+        "supervisor_username": supervisor.username, "supervisor_password": password,
+    }
+
+    run = (await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )).json()
+    bad = (await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT, run_id=run["id"],
+    )).json()
+    await client.post(
+        f"/api/measurements/{bad['id']}/authorise",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )
+
+    closed = (await client.post(
+        f"/api/production-runs/{run['id']}/close",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )).json()
+
+    text = (tmp_path / closed["statistics_path"]).read_text(encoding="utf-8-sig")
+    header, row = [line for line in text.splitlines() if line.strip()]
+    assert "authorised_by" in header and "authorised_at" in header
+    assert str(supervisor.id) in row
+
+
+async def test_a_fail_from_another_production_does_not_block_this_one(
+    client: AsyncClient, maker_user, measurement_tec_user, admin_user, db_session,
+    tmp_path, monkeypatch,
+):
+    """The run is the boundary: an old production must not stop a new one."""
+    # Closing a run writes the statistics file; pointed at tmp_path so the test
+    # does not leave one in the repository's uploads directory.
+    from src.backend.config import settings
+
+    monkeypatch.setattr(
+        type(settings), "upload_path", property(lambda self: tmp_path), raising=False,
+    )
+
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TOL-SCOPE")
+    station = Station(
+        code="ST-SCOPE", name="Stazione", active=True, created_by=admin_user.id,
+    )
+    db_session.add(station)
+    await db_session.commit()
+    version_id, first, second = await _two_quotes(db_session, recipe.id)
+
+    first_run = (await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )).json()
+    await _measure(
+        client, measurement_tec_user, first.id, version_id, OUT,
+        run_id=first_run["id"],
+    )
+
+    supervisor, password = await _supervisor(db_session, username="capo-scope")
+    await client.post(
+        f"/api/production-runs/{first_run['id']}/close",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": supervisor.username,
+            "supervisor_password": password,
+        },
+    )
+
+    second_run = (await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )).json()
+    resp = await _measure(
+        client, measurement_tec_user, second.id, version_id, GOOD,
+        run_id=second_run["id"],
+    )
+    assert resp.status_code == 200

src/backend/tests/test_production_runs.py

@@ -0,0 +1,747 @@
+"""Integration tests for /api/production-runs.
+
+The point of these tables is that a production outlives the page it was started
+from, so the tests lean on what must survive: the deadline, the cycle count, the
+trace of what happened, and the rule that a station has one production at a time.
+"""
+from datetime import datetime, timedelta
+
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from src.backend.models.orm.production import ProductionRun
+from src.backend.models.orm.recipe import Recipe
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import RecipeTask
+from src.backend.services import auth_service, production_service
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+async def _station(db_session, user_id: int, code: str = "ST-PROD") -> Station:
+    station = Station(code=code, name=f"Stazione {code}", active=True, created_by=user_id)
+    db_session.add(station)
+    await db_session.commit()
+    await db_session.refresh(station)
+    return station
+
+
+async def _recipe_with_interval(db_session, user_id: int, minutes, code="REC-PROD") -> Recipe:
+    recipe = await create_test_recipe(db_session, user_id=user_id, code=code)
+    recipe.measurement_interval_minutes = minutes
+    await db_session.commit()
+    await db_session.refresh(recipe)
+    return recipe
+
+
+async def _supervisor(db_session, username="capoturno", password="TurnoPwd1"):
+    user = await auth_service.create_user(
+        db_session,
+        username=username,
+        password=password,
+        display_name="Capo Turno",
+        roles=["Supervisor"],
+    )
+    await db_session.commit()
+    return user, password
+
+
+async def _advance_clock(db_session, run_id: int, delta: timedelta) -> None:
+    """Simulate wall-clock time passing, without sleeping in a test.
+
+    Winding every instant stored on the run back by D is indistinguishable from D
+    having elapsed. Moving only one of them would not be time passing - it would be
+    rewriting history, and would measure the wrong thing.
+    """
+    run = (await db_session.execute(
+        select(ProductionRun).where(ProductionRun.id == run_id)
+    )).scalar_one()
+    run.started_at = run.started_at - delta
+    if run.next_measurement_at is not None:
+        run.next_measurement_at = run.next_measurement_at - delta
+    if run.paused_at is not None:
+        run.paused_at = run.paused_at - delta
+    await db_session.commit()
+
+
+async def _only_task(db_session, version_id: int):
+    """The single task the fixture recipe comes with - a measurement one."""
+    row = await db_session.execute(
+        select(RecipeTask)
+        .where(RecipeTask.version_id == version_id)
+        .order_by(RecipeTask.order_index)
+    )
+    return row.scalars().first()
+
+
+async def _add_task(
+    db_session, version_id: int, order_index: int, title: str, task_type: str = "measure",
+) -> RecipeTask:
+    """A further task in the sequence, of a declared type."""
+    task = RecipeTask(
+        version_id=version_id, order_index=order_index, title=title, task_type=task_type,
+    )
+    db_session.add(task)
+    await db_session.commit()
+    await db_session.refresh(task)
+    return task
+
+
+async def _open(client, user, station, recipe, **extra):
+    return await client.post(
+        "/api/production-runs",
+        headers=auth_headers(user),
+        json={"station_code": station.code, "recipe_id": recipe.id, **extra},
+    )
+
+
+# ---------------------------------------------------------------------------
+# Opening
+# ---------------------------------------------------------------------------
+
+
+async def test_open_run_requires_auth(client: AsyncClient):
+    resp = await client.post("/api/production-runs", json={"station_code": "X", "recipe_id": 1})
+    assert resp.status_code == 401
+
+
+async def test_open_run_starts_the_countdown(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id)
+    recipe = await _recipe_with_interval(db_session, admin_user.id, minutes=30)
+
+    resp = await _open(client, measurement_tec_user, station, recipe)
+    assert resp.status_code == 201, resp.text
+    body = resp.json()
+    assert body["status"] == "running"
+    assert body["cycle_count"] == 0
+    assert body["measurement_interval_minutes"] == 30
+    # About half an hour out, allowing for the round trip.
+    assert 29 * 60 <= body["seconds_to_next_measurement"] <= 30 * 60
+    assert body["overdue"] is False
+
+
+async def test_open_run_without_interval_has_no_deadline(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-NOINT")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, None, code="REC-NOINT")
+
+    body = (await _open(client, measurement_tec_user, station, recipe)).json()
+    assert body["next_measurement_at"] is None
+    assert body["seconds_to_next_measurement"] is None
+    assert body["overdue"] is False
+
+
+async def test_open_run_snapshots_the_interval(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """Editing the recipe mid-production must not move a running deadline."""
+    station = await _station(db_session, admin_user.id, code="ST-SNAP")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 20, code="REC-SNAP")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+
+    recipe.measurement_interval_minutes = 90
+    await db_session.commit()
+
+    resp = await client.get(
+        f"/api/production-runs/{run_id}", headers=auth_headers(measurement_tec_user),
+    )
+    assert resp.json()["measurement_interval_minutes"] == 20
+
+
+async def test_station_cannot_have_two_open_runs(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-ONE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 15, code="REC-ONE")
+    first = await _open(client, measurement_tec_user, station, recipe)
+    assert first.status_code == 201
+
+    second = await _open(client, measurement_tec_user, station, recipe)
+    assert second.status_code == 409
+
+
+async def test_open_run_rejects_unknown_station(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 15, code="REC-NOSTAT")
+    resp = await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": "ST-GHOST", "recipe_id": recipe.id},
+    )
+    assert resp.status_code == 404
+
+
+# ---------------------------------------------------------------------------
+# Reading the current run - what replaces the in-page state
+# ---------------------------------------------------------------------------
+
+
+async def test_current_run_is_null_when_nothing_is_running(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-IDLE")
+    resp = await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )
+    assert resp.status_code == 200
+    assert resp.json() is None
+
+
+async def test_current_run_survives_and_keeps_counting_down(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """The deadline is stored, so a fresh page load rejoins the same countdown."""
+    station = await _station(db_session, admin_user.id, code="ST-KEEP")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 45, code="REC-KEEP")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+
+    later = await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )
+    body = later.json()
+    assert body["id"] == opened["id"]
+    assert body["next_measurement_at"] == opened["next_measurement_at"]
+    assert body["seconds_to_next_measurement"] <= opened["seconds_to_next_measurement"]
+
+
+# ---------------------------------------------------------------------------
+# Cycles
+# ---------------------------------------------------------------------------
+
+
+async def test_cycle_restarts_the_interval_and_counts(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-CYC")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-CYC")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+
+    resp = await client.post(
+        f"/api/production-runs/{run_id}/cycle", headers=auth_headers(measurement_tec_user),
+    )
+    assert resp.status_code == 200, resp.text
+    body = resp.json()
+    assert body["cycle_count"] == 1
+    assert 9 * 60 <= body["seconds_to_next_measurement"] <= 10 * 60
+
+
+async def test_measurement_tasks_travel_with_the_run(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """Where the measurement is cannot be guessed by the client: the run says so.
+
+    This is what lets a screen bring the operator back to the measurement when the
+    interval expires, whatever page they wandered off to.
+    """
+    station = await _station(db_session, admin_user.id, code="ST-SEQ")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-SEQ")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    first = await _only_task(db_session, opened["version_id"])
+    # A drawing between the two measurements, to prove it is left out.
+    await _add_task(db_session, opened["version_id"], 1, "Disegno", "drawing")
+    second = await _add_task(db_session, opened["version_id"], 2, "Seconda misura")
+
+    body = (await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )).json()
+    assert body["measurement_task_ids"] == [first.id, second.id]
+
+
+async def test_cycle_closes_only_on_the_last_measurement_task(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """Restarting at the first would measure three times as often as configured."""
+    station = await _station(db_session, admin_user.id, code="ST-MULTI")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-MULTI")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    run_id = opened["id"]
+    first = await _only_task(db_session, opened["version_id"])
+    second = await _add_task(db_session, opened["version_id"], 1, "Seconda misura")
+    headers = auth_headers(measurement_tec_user)
+
+    # Four minutes of the interval have gone by when the first task is finished.
+    await _advance_clock(db_session, run_id, timedelta(minutes=4))
+
+    mid = (await client.post(
+        f"/api/production-runs/{run_id}/cycle", headers=headers, json={"task_id": first.id},
+    )).json()
+    assert mid["cycle_count"] == 0, "il ciclo non e' ancora finito"
+    assert 5 * 60 <= mid["seconds_to_next_measurement"] <= 6 * 60, (
+        "l'intervallo non deve ripartire a meta' ciclo"
+    )
+
+    end = (await client.post(
+        f"/api/production-runs/{run_id}/cycle", headers=headers, json={"task_id": second.id},
+    )).json()
+    assert end["cycle_count"] == 1
+    assert 9 * 60 <= end["seconds_to_next_measurement"] <= 10 * 60
+
+    events = (await client.get(
+        f"/api/production-runs/{run_id}", headers=headers,
+    )).json()["events"]
+    by_type = {e["event_type"]: e for e in events}
+    assert by_type["task_measured"]["task_id"] == first.id
+    assert by_type["cycle_completed"]["task_id"] == second.id
+
+
+async def test_cycle_without_a_task_closes_as_before(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """A caller that knows nothing of the sequence keeps the old behaviour."""
+    station = await _station(db_session, admin_user.id, code="ST-PLAIN")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-PLAIN")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    await _add_task(db_session, opened["version_id"], 1, "Seconda misura")
+
+    body = (await client.post(
+        f"/api/production-runs/{opened['id']}/cycle",
+        headers=auth_headers(measurement_tec_user),
+    )).json()
+    assert body["cycle_count"] == 1
+
+
+# ---------------------------------------------------------------------------
+# Rimisura - turning the piece over
+# ---------------------------------------------------------------------------
+
+
+async def test_remeasure_does_not_buy_another_interval(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """A second reading of the same part is not a new cycle."""
+    station = await _station(db_session, admin_user.id, code="ST-AGAIN")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-AGAIN")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    run_id = opened["id"]
+    task = await _only_task(db_session, opened["version_id"])
+    headers = auth_headers(measurement_tec_user)
+
+    await _advance_clock(db_session, run_id, timedelta(minutes=7))
+
+    body = (await client.post(
+        f"/api/production-runs/{run_id}/remeasure", headers=headers,
+        json={"task_id": task.id},
+    )).json()
+    assert body["cycle_count"] == 0
+    assert 2 * 60 <= body["seconds_to_next_measurement"] <= 3 * 60, (
+        "rigirare il pezzo non deve spostare la scadenza"
+    )
+
+    events = (await client.get(
+        f"/api/production-runs/{run_id}", headers=headers,
+    )).json()["events"]
+    remeasured = [e for e in events if e["event_type"] == "remeasure"]
+    assert len(remeasured) == 1
+    assert remeasured[0]["task_id"] == task.id
+
+
+async def test_remeasure_refused_while_the_line_is_stopped(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-AGPAUSE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-AGPAUSE")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-again")
+
+    await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )
+    resp = await client.post(
+        f"/api/production-runs/{run_id}/remeasure",
+        headers=auth_headers(measurement_tec_user),
+    )
+    assert resp.status_code == 409
+
+
+async def test_overdue_run_reports_negative_seconds(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """Past the interval the countdown keeps going the other way, so the delay shows."""
+    station = await _station(db_session, admin_user.id, code="ST-LATE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 5, code="REC-LATE")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+
+    await _advance_clock(db_session, run_id, timedelta(minutes=8))
+
+    body = (await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )).json()
+    assert body["overdue"] is True
+    assert -190 <= body["seconds_to_next_measurement"] <= -170
+
+
+# ---------------------------------------------------------------------------
+# Fermo linea / ripresa / chiusura
+# ---------------------------------------------------------------------------
+
+
+async def test_pause_requires_supervisor_credentials(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-PAUSE1")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-PAUSE1")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+
+    resp = await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": "nobody", "supervisor_password": "wrong"},
+    )
+    assert resp.status_code == 401
+
+
+async def test_pause_rejects_non_supervisor(
+    client: AsyncClient, measurement_tec_user, maker_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-PAUSE2")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-PAUSE2")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+
+    resp = await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "supervisor_username": maker_user.username,
+            "supervisor_password": "testpassword123",
+        },
+    )
+    assert resp.status_code in (401, 403)
+
+
+async def test_pause_freezes_the_countdown(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """A fermo linea must not eat the measurement interval."""
+    station = await _station(db_session, admin_user.id, code="ST-FREEZE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 30, code="REC-FREEZE")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-freeze")
+
+    paused = (await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )).json()
+    assert paused["status"] == "paused"
+    frozen = paused["seconds_to_next_measurement"]
+
+    # An hour goes by with the line stopped: were the countdown still ticking, the
+    # run would now be deeply overdue.
+    await _advance_clock(db_session, run_id, timedelta(hours=1))
+
+    still = (await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )).json()
+    assert still["status"] == "paused"
+    assert still["seconds_to_next_measurement"] == frozen
+    assert still["overdue"] is False
+
+
+async def test_resume_gives_back_the_time_the_stop_took(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-RESUME")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 30, code="REC-RESUME")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-resume")
+    creds = {"supervisor_username": supervisor.username, "supervisor_password": password}
+
+    frozen = (await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )).json()["seconds_to_next_measurement"]
+
+    # The line stays down for ten minutes.
+    await _advance_clock(db_session, run_id, timedelta(minutes=10))
+
+    resumed = (await client.post(
+        f"/api/production-runs/{run_id}/resume",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )).json()
+    assert resumed["status"] == "running"
+    # The ten minutes of stop were handed back, not charged to the operator.
+    assert abs(resumed["seconds_to_next_measurement"] - frozen) <= 2
+
+
+async def test_cycle_refused_while_paused(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-PCYC")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-PCYC")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-pcyc")
+
+    await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )
+    resp = await client.post(
+        f"/api/production-runs/{run_id}/cycle", headers=auth_headers(measurement_tec_user),
+    )
+    assert resp.status_code == 409
+
+
+async def test_close_stops_the_timer_and_frees_the_station(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-CLOSE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-CLOSE")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-close")
+
+    closed = (await client.post(
+        f"/api/production-runs/{run_id}/close",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )).json()
+    assert closed["status"] == "closed"
+    assert closed["closed_by"] == supervisor.id
+    assert closed["seconds_to_next_measurement"] is None
+
+    # Nothing is open at the station any more...
+    current = await client.get(
+        "/api/production-runs/current",
+        params={"station_code": station.code},
+        headers=auth_headers(measurement_tec_user),
+    )
+    assert current.json() is None
+
+    # ...and a new production can start there.
+    again = await _open(client, measurement_tec_user, station, recipe)
+    assert again.status_code == 201
+
+
+async def test_closed_run_refuses_further_actions(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-DEAD")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-DEAD")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-dead")
+    creds = {"supervisor_username": supervisor.username, "supervisor_password": password}
+
+    await client.post(
+        f"/api/production-runs/{run_id}/close",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )
+    cycle = await client.post(
+        f"/api/production-runs/{run_id}/cycle", headers=auth_headers(measurement_tec_user),
+    )
+    assert cycle.status_code == 409
+    pause = await client.post(
+        f"/api/production-runs/{run_id}/pause",
+        headers=auth_headers(measurement_tec_user), json=creds,
+    )
+    assert pause.status_code == 409
+
+
+# ---------------------------------------------------------------------------
+# The trace
+# ---------------------------------------------------------------------------
+
+
+async def test_run_records_what_happened(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """The history a production never had: start, cycles, stop, resume, close."""
+    station = await _station(db_session, admin_user.id, code="ST-TRACE")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-TRACE")
+    run_id = (await _open(client, measurement_tec_user, station, recipe)).json()["id"]
+    supervisor, password = await _supervisor(db_session, username="capo-trace")
+    creds = {"supervisor_username": supervisor.username, "supervisor_password": password}
+    headers = auth_headers(measurement_tec_user)
+
+    await client.post(f"/api/production-runs/{run_id}/cycle", headers=headers)
+    await client.post(f"/api/production-runs/{run_id}/pause", headers=headers, json=creds)
+    await client.post(f"/api/production-runs/{run_id}/resume", headers=headers, json=creds)
+    await client.post(f"/api/production-runs/{run_id}/close", headers=headers, json=creds)
+
+    events = (await client.get(
+        f"/api/production-runs/{run_id}", headers=headers,
+    )).json()["events"]
+    assert [e["event_type"] for e in events] == [
+        "start", "cycle_completed", "line_stop", "resume", "close",
+    ]
+    # Who authorised what is on the record.
+    by_type = {e["event_type"]: e for e in events}
+    assert by_type["line_stop"]["supervisor_id"] == supervisor.id
+    assert by_type["close"]["supervisor_id"] == supervisor.id
+    assert by_type["cycle_completed"]["supervisor_id"] is None
+
+
+async def test_traceability_travels_with_the_run(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    station = await _station(db_session, admin_user.id, code="ST-LOT")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-LOT")
+    body = (await _open(
+        client, measurement_tec_user, station, recipe,
+        lot_number="LOT-42", serial_number="SN-7",
+    )).json()
+    assert body["lot_number"] == "LOT-42"
+    assert body["serial_number"] == "SN-7"
+
+
+# ---------------------------------------------------------------------------
+# Fine produzione: the statistics file (point 6)
+# ---------------------------------------------------------------------------
+
+
+async def _measure(client, user, run_id, subtask_id, version_id, value):
+    return await client.post(
+        "/api/measurements/",
+        headers=auth_headers(user),
+        json={
+            "subtask_id": subtask_id,
+            "version_id": version_id,
+            "value": value,
+            "lot_number": "LOT-STAT",
+            "production_run_id": run_id,
+        },
+    )
+
+
+async def _first_subtask(db_session, recipe_id: int):
+    from src.backend.models.orm.recipe import RecipeVersion
+    from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+
+    row = await db_session.execute(
+        select(RecipeSubtask)
+        .join(RecipeTask, RecipeTask.id == RecipeSubtask.task_id)
+        .join(RecipeVersion, RecipeVersion.id == RecipeTask.version_id)
+        .where(RecipeVersion.recipe_id == recipe_id)
+    )
+    return row.scalars().first()
+
+
+async def test_measurement_is_tied_to_the_run(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session,
+):
+    """Without this link 'the measurements of this production' is not a query."""
+    station = await _station(db_session, admin_user.id, code="ST-LINK")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-LINK")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    subtask = await _first_subtask(db_session, recipe.id)
+
+    resp = await _measure(
+        client, measurement_tec_user, opened["id"], subtask.id, opened["version_id"], 10.0,
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["production_run_id"] == opened["id"]
+
+
+async def test_close_emits_the_statistics_file(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session, tmp_path,
+    monkeypatch,
+):
+    from src.backend.config import settings
+    from src.backend.services import production_export_service
+
+    monkeypatch.setattr(
+        type(settings), "upload_path", property(lambda self: tmp_path), raising=False,
+    )
+
+    station = await _station(db_session, admin_user.id, code="ST-STAT")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-STAT")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    subtask = await _first_subtask(db_session, recipe.id)
+    await _measure(
+        client, measurement_tec_user, opened["id"], subtask.id, opened["version_id"], 10.0,
+    )
+    supervisor, password = await _supervisor(db_session, username="capo-stat")
+
+    closed = (await client.post(
+        f"/api/production-runs/{opened['id']}/close",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )).json()
+
+    assert closed["statistics_path"], "la chiusura deve produrre il file di statistica"
+    written = tmp_path / closed["statistics_path"]
+    assert written.exists()
+
+    text = written.read_text(encoding="utf-8-sig")
+    lines = [line for line in text.splitlines() if line.strip()]
+    assert len(lines) == 2, "intestazione piu' una misura"
+    assert "production_run_id" in lines[0]
+    # The tolerance limits travel with the row: a pass/fail must stay re-derivable.
+    for column in ("ltl", "lwl", "uwl", "utl", "nominal"):
+        assert column in lines[0]
+    assert "LOT-STAT" in lines[1]
+    assert str(opened["id"]) in lines[1]
+
+
+async def test_close_marks_measurements_as_exported(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session, tmp_path,
+    monkeypatch,
+):
+    from src.backend.config import settings
+    from src.backend.models.orm.measurement import Measurement
+
+    monkeypatch.setattr(
+        type(settings), "upload_path", property(lambda self: tmp_path), raising=False,
+    )
+
+    station = await _station(db_session, admin_user.id, code="ST-SYNC")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-SYNC")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    subtask = await _first_subtask(db_session, recipe.id)
+    created = (await _measure(
+        client, measurement_tec_user, opened["id"], subtask.id, opened["version_id"], 10.0,
+    )).json()
+    assert created["synced_to_csv"] is False
+
+    supervisor, password = await _supervisor(db_session, username="capo-sync")
+    await client.post(
+        f"/api/production-runs/{opened['id']}/close",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )
+
+    row = (await db_session.execute(
+        select(Measurement).where(Measurement.id == created["id"])
+    )).scalar_one()
+    await db_session.refresh(row)
+    assert row.synced_to_csv is True
+
+
+async def test_close_without_measurements_writes_no_file(
+    client: AsyncClient, measurement_tec_user, admin_user, db_session, tmp_path,
+    monkeypatch,
+):
+    """An empty file would be noise in the folder, not evidence."""
+    from src.backend.config import settings
+
+    monkeypatch.setattr(
+        type(settings), "upload_path", property(lambda self: tmp_path), raising=False,
+    )
+
+    station = await _station(db_session, admin_user.id, code="ST-EMPTY")
+    recipe = await _recipe_with_interval(db_session, admin_user.id, 10, code="REC-EMPTY")
+    opened = (await _open(client, measurement_tec_user, station, recipe)).json()
+    supervisor, password = await _supervisor(db_session, username="capo-empty")
+
+    closed = (await client.post(
+        f"/api/production-runs/{opened['id']}/close",
+        headers=auth_headers(measurement_tec_user),
+        json={"supervisor_username": supervisor.username, "supervisor_password": password},
+    )).json()
+    assert closed["statistics_path"] is None
+    assert not (tmp_path / "statistics").exists()

src/backend/tests/test_recipe_rules.py

@@ -0,0 +1,280 @@
+"""What a recipe is allowed to demand of the operator (points 8 and 9).
+
+Two rules that used to be nobody's: whether the lot and the serial are compulsory,
+and whether a value may be typed rather than read from the caliper. Both are
+declared on the recipe and enforced on the server, because a rule enforced only by
+the screen is a suggestion - the keypad can be hidden and the same request still
+sent by hand.
+"""
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from src.backend.models.orm.recipe import Recipe, RecipeVersion
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+async def _rules(db_session, recipe: Recipe, **flags) -> Recipe:
+    for key, value in flags.items():
+        setattr(recipe, key, value)
+    await db_session.commit()
+    await db_session.refresh(recipe)
+    return recipe
+
+
+async def _first_subtask(db_session, recipe_id: int) -> RecipeSubtask:
+    row = await db_session.execute(
+        select(RecipeSubtask)
+        .join(RecipeTask, RecipeTask.id == RecipeSubtask.task_id)
+        .join(RecipeVersion, RecipeVersion.id == RecipeTask.version_id)
+        .where(RecipeVersion.recipe_id == recipe_id)
+    )
+    return row.scalars().first()
+
+
+async def _current_version_id(db_session, recipe_id: int) -> int:
+    row = await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )
+    return row.scalar_one().id
+
+
+async def _measure(client, user, subtask_id, version_id, **extra):
+    return await client.post(
+        "/api/measurements/",
+        headers=auth_headers(user),
+        json={
+            "subtask_id": subtask_id, "version_id": version_id, "value": 10.0, **extra,
+        },
+    )
+
+
+# ---------------------------------------------------------------------------
+# The rules are part of the recipe
+# ---------------------------------------------------------------------------
+
+
+async def test_new_recipe_forbids_typing_by_default(client: AsyncClient, maker_user):
+    """The caliper is the instrument: typing is what has to be asked for."""
+    resp = await client.post(
+        "/api/recipes",
+        headers=auth_headers(maker_user),
+        json={"code": "REG-DEF", "name": "Regole di default"},
+    )
+    assert resp.status_code == 201, resp.text
+    body = resp.json()
+    assert body["allow_manual_input"] is False
+    assert body["requires_lot"] is False
+    assert body["requires_serial"] is False
+
+
+async def test_recipe_carries_the_rules_it_was_created_with(
+    client: AsyncClient, maker_user,
+):
+    resp = await client.post(
+        "/api/recipes",
+        headers=auth_headers(maker_user),
+        json={
+            "code": "REG-SET", "name": "Con regole",
+            "requires_lot": True, "requires_serial": True, "allow_manual_input": True,
+        },
+    )
+    body = resp.json()
+    assert (body["requires_lot"], body["requires_serial"], body["allow_manual_input"]) \
+        == (True, True, True)
+
+
+async def test_rules_can_be_switched_off_again(
+    client: AsyncClient, maker_user, db_session,
+):
+    """False must reach the database: a flag that only ever turns on is a trap."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="REG-OFF")
+    await _rules(db_session, recipe, requires_lot=True, allow_manual_input=True)
+
+    resp = await client.put(
+        f"/api/recipes/{recipe.id}",
+        headers=auth_headers(maker_user),
+        json={"requires_lot": False, "allow_manual_input": False},
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["requires_lot"] is False
+    assert resp.json()["allow_manual_input"] is False
+
+
+async def test_rules_survive_a_new_version(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """Editing a recipe with measurements copies it; the rules must come along."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="REG-VER")
+    await _rules(db_session, recipe, requires_lot=True)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    # A measurement on the current version forces copy-on-write on the next edit.
+    await _measure(
+        client, measurement_tec_user, subtask.id, version_id, lot_number="L-1",
+    )
+    resp = await client.put(
+        f"/api/recipes/{recipe.id}",
+        headers=auth_headers(maker_user),
+        json={"name": "Rinominata"},
+    )
+    assert resp.status_code == 200
+    assert resp.json()["requires_lot"] is True
+
+
+# ---------------------------------------------------------------------------
+# Point 8 - traceability
+# ---------------------------------------------------------------------------
+
+
+async def test_measurement_without_the_required_lot_is_refused(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TRC-LOT")
+    await _rules(db_session, recipe, requires_lot=True)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    refused = await _measure(client, measurement_tec_user, subtask.id, version_id)
+    assert refused.status_code == 422
+
+    # Blank is not a lot number either.
+    blank = await _measure(
+        client, measurement_tec_user, subtask.id, version_id, lot_number="   ",
+    )
+    assert blank.status_code == 422
+
+    accepted = await _measure(
+        client, measurement_tec_user, subtask.id, version_id, lot_number="LOT-9",
+    )
+    assert accepted.status_code == 200
+
+
+async def test_measurement_without_the_required_serial_is_refused(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TRC-SER")
+    await _rules(db_session, recipe, requires_serial=True)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    assert (await _measure(
+        client, measurement_tec_user, subtask.id, version_id,
+    )).status_code == 422
+    assert (await _measure(
+        client, measurement_tec_user, subtask.id, version_id, serial_number="SN-1",
+    )).status_code == 200
+
+
+async def test_optional_traceability_still_accepts_nothing(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """Recipes that do not ask for a lot must go on working as they did."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TRC-OPT")
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    assert (await _measure(
+        client, measurement_tec_user, subtask.id, version_id,
+    )).status_code == 200
+
+
+async def test_production_cannot_start_without_the_required_lot(
+    client: AsyncClient, maker_user, measurement_tec_user, admin_user, db_session,
+):
+    """The lot has to be there before the first value, not after it."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="TRC-RUN")
+    await _rules(db_session, recipe, requires_lot=True)
+    station = Station(
+        code="ST-TRC", name="Stazione", active=True, created_by=admin_user.id,
+    )
+    db_session.add(station)
+    await db_session.commit()
+
+    refused = await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )
+    assert refused.status_code == 422
+
+    started = await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "station_code": station.code, "recipe_id": recipe.id, "lot_number": "L-7",
+        },
+    )
+    assert started.status_code == 201
+
+
+# ---------------------------------------------------------------------------
+# Point 9 - typed values
+# ---------------------------------------------------------------------------
+
+
+async def test_typed_value_is_refused_when_the_recipe_forbids_it(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="MAN-NO")
+    await _rules(db_session, recipe, allow_manual_input=False)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    refused = await _measure(
+        client, measurement_tec_user, subtask.id, version_id, input_method="manual",
+    )
+    assert refused.status_code == 422
+    assert "caliper" in refused.json()["detail"].lower()
+
+
+async def test_the_caliper_is_always_welcome(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="MAN-USB")
+    await _rules(db_session, recipe, allow_manual_input=False)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    resp = await _measure(
+        client, measurement_tec_user, subtask.id, version_id,
+        input_method="usb_caliper",
+    )
+    assert resp.status_code == 200
+    assert resp.json()["input_method"] == "usb_caliper"
+
+
+async def test_typing_is_allowed_where_the_recipe_says_so(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id, code="MAN-YES")
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    resp = await _measure(
+        client, measurement_tec_user, subtask.id, version_id, input_method="manual",
+    )
+    assert resp.status_code == 200
+
+
+async def test_the_default_input_method_is_covered_by_the_rule(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """A request that says nothing about how the value was entered counts as typed.
+
+    Otherwise omitting the field would be the way around the rule.
+    """
+    recipe = await create_test_recipe(db_session, maker_user.id, code="MAN-DEF")
+    await _rules(db_session, recipe, allow_manual_input=False)
+    subtask = await _first_subtask(db_session, recipe.id)
+    version_id = await _current_version_id(db_session, recipe.id)
+
+    assert (await _measure(
+        client, measurement_tec_user, subtask.id, version_id,
+    )).status_code == 422

src/backend/tests/test_stations_api.py

@@ -201,3 +201,176 @@ async def test_duplicate_assignment_returns_409(
         json={"recipe_id": recipe.id},
     )
     assert second.status_code == 409
+
+
+# ---------------------------------------------------------------------------
+# Station list carries assigned recipes (point 7)
+# ---------------------------------------------------------------------------
+
+
+async def test_list_stations_includes_assigned_recipes(
+    client: AsyncClient, admin_user, db_session,
+):
+    """The admin list shows which recipes a station handles, without a call per row."""
+    recipe = await create_test_recipe(db_session, user_id=admin_user.id, code="REC-LIST")
+    await db_session.commit()
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-WITHREC", "name": "Con ricette"},
+    )
+    sid = created.json()["id"]
+    await client.post(
+        f"/api/stations/{sid}/recipes",
+        headers=auth_headers(admin_user),
+        json={"recipe_id": recipe.id},
+    )
+
+    resp = await client.get("/api/stations", headers=auth_headers(admin_user))
+    assert resp.status_code == 200
+    station = next(s for s in resp.json() if s["id"] == sid)
+    assert [r["code"] for r in station["recipes"]] == ["REC-LIST"]
+
+
+async def test_list_stations_empty_recipes_when_unassigned(
+    client: AsyncClient, admin_user,
+):
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-NOREC", "name": "Senza ricette"},
+    )
+    sid = created.json()["id"]
+    resp = await client.get("/api/stations", headers=auth_headers(admin_user))
+    station = next(s for s in resp.json() if s["id"] == sid)
+    assert station["recipes"] == []
+
+
+async def test_list_stations_hides_inactive_recipes(
+    client: AsyncClient, admin_user, db_session,
+):
+    """An assigned but deactivated recipe must not show: the operator won't see it."""
+    recipe = await create_test_recipe(db_session, user_id=admin_user.id, code="REC-OFF")
+    await db_session.commit()
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-INACT", "name": "Con ricetta spenta"},
+    )
+    sid = created.json()["id"]
+    await client.post(
+        f"/api/stations/{sid}/recipes",
+        headers=auth_headers(admin_user),
+        json={"recipe_id": recipe.id},
+    )
+
+    recipe.active = False
+    await db_session.commit()
+
+    resp = await client.get("/api/stations", headers=auth_headers(admin_user))
+    station = next(s for s in resp.json() if s["id"] == sid)
+    assert station["recipes"] == []
+
+
+# ---------------------------------------------------------------------------
+# Per-row station reset (point 7)
+# ---------------------------------------------------------------------------
+
+
+async def test_reset_station_removes_all_assignments(
+    client: AsyncClient, admin_user, db_session,
+):
+    first = await create_test_recipe(db_session, user_id=admin_user.id, code="REC-R1")
+    second = await create_test_recipe(db_session, user_id=admin_user.id, code="REC-R2")
+    await db_session.commit()
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-RESET", "name": "Da azzerare"},
+    )
+    sid = created.json()["id"]
+    for recipe in (first, second):
+        await client.post(
+            f"/api/stations/{sid}/recipes",
+            headers=auth_headers(admin_user),
+            json={"recipe_id": recipe.id},
+        )
+
+    resp = await client.delete(
+        f"/api/stations/{sid}/recipes", headers=auth_headers(admin_user),
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json() == {"station_id": sid, "removed": 2}
+
+    left = await client.get(
+        f"/api/stations/{sid}/recipes", headers=auth_headers(admin_user),
+    )
+    assert left.json() == []
+
+
+async def test_reset_station_keeps_the_station(
+    client: AsyncClient, admin_user, db_session,
+):
+    """Reset clears assignments only - the station stays and can be reassigned."""
+    recipe = await create_test_recipe(db_session, user_id=admin_user.id, code="REC-KEEP")
+    await db_session.commit()
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-KEEP", "name": "Resta"},
+    )
+    sid = created.json()["id"]
+    await client.post(
+        f"/api/stations/{sid}/recipes",
+        headers=auth_headers(admin_user),
+        json={"recipe_id": recipe.id},
+    )
+    await client.delete(f"/api/stations/{sid}/recipes", headers=auth_headers(admin_user))
+
+    still_there = await client.get(
+        f"/api/stations/{sid}", headers=auth_headers(admin_user),
+    )
+    assert still_there.status_code == 200
+
+    reassigned = await client.post(
+        f"/api/stations/{sid}/recipes",
+        headers=auth_headers(admin_user),
+        json={"recipe_id": recipe.id},
+    )
+    assert reassigned.status_code == 201
+
+
+async def test_reset_empty_station_is_idempotent(client: AsyncClient, admin_user):
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-IDEM", "name": "Gia vuota"},
+    )
+    sid = created.json()["id"]
+    resp = await client.delete(
+        f"/api/stations/{sid}/recipes", headers=auth_headers(admin_user),
+    )
+    assert resp.status_code == 200
+    assert resp.json()["removed"] == 0
+
+
+async def test_reset_station_non_admin_is_403(
+    client: AsyncClient, admin_user, maker_user,
+):
+    created = await client.post(
+        "/api/stations",
+        headers=auth_headers(admin_user),
+        json={"code": "ST-RBAC", "name": "Protetta"},
+    )
+    sid = created.json()["id"]
+    resp = await client.delete(
+        f"/api/stations/{sid}/recipes", headers=auth_headers(maker_user),
+    )
+    assert resp.status_code == 403
+
+
+async def test_reset_unknown_station_is_404(client: AsyncClient, admin_user):
+    resp = await client.delete(
+        "/api/stations/999999/recipes", headers=auth_headers(admin_user),
+    )
+    assert resp.status_code == 404

src/backend/tests/test_task_progress.py

@@ -0,0 +1,246 @@
+"""Point 10: a task left half done has to look different from one never opened.
+
+The task list showed both the same way, so an operator coming back to a recipe
+had to remember where they had got to - or measure something twice. These tests
+are about the count the list now puts on each card.
+"""
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.station import Station
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+async def _version_and_task(db_session, recipe_id: int):
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    return version, task
+
+
+async def _add_quote(db_session, task_id: int, marker: int):
+    quote = RecipeSubtask(
+        task_id=task_id, marker_number=marker, description=f"Quota {marker}",
+        nominal=10.0, utl=10.5, uwl=10.3, lwl=9.7, ltl=9.5, unit="mm",
+    )
+    db_session.add(quote)
+    await db_session.commit()
+    await db_session.refresh(quote)
+    return quote
+
+
+async def _measure(client, user, subtask_id, version_id, value=10.0):
+    return await client.post(
+        "/api/measurements/", headers=auth_headers(user),
+        json={"subtask_id": subtask_id, "version_id": version_id, "value": value},
+    )
+
+
+async def _progress(client, user, version_id):
+    resp = await client.get(
+        "/api/measurements/task-progress",
+        headers=auth_headers(user),
+        params={"version_id": version_id},
+    )
+    assert resp.status_code == 200
+    return {row["task_id"]: row for row in resp.json()["tasks"]}
+
+
+async def test_untouched_task_reads_as_none(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[task.id]["state"] == "none"
+    assert rows[task.id]["measured"] == 0
+
+
+async def test_half_measured_task_reads_as_partial(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    await _add_quote(db_session, task.id, 2)
+
+    await _measure(client, measurement_tec_user, first.id, version.id)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[task.id] == {
+        "task_id": task.id, "quotes": 2, "measured": 1, "state": "partial",
+    }
+
+
+async def test_every_quote_taken_reads_as_complete(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    second = await _add_quote(db_session, task.id, 2)
+
+    await _measure(client, measurement_tec_user, first.id, version.id)
+    await _measure(client, measurement_tec_user, second.id, version.id)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[task.id]["state"] == "complete"
+    assert rows[task.id]["measured"] == 2
+
+
+async def test_measuring_the_same_quote_twice_does_not_inflate_the_count(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    """Two readings of one quote are one quote done, not two.
+
+    Re-measuring is allowed - it is how an operator recovers from a caliper that
+    slipped - and counting attempts would show a task of two quotes as finished
+    after one of them had been measured twice.
+    """
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    await _add_quote(db_session, task.id, 2)
+
+    await _measure(client, measurement_tec_user, first.id, version.id, 10.0)
+    await _measure(client, measurement_tec_user, first.id, version.id, 10.1)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[task.id]["measured"] == 1
+    assert rows[task.id]["state"] == "partial"
+
+
+async def test_documental_tasks_are_left_out(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    """A note cannot be half read; carrying it as unfinished for ever is noise."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    note = RecipeTask(
+        version_id=version.id, order_index=5, title="Leggere prima",
+        task_type="note",
+    )
+    db_session.add(note)
+    await db_session.commit()
+    await db_session.refresh(note)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert note.id not in rows
+    assert task.id in rows
+
+
+async def test_a_measurement_task_without_quotes_is_not_reported_as_missing(
+    client: AsyncClient, db_session, measurement_tec_user,
+):
+    """Declared as a measurement, no quotes entered yet: nothing to take."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, _task = await _version_and_task(db_session, recipe.id)
+    empty = RecipeTask(
+        version_id=version.id, order_index=6, title="Da compilare",
+        task_type="measure",
+    )
+    db_session.add(empty)
+    await db_session.commit()
+    await db_session.refresh(empty)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[empty.id]["state"] == "empty"
+    assert rows[empty.id]["quotes"] == 0
+
+
+async def test_another_operators_work_is_not_counted_as_yours(
+    client: AsyncClient, db_session, measurement_tec_user, admin_user,
+):
+    """Outside a production the boundary is the operator, as it is everywhere else."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+
+    await _measure(client, admin_user, first.id, version.id)
+
+    rows = await _progress(client, measurement_tec_user, version.id)
+    assert rows[task.id]["state"] == "none"
+
+
+async def test_inside_a_production_the_run_is_the_boundary(
+    client: AsyncClient, db_session, measurement_tec_user, admin_user,
+):
+    """A quote measured before the production opened is not part of it.
+
+    The list is read during the run, and carrying yesterday's readings into it
+    would show a task as finished that this production has not touched.
+    """
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task = await _version_and_task(db_session, recipe.id)
+    first = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+
+    # Measured outside any production.
+    await _measure(client, measurement_tec_user, first.id, version.id)
+
+    station = Station(
+        code="ST-PROG", name="Stazione", active=True, created_by=admin_user.id,
+    )
+    db_session.add(station)
+    await db_session.commit()
+    run = (await client.post(
+        "/api/production-runs",
+        headers=auth_headers(measurement_tec_user),
+        json={"station_code": station.code, "recipe_id": recipe.id},
+    )).json()
+
+    resp = await client.get(
+        "/api/measurements/task-progress",
+        headers=auth_headers(measurement_tec_user),
+        params={"version_id": version.id, "production_run_id": run["id"]},
+    )
+    assert resp.status_code == 200
+    rows = {row["task_id"]: row for row in resp.json()["tasks"]}
+    assert rows[task.id]["state"] == "none"
+
+    await client.post(
+        "/api/measurements/", headers=auth_headers(measurement_tec_user),
+        json={
+            "subtask_id": first.id, "version_id": version.id, "value": 10.0,
+            "production_run_id": run["id"],
+        },
+    )
+    resp = await client.get(
+        "/api/measurements/task-progress",
+        headers=auth_headers(measurement_tec_user),
+        params={"version_id": version.id, "production_run_id": run["id"]},
+    )
+    rows = {row["task_id"]: row for row in resp.json()["tasks"]}
+    assert rows[task.id]["state"] == "complete"
+
+
+async def test_progress_needs_a_measurement_role(
+    client: AsyncClient, db_session, maker_user, measurement_tec_user,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, _task = await _version_and_task(db_session, recipe.id)
+
+    resp = await client.get(
+        "/api/measurements/task-progress",
+        headers=auth_headers(maker_user),
+        params={"version_id": version.id},
+    )
+    assert resp.status_code == 403

src/backend/tests/test_task_type.py

@@ -0,0 +1,171 @@
+"""Tests for the explicit task type (point 2).
+
+The type used to be deduced - quotes present meant a measurement - so a measurement
+task whose quotes had not been entered yet was treated as a note, and the system
+behaved differently depending on how complete the recipe was. These tests pin the
+type down as something declared and carried.
+"""
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import RecipeTask
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+async def _current_version(db_session, recipe_id: int) -> RecipeVersion:
+    row = await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )
+    return row.scalar_one()
+
+
+async def test_task_defaults_to_note(client: AsyncClient, maker_user, db_session):
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-T1")
+    await db_session.commit()
+
+    resp = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": "Solo una nota"},
+    )
+    assert resp.status_code == 201, resp.text
+    assert resp.json()["task_type"] == "note"
+
+
+@pytest.mark.parametrize("declared", ["note", "measure", "drawing"])
+async def test_task_keeps_the_declared_type(
+    client: AsyncClient, maker_user, db_session, declared,
+):
+    recipe = await create_test_recipe(
+        db_session, user_id=maker_user.id, code=f"REC-{declared}",
+    )
+    await db_session.commit()
+
+    resp = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": f"Task {declared}", "task_type": declared},
+    )
+    assert resp.status_code == 201, resp.text
+    assert resp.json()["task_type"] == declared
+
+
+async def test_measure_task_without_quotes_is_still_a_measure_task(
+    client: AsyncClient, maker_user, db_session,
+):
+    """The criterion from the document: an unfinished recipe must not change kind."""
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-NOQ")
+    await db_session.commit()
+
+    created = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": "Misura da compilare", "task_type": "measure"},
+    )
+    task_id = created.json()["id"]
+
+    fetched = await client.get(f"/api/tasks/{task_id}", headers=auth_headers(maker_user))
+    body = fetched.json()
+    assert body["subtasks"] == [], "il task non ha ancora quote"
+    assert body["task_type"] == "measure", "e resta comunque un task di misura"
+
+
+async def test_unknown_type_is_rejected(client: AsyncClient, maker_user, db_session):
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-BAD")
+    await db_session.commit()
+
+    resp = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": "x", "task_type": "qualcosa"},
+    )
+    assert resp.status_code == 422
+
+
+async def test_type_can_be_changed(client: AsyncClient, maker_user, db_session):
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-CHG")
+    await db_session.commit()
+    created = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": "Nasce nota", "task_type": "note"},
+    )
+    task_id = created.json()["id"]
+
+    resp = await client.put(
+        f"/api/tasks/{task_id}",
+        headers=auth_headers(maker_user),
+        json={"task_type": "measure"},
+    )
+    assert resp.status_code == 200, resp.text
+    assert resp.json()["task_type"] == "measure"
+
+
+async def test_type_survives_copy_on_write_versioning(
+    client: AsyncClient, maker_user, measurement_tec_user, db_session,
+):
+    """A new version must not silently reclassify every task of the recipe."""
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-COW")
+    await db_session.commit()
+
+    version = await _current_version(db_session, recipe.id)
+    row = await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )
+    first_task = row.scalars().first()
+
+    # Declare it a drawing. The point is that this survives, not how it got set.
+    await client.put(
+        f"/api/tasks/{first_task.id}",
+        headers=auth_headers(maker_user),
+        json={"task_type": "drawing"},
+    )
+
+    # Measure on the current version, which forces the next edit to copy-on-write.
+    await client.post(
+        "/api/measurements/",
+        headers=auth_headers(measurement_tec_user),
+        json={
+            "subtask_id": first_task.subtasks[0].id,
+            "version_id": version.id,
+            "value": 10.0,
+        },
+    )
+
+    # Editing now must create a new version.
+    await client.put(
+        f"/api/recipes/{recipe.id}",
+        headers=auth_headers(maker_user),
+        json={"name": "Rinominata"},
+    )
+
+    new_version = await _current_version(db_session, recipe.id)
+    assert new_version.id != version.id, "l'edit doveva creare una nuova versione"
+
+    copied = await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == new_version.id)
+    )
+    types = sorted(t.task_type for t in copied.scalars().all())
+    assert "drawing" in types, "il tipo dichiarato non e' sopravvissuto alla copia"
+
+
+async def test_task_list_exposes_the_type(client: AsyncClient, maker_user, db_session):
+    """The operator's task list reads the type, so it must come over the wire."""
+    recipe = await create_test_recipe(db_session, user_id=maker_user.id, code="REC-LST")
+    await db_session.commit()
+    await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={"title": "Misura", "task_type": "measure"},
+    )
+
+    resp = await client.get(
+        f"/api/recipes/{recipe.id}/tasks", headers=auth_headers(maker_user),
+    )
+    assert resp.status_code == 200
+    assert all("task_type" in t for t in resp.json())

src/backend/tests/test_vision_boundary.py

@@ -0,0 +1,106 @@
+"""Il server non importa VisionSuite, e questo lo verifica.
+
+Senza qualcuno che lo controlli, il confine si perde in silenzio: basta un
+import comodo perché l'immagine dell'API torni da cinque gigabyte e il motivo
+per cui il worker esiste svanisca senza che nessuna prova diventi rossa.
+"""
+import ast
+from pathlib import Path
+
+BACKEND = Path(__file__).resolve().parents[1]
+
+# Full dotted prefixes, not bare roots: "src" alone would forbid every import
+# the backend legitimately makes of itself (src.backend.*), so the entries
+# for the vision tree are multi-segment ("src.vision", "src.vision_worker")
+# while VisionSuite's own top-level packages stay single-segment.
+#
+# M2: "Acquire" used to sit here as a stand-in for vs-camera, but it is a
+# class/SDK name (Balluff's "mvIMPACT Acquire" SDK), never an import root -
+# it could never match. vs-camera imports under the same `visionsuite`
+# namespace as vs-core (see
+# vendor/visionsuite/packages/vs-camera/pyproject.toml: "il codice si importa
+# come visionsuite.camera...."), which "visionsuite" below already forbids,
+# so there is nothing separate to add for it.
+FORBIDDEN = {
+    "visionsuite", "pm2d", "dxf_compare", "torch",
+    "src.vision", "src.vision_worker",
+}
+
+
+def _imported_modules(source: Path) -> set[str]:
+    """Full dotted module paths this file imports - not just their first
+    segment. `from src.vision.runner import ...` must be checked against
+    "src.vision.runner" (inside the forbidden "src.vision"), not against
+    "src" alone, which is never forbidden since the backend imports itself
+    constantly.
+    """
+    tree = ast.parse(source.read_text(encoding="utf-8"), filename=str(source))
+    modules: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            modules.update(alias.name for alias in node.names)
+        elif isinstance(node, ast.ImportFrom) and node.module and node.level == 0:
+            modules.add(node.module)
+    return modules
+
+
+def _forbidden_hits(modules: set[str]) -> set[str]:
+    """Which FORBIDDEN entries a set of imported module paths triggers.
+
+    A module counts as forbidden if it equals a forbidden entry exactly, or
+    lives inside it ("src.vision.runner" is inside "src.vision").
+    """
+    hits: set[str] = set()
+    for module in modules:
+        for entry in FORBIDDEN:
+            if module == entry or module.startswith(entry + "."):
+                hits.add(entry)
+    return hits
+
+
+def test_the_server_never_imports_visionsuite():
+    offenders = []
+    for source in BACKEND.rglob("*.py"):
+        if "tests" in source.parts or "migrations" in source.parts:
+            continue
+        forbidden = _forbidden_hits(_imported_modules(source))
+        if forbidden:
+            offenders.append(f"{source.relative_to(BACKEND)}: {sorted(forbidden)}")
+
+    assert offenders == [], (
+        "il server deve restare cieco alla visione; usa il worker via HTTP:\n"
+        + "\n".join(offenders)
+    )
+
+
+def test_the_submodule_is_pinned():
+    """Server e stazione devono montare lo stesso commit: qui c'è la fonte."""
+    gitmodules = Path(__file__).resolve().parents[3] / ".gitmodules"
+    assert gitmodules.exists(), "vendor/visionsuite non è un sottomodulo"
+    assert "vendor/visionsuite" in gitmodules.read_text(encoding="utf-8")
+
+
+def test_a_full_module_path_import_is_caught_not_just_its_root(tmp_path):
+    """M2: `from src.vision.runner import ...` inside src/backend/** must be
+    caught. Checking only the first dotted segment ("src") would miss it -
+    "src" itself is never forbidden, the backend imports itself all the time.
+    This is the blind spot the reviewer found: same repository, no install
+    needed, the most tempting shortcut of all.
+    """
+    offender = tmp_path / "sneaky.py"
+    offender.write_text("from src.vision.runner import run_graph\n")
+    assert _forbidden_hits(_imported_modules(offender)) == {"src.vision"}
+
+
+def test_a_vision_worker_import_is_also_caught(tmp_path):
+    offender = tmp_path / "sneaky.py"
+    offender.write_text("import src.vision_worker.main\n")
+    assert _forbidden_hits(_imported_modules(offender)) == {"src.vision_worker"}
+
+
+def test_an_unrelated_src_backend_import_is_not_caught(tmp_path):
+    """The fix must not turn every self-import of the backend into an
+    offender - only the vision subtrees are forbidden."""
+    innocent = tmp_path / "innocent.py"
+    innocent.write_text("from src.backend.config import settings\n")
+    assert _forbidden_hits(_imported_modules(innocent)) == set()

src/backend/tests/test_vision_copy_on_write.py

@@ -0,0 +1,167 @@
+"""C1: the vision graph must survive copy-on-write, and creation must save it.
+
+The spec assumed versioning was free for `vision_json`/`vision_output` - "un
+campo in piu nella copia, non una logica nuova". `_copy_tasks_to_version`
+copies an explicit field list that omitted both fields, so the first edit to
+a recipe that already has measurements silently drops the camera graph from
+the new version. The creation paths (`POST /api/recipes/{id}/tasks` and
+`POST /api/tasks/{id}/subtasks`) have the same shape of bug: `TaskCreate` and
+`SubtaskCreate` declare the fields, but the routers never pass them to the
+ORM constructors, so the API answers 200 and silently does not store them.
+
+Each test file here reads on its own - the same idiom `test_vision_graph.py`
+and `test_recipes.py` already use.
+"""
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.models.orm.user import User
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+async def _seed_measurement(db: AsyncSession, recipe_id: int, user_id: int) -> None:
+    """Give the current version a measurement, so an update takes the
+    copy-on-write path instead of updating in place."""
+    version_id = (await db.execute(
+        select(RecipeVersion.id).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    subtask_id = (await db.execute(
+        select(RecipeSubtask.id)
+        .join(RecipeSubtask.task)
+        .where(RecipeSubtask.task.has(version_id=version_id))
+    )).scalar_one()
+    db.add(Measurement(
+        subtask_id=subtask_id,
+        version_id=version_id,
+        measured_by=user_id,
+        value=10.0,
+        pass_fail="pass",
+        deviation=0.0,
+        input_method="manual",
+    ))
+    await db.flush()
+
+
+async def _current_task_and_subtask(db: AsyncSession, recipe_id: int):
+    version = (await db.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    subtask = (await db.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    return version, task, subtask
+
+
+@pytest.mark.asyncio
+async def test_copy_on_write_preserves_the_vision_graph(
+    client: AsyncClient, db_session: AsyncSession, maker_user: User,
+):
+    """A recipe edit that triggers copy-on-write must not drop the camera graph."""
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    _, task, subtask = await _current_task_and_subtask(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.flush()
+
+    # A measurement on v1 forces the next PUT onto the copy-on-write path.
+    await _seed_measurement(db_session, recipe.id, maker_user.id)
+    db_session.expunge_all()
+
+    resp = await client.put(
+        f"/api/recipes/{recipe.id}",
+        headers=auth_headers(maker_user),
+        json={"change_notes": "Edit that must not drop the graph"},
+    )
+    assert resp.status_code == 200
+    db_session.expunge_all()
+
+    version, new_task, new_subtask = await _current_task_and_subtask(db_session, recipe.id)
+    assert version.version_number == 2, "the edit should have created v2"
+    assert new_task.vision_json == {"schema_version": 1, "name": "p", "tools": []}
+    assert new_subtask.vision_output == "t1.diameter"
+
+
+@pytest.mark.asyncio
+async def test_creating_a_task_persists_its_vision_graph(
+    client: AsyncClient, db_session: AsyncSession, maker_user: User,
+):
+    """POST /api/recipes/{id}/tasks must store vision_json and vision_output,
+    not just accept and echo them back."""
+    recipe = await create_test_recipe(db_session, maker_user.id)
+
+    resp = await client.post(
+        f"/api/recipes/{recipe.id}/tasks",
+        headers=auth_headers(maker_user),
+        json={
+            "title": "Camera task",
+            "task_type": "camera_measure",
+            "vision_json": {"schema_version": 1, "name": "p", "tools": []},
+            "subtasks": [
+                {
+                    "marker_number": 1,
+                    "description": "Diameter",
+                    "nominal": 10.0,
+                    "vision_output": "t1.diameter",
+                }
+            ],
+        },
+    )
+    assert resp.status_code == 201
+    body = resp.json()
+    assert body["vision_json"] == {"schema_version": 1, "name": "p", "tools": []}
+    assert body["subtasks"][0]["vision_output"] == "t1.diameter"
+
+    # Reload from the database - the response reflects the ORM object right
+    # after flush, so this catches a constructor that dropped the field.
+    db_session.expunge_all()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.id == body["id"])
+    )).scalar_one()
+    sub = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().one()
+    assert task.vision_json == {"schema_version": 1, "name": "p", "tools": []}
+    assert sub.vision_output == "t1.diameter"
+
+
+@pytest.mark.asyncio
+async def test_creating_a_subtask_persists_its_vision_output(
+    client: AsyncClient, db_session: AsyncSession, maker_user: User,
+):
+    """POST /api/tasks/{id}/subtasks (the standalone endpoint) must also store it."""
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    _, task, _ = await _current_task_and_subtask(db_session, recipe.id)
+
+    resp = await client.post(
+        f"/api/tasks/{task.id}/subtasks",
+        headers=auth_headers(maker_user),
+        json={
+            "marker_number": 2,
+            "description": "Radius",
+            "nominal": 5.0,
+            "vision_output": "t1.radius",
+        },
+    )
+    assert resp.status_code == 201
+    assert resp.json()["vision_output"] == "t1.radius"
+
+    db_session.expunge_all()
+    sub = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.marker_number == 2)
+    )).scalar_one()
+    assert sub.vision_output == "t1.radius"

src/backend/tests/test_vision_execute.py

@@ -0,0 +1,329 @@
+"""A quote from a camera is a measurement like any other: same verdict, same gate."""
+import io
+
+import pytest
+from PIL import Image
+from sqlalchemy import select
+
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.models.orm.vision import VisionResult
+from src.backend.services import measurement_service
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+GOOD = 10.0     # inside every limit of the fixture quote
+OUT = 99.0      # far outside
+
+
+def an_image() -> bytes:
+    buffer = io.BytesIO()
+    Image.new("L", (64, 64), color=0).save(buffer, format="PNG")
+    return buffer.getvalue()
+
+
+async def recipe_parts(db_session, recipe_id: int):
+    """`create_test_recipe` returns only the recipe: here it is broken apart."""
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    subtask = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    return version, task, subtask
+
+
+async def two_vision_quotes(db_session, recipe_id: int):
+    """The fixture recipe's task, with a second quote mapped to a vision output.
+
+    Same shape as `test_out_of_tolerance.py::_two_quotes`, adapted so both
+    quotes belong to a single camera acquisition instead of two separate
+    manual readings.
+    """
+    version, task, first = await recipe_parts(db_session, recipe_id)
+    first.vision_output = "t1.diameter"
+
+    second = RecipeSubtask(
+        task_id=task.id, marker_number=2, description="Second quote",
+        nominal=10.0, utl=10.5, uwl=10.3, lwl=9.7, ltl=9.5, unit="mm",
+        vision_output="t1.radius",
+    )
+    db_session.add(second)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    await db_session.commit()
+    await db_session.refresh(second)
+    return version, task, first, second
+
+
+@pytest.fixture
+def worker_says(monkeypatch):
+    """The worker does not run in tests: its response is faked here."""
+    def _install(outputs: dict, failures: list | None = None):
+        async def fake_call(image_bytes: bytes, graph: dict) -> dict:
+            return {
+                "outputs": outputs,
+                "failures": failures or [],
+                "engine_version": "a" * 40,
+                "duration_ms": 12,
+            }
+        monkeypatch.setattr(
+            "src.backend.services.vision_service._call_worker", fake_call,
+        )
+    return _install
+
+
+@pytest.mark.asyncio
+async def test_a_camera_quote_becomes_a_measurement(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    # The fixture creates the quote with nominal 10 and tight limits around it.
+    await db_session.commit()
+    worker_says({"t1.diameter": float(subtask.nominal)})
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id), "lot_number": "L1"},
+    )
+
+    assert response.status_code == 200
+    saved = (await db_session.execute(
+        select(Measurement).where(Measurement.subtask_id == subtask.id)
+    )).scalars().all()
+    assert len(saved) == 1
+    assert saved[0].pass_fail == "pass"
+    assert saved[0].input_method == "camera"
+    assert saved[0].vision_result_id is not None
+
+
+@pytest.mark.asyncio
+async def test_the_execution_records_which_engine_produced_it(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.commit()
+    worker_says({"t1.diameter": float(subtask.nominal)})
+
+    await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id), "lot_number": "L1"},
+    )
+
+    result = (await db_session.execute(select(VisionResult))).scalars().one()
+    assert result.engine_version == "a" * 40
+    assert result.executed_on == "server"
+    assert result.graph_snapshot == task.vision_json
+
+
+@pytest.mark.asyncio
+async def test_a_missing_output_names_itself(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    """No half measurement: if an output is missing, the error names it."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.commit()
+    worker_says({"t1.radius": 5.0})
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id), "lot_number": "L1"},
+    )
+
+    assert response.status_code == 422
+    assert "t1.diameter" in response.json()["detail"]
+    assert (await db_session.execute(select(Measurement))).scalars().all() == []
+
+
+@pytest.mark.asyncio
+async def test_a_task_without_a_graph_is_refused(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    await db_session.commit()
+    worker_says({})
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 422
+    assert "graph" in response.json()["detail"].lower()
+
+
+# ---------------------------------------------------------------------------
+# Ruling R10: the gate is evaluated once per acquisition, not once per quote.
+# A camera acquisition produces every quote of the task at the same instant -
+# there is no "moving on" between them, so the sequential gate must not fire
+# inside the batch. It still has to fire between two separate acquisitions.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_a_non_last_out_of_tolerance_quote_does_not_break_the_batch(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    """The quote evaluated first fails; the one after it must still be saved."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, first, second = await two_vision_quotes(db_session, recipe.id)
+    worker_says({"t1.diameter": OUT, "t1.radius": GOOD})
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 200
+    by_subtask = {
+        m.subtask_id: m.pass_fail
+        for m in (await db_session.execute(select(Measurement))).scalars().all()
+    }
+    assert by_subtask == {first.id: "fail", second.id: "pass"}
+
+
+@pytest.mark.asyncio
+async def test_a_pending_fail_outside_the_batch_still_blocks(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    """A fail left over from something else this camera task does not cover."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, first, second = await two_vision_quotes(db_session, recipe.id)
+
+    # A third quote on the same task, not part of this acquisition (no vision
+    # output), already failed and is waiting on a supervisor.
+    third = RecipeSubtask(
+        task_id=task.id, marker_number=3, description="Not a vision quote",
+        nominal=10.0, utl=10.5, uwl=10.3, lwl=9.7, ltl=9.5, unit="mm",
+    )
+    db_session.add(third)
+    await db_session.commit()
+    await db_session.refresh(third)
+    await measurement_service.save_measurement(
+        db_session, subtask_id=third.id, version_id=version.id,
+        measured_by=measurement_tec_user.id, value=OUT,
+    )
+    await db_session.commit()
+
+    worker_says({"t1.diameter": GOOD, "t1.radius": GOOD})
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 409
+    assert "supervisor" in response.json()["detail"].lower()
+    assert (await db_session.execute(
+        select(Measurement).where(Measurement.subtask_id.in_([first.id, second.id]))
+    )).scalars().all() == []
+
+
+@pytest.mark.asyncio
+async def test_a_pending_fail_blocks_before_the_worker_is_called(
+    client, db_session, measurement_tec_user, monkeypatch,
+):
+    """M3: a blocked operator must not burn a whole vision execution to learn
+    they are blocked. The gate is checked before the worker runs, not after."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, first, second = await two_vision_quotes(db_session, recipe.id)
+
+    third = RecipeSubtask(
+        task_id=task.id, marker_number=3, description="Not a vision quote",
+        nominal=10.0, utl=10.5, uwl=10.3, lwl=9.7, ltl=9.5, unit="mm",
+    )
+    db_session.add(third)
+    await db_session.commit()
+    await db_session.refresh(third)
+    await measurement_service.save_measurement(
+        db_session, subtask_id=third.id, version_id=version.id,
+        measured_by=measurement_tec_user.id, value=OUT,
+    )
+    await db_session.commit()
+
+    async def worker_must_not_be_called(image_bytes: bytes, graph: dict) -> dict:
+        raise AssertionError("the worker must not run while the operator is blocked")
+    monkeypatch.setattr(
+        "src.backend.services.vision_service._call_worker", worker_must_not_be_called,
+    )
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 409
+    assert "supervisor" in response.json()["detail"].lower()
+
+
+@pytest.mark.asyncio
+async def test_a_new_acquisition_can_remeasure_its_own_pending_quote(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    """Re-shooting the same task is the way out, same as re-measuring by hand."""
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, first, second = await two_vision_quotes(db_session, recipe.id)
+
+    worker_says({"t1.diameter": OUT, "t1.radius": GOOD})
+    first_shot = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p1.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+    assert first_shot.status_code == 200
+
+    # first.id is now pending authorisation - but it belongs to this same
+    # task, so a new acquisition of it is not "moving on".
+    worker_says({"t1.diameter": GOOD, "t1.radius": GOOD})
+    second_shot = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p2.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+    assert second_shot.status_code == 200
+
+    readings = (await db_session.execute(
+        select(Measurement)
+        .where(Measurement.subtask_id == first.id)
+        .order_by(Measurement.id)
+    )).scalars().all()
+    assert [m.pass_fail for m in readings] == ["fail", "pass"]
+
+    assert await measurement_service.pending_authorisation(
+        db_session, version.id, measurement_tec_user.id,
+    ) is None

src/backend/tests/test_vision_graph.py

@@ -0,0 +1,71 @@
+"""Il grafo di visione vive sul task, e la quota sa quale uscita la riempie."""
+import pytest
+from sqlalchemy import select
+
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import TASK_TYPES, RecipeSubtask, RecipeTask
+from src.backend.tests.conftest import create_test_recipe
+
+
+async def recipe_parts(db_session, recipe_id: int):
+    """`create_test_recipe` restituisce la sola ricetta: qui si scompone.
+
+    Versione corrente, il suo unico task, la sua prima quota. È l'idioma che
+    `test_out_of_tolerance.py` usa già, ripetuto perché ogni file di prova si
+    legge da solo.
+    """
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    subtask = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    return version, task, subtask
+
+
+@pytest.mark.asyncio
+async def test_task_carries_a_vision_graph(db_session, maker_user):
+    """Il grafo si salva e si rilegge identico: il server non lo interpreta."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="VIS-1")
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    graph = {
+        "schema_version": 1,
+        "name": "diametro del foro",
+        "tools": [{"id": "t1", "kind": "DetectCircle", "name": "foro"}],
+    }
+    task.vision_json = graph
+    task.task_type = "camera_measure"
+    await db_session.flush()
+    db_session.expunge_all()
+
+    reloaded = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.id == task.id)
+    )).scalar_one()
+    assert reloaded.vision_json == graph
+
+
+@pytest.mark.asyncio
+async def test_subtask_names_the_output_that_fills_it(db_session, maker_user):
+    """La quota dichiara quale uscita del grafo le compete."""
+    recipe = await create_test_recipe(db_session, maker_user.id, code="VIS-2")
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    subtask.vision_output = "t1.diameter"
+    await db_session.flush()
+    db_session.expunge_all()
+
+    reloaded = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.id == subtask.id)
+    )).scalar_one()
+    assert reloaded.vision_output == "t1.diameter"
+
+
+def test_the_typo_is_gone():
+    """`xf_compare` era un refuso: si corregge finché nessuna riga lo usa."""
+    assert "dxf_compare" in TASK_TYPES
+    assert "xf_compare" not in TASK_TYPES

src/backend/tests/test_vision_reference.py

@@ -0,0 +1,110 @@
+"""The Maker composes at a desk: image saved, graph tried, no measurement."""
+import io
+
+import pytest
+from PIL import Image
+from sqlalchemy import select
+
+from src.backend.config import settings
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.models.orm.vision import VisionReferenceImage
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+def an_image() -> bytes:
+    buffer = io.BytesIO()
+    Image.new("L", (64, 64), color=0).save(buffer, format="PNG")
+    return buffer.getvalue()
+
+
+async def recipe_parts(db_session, recipe_id: int):
+    """`create_test_recipe` returns only the recipe: here it is broken apart."""
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    subtask = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    return version, task, subtask
+
+
+@pytest.fixture
+def worker_says(monkeypatch):
+    def _install(outputs: dict):
+        async def fake_call(image_bytes: bytes, graph: dict) -> dict:
+            return {"outputs": outputs, "failures": [],
+                    "engine_version": "b" * 40, "duration_ms": 7}
+        monkeypatch.setattr(
+            "src.backend.services.vision_service._call_worker", fake_call,
+        )
+    return _install
+
+
+@pytest.fixture(autouse=True)
+def redirect_uploads(monkeypatch, tmp_path):
+    """Keep reference images out of the real repo tree during tests.
+
+    `settings.upload_path` is `parents[2] / self.upload_dir`, and pathlib
+    replaces the left operand of `/` when the right one is absolute - so
+    redirecting the bare `upload_dir` setting to pytest's absolute `tmp_path`
+    is enough to redirect the resolved path too.
+    """
+    monkeypatch.setattr(settings, "upload_dir", str(tmp_path))
+
+
+@pytest.mark.asyncio
+async def test_a_reference_image_keeps_its_provenance(
+    client, db_session, maker_user,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    await db_session.commit()
+
+    response = await client.post(
+        "/api/vision/reference-images",
+        headers=auth_headers(maker_user),
+        files={"image": ("ref.png", an_image(), "image/png")},
+        data={"task_id": str(task.id), "note": "sample part"},
+    )
+
+    assert response.status_code == 200
+    saved = (await db_session.execute(
+        select(VisionReferenceImage)
+    )).scalars().one()
+    assert saved.task_id == task.id
+    assert saved.note == "sample part"
+    assert saved.acquired_at is not None
+    assert saved.path
+
+
+@pytest.mark.asyncio
+async def test_preview_runs_the_graph_and_saves_no_measurement(
+    client, db_session, maker_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.commit()
+    worker_says({"t1.diameter": 42.0})
+
+    response = await client.post(
+        "/api/vision/preview",
+        headers=auth_headers(maker_user),
+        files={"image": ("ref.png", an_image(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 200
+    assert response.json()["outputs"]["t1.diameter"] == 42.0
+    # This is the whole point of this test: composing does not measure.
+    assert (await db_session.execute(select(Measurement))).scalars().all() == []

src/backend/tests/test_vision_upload_limits.py

@@ -0,0 +1,190 @@
+"""I3: the three vision endpoints must not accept an unbounded or wrong-typed
+upload. `/execute`, `/preview` and `/reference-images` read the whole image
+into memory with no size cap, and `/reference-images` wrote whatever arrived
+to disk with a `.png` extension regardless of its real type - a PDF is not
+something a vision graph can measure.
+
+Reuses the project's existing convention (`validate_file_size`,
+`ALLOWED_IMAGE_TYPES` from `files.py`) instead of inventing a second one.
+"""
+import io
+
+import pytest
+from PIL import Image
+from sqlalchemy import select
+
+from src.backend.config import settings
+from src.backend.models.orm.measurement import Measurement
+from src.backend.models.orm.recipe import RecipeVersion
+from src.backend.models.orm.task import RecipeSubtask, RecipeTask
+from src.backend.models.orm.vision import VisionReferenceImage, VisionResult
+from src.backend.tests.conftest import auth_headers, create_test_recipe
+
+
+@pytest.fixture(autouse=True)
+def redirect_uploads(monkeypatch, tmp_path):
+    """Keep the reference image this file stores out of the real repo tree.
+
+    The same guard `test_vision_reference.py` carries: without it, the test that
+    proves a jpeg is stored with its own extension leaves a real file under
+    `uploads/vision/reference/`, so every suite run dirties the working tree.
+
+    `settings.upload_path` is `parents[2] / self.upload_dir`, and pathlib
+    replaces the left operand of `/` when the right one is absolute - so
+    redirecting the bare `upload_dir` setting is enough to redirect the
+    resolved path too.
+    """
+    monkeypatch.setattr(settings, "upload_dir", str(tmp_path))
+
+
+def a_png() -> bytes:
+    buffer = io.BytesIO()
+    Image.new("L", (64, 64), color=0).save(buffer, format="PNG")
+    return buffer.getvalue()
+
+
+def a_jpeg() -> bytes:
+    buffer = io.BytesIO()
+    Image.new("RGB", (64, 64), color=0).save(buffer, format="JPEG")
+    return buffer.getvalue()
+
+
+async def recipe_parts(db_session, recipe_id: int):
+    """`create_test_recipe` returns only the recipe: here it is broken apart."""
+    version = (await db_session.execute(
+        select(RecipeVersion).where(
+            RecipeVersion.recipe_id == recipe_id,
+            RecipeVersion.is_current == True,  # noqa: E712
+        )
+    )).scalar_one()
+    task = (await db_session.execute(
+        select(RecipeTask).where(RecipeTask.version_id == version.id)
+    )).scalars().first()
+    subtask = (await db_session.execute(
+        select(RecipeSubtask).where(RecipeSubtask.task_id == task.id)
+    )).scalars().first()
+    return version, task, subtask
+
+
+@pytest.fixture
+def worker_says(monkeypatch):
+    def _install(outputs: dict):
+        async def fake_call(image_bytes: bytes, graph: dict) -> dict:
+            return {"outputs": outputs, "failures": [],
+                    "engine_version": "c" * 40, "duration_ms": 5}
+        monkeypatch.setattr(
+            "src.backend.services.vision_service._call_worker", fake_call,
+        )
+    return _install
+
+
+@pytest.mark.asyncio
+async def test_execute_rejects_a_non_image_upload(
+    client, db_session, measurement_tec_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.commit()
+    worker_says({"t1.diameter": float(subtask.nominal)})
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.pdf", b"%PDF-1.4 not an image", "application/pdf")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 400
+    assert (await db_session.execute(select(VisionResult))).scalars().all() == []
+    assert (await db_session.execute(select(Measurement))).scalars().all() == []
+
+
+@pytest.mark.asyncio
+async def test_execute_rejects_an_oversized_upload(
+    client, db_session, measurement_tec_user, worker_says, monkeypatch,
+):
+    recipe = await create_test_recipe(db_session, measurement_tec_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    subtask.vision_output = "t1.diameter"
+    await db_session.commit()
+    worker_says({"t1.diameter": float(subtask.nominal)})
+    monkeypatch.setattr(settings, "max_upload_size_mb", 0)
+
+    response = await client.post(
+        "/api/vision/execute",
+        headers=auth_headers(measurement_tec_user),
+        files={"image": ("p.png", a_png(), "image/png")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 400
+    assert (await db_session.execute(select(VisionResult))).scalars().all() == []
+
+
+@pytest.mark.asyncio
+async def test_preview_rejects_a_non_image_upload(
+    client, db_session, maker_user, worker_says,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    version, task, subtask = await recipe_parts(db_session, recipe.id)
+    task.task_type = "camera_measure"
+    task.vision_json = {"schema_version": 1, "name": "p", "tools": []}
+    await db_session.commit()
+    worker_says({"t1.diameter": 10.0})
+
+    response = await client.post(
+        "/api/vision/preview",
+        headers=auth_headers(maker_user),
+        files={"image": ("p.pdf", b"%PDF-1.4 not an image", "application/pdf")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 400
+
+
+@pytest.mark.asyncio
+async def test_reference_images_rejects_a_non_image_upload(
+    client, db_session, maker_user,
+):
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    _, task, _ = await recipe_parts(db_session, recipe.id)
+    await db_session.commit()
+
+    response = await client.post(
+        "/api/vision/reference-images",
+        headers=auth_headers(maker_user),
+        files={"image": ("p.pdf", b"%PDF-1.4 not an image", "application/pdf")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 400
+    assert (await db_session.execute(
+        select(VisionReferenceImage)
+    )).scalars().all() == []
+
+
+@pytest.mark.asyncio
+async def test_reference_image_keeps_the_extension_matching_its_real_type(
+    client, db_session, maker_user,
+):
+    """A jpeg upload must not be written to disk as `<uuid>.png`."""
+    recipe = await create_test_recipe(db_session, maker_user.id)
+    _, task, _ = await recipe_parts(db_session, recipe.id)
+    await db_session.commit()
+
+    response = await client.post(
+        "/api/vision/reference-images",
+        headers=auth_headers(maker_user),
+        files={"image": ("p.jpg", a_jpeg(), "image/jpeg")},
+        data={"task_id": str(task.id)},
+    )
+
+    assert response.status_code == 200
+    saved = (await db_session.execute(select(VisionReferenceImage))).scalars().one()
+    assert saved.path.endswith(".jpg")
+    assert (settings.upload_path / saved.path).exists()

src/backend/tests/test_vision_worker_transport.py

@@ -0,0 +1,117 @@
+"""I2: a broken worker must answer with a named status, never a naked 500.
+
+`_call_worker` mapped a non-200 HTTP response to 502, but let
+`httpx.ConnectError` (worker stopped) and `httpx.ReadTimeout` (job past the
+120s budget) propagate unhandled - with no exception handler in main.py that
+turns those into anything but FastAPI's generic 500. It also read
+`report["outputs"]`, `report["failures"]` and `report["engine_version"]`
+without checking the response had that shape at all, so a proxy answering 200
+with something else (HTML, an empty body) would raise a bare KeyError instead
+of a diagnosable error.
+
+The spec's edge-case table: "il worker isola il guasto, l'API resta in
+piedi, il task riporta l'errore" - the worker's fault must not become the
+API's fault.
+"""
+import httpx
+import pytest
+from fastapi import HTTPException
+
+from src.backend.services.vision_service import _call_worker
+
+
+class _FakeResponse:
+    """Stands in for httpx.Response: only what `_call_worker` touches."""
+
+    def __init__(self, status_code: int, body=..., text: str = "", raise_on_json: bool = False):
+        self.status_code = status_code
+        self._body = body
+        self.text = text
+        self._raise_on_json = raise_on_json
+
+    def json(self):
+        if self._raise_on_json:
+            raise ValueError("not JSON")
+        return self._body
+
+
+class _FakeAsyncClient:
+    """Stands in for httpx.AsyncClient: post() does whatever the test wants."""
+
+    def __init__(self, *, response=None, raises=None):
+        self._response = response
+        self._raises = raises
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *exc_info):
+        return False
+
+    async def post(self, *args, **kwargs):
+        if self._raises is not None:
+            raise self._raises
+        return self._response
+
+
+def _install(monkeypatch, *, response=None, raises=None):
+    monkeypatch.setattr(
+        "src.backend.services.vision_service.httpx.AsyncClient",
+        lambda *a, **k: _FakeAsyncClient(response=response, raises=raises),
+    )
+
+
+@pytest.mark.asyncio
+async def test_a_stopped_worker_answers_502_not_a_naked_500(monkeypatch):
+    _install(monkeypatch, raises=httpx.ConnectError("Connection refused"))
+
+    with pytest.raises(HTTPException) as excinfo:
+        await _call_worker(b"img", {"schema_version": 1, "tools": []})
+
+    assert excinfo.value.status_code == 502
+    assert "unreachable" in excinfo.value.detail.lower()
+
+
+@pytest.mark.asyncio
+async def test_a_job_past_the_time_budget_answers_504(monkeypatch):
+    _install(monkeypatch, raises=httpx.ReadTimeout("timed out"))
+
+    with pytest.raises(HTTPException) as excinfo:
+        await _call_worker(b"img", {"schema_version": 1, "tools": []})
+
+    assert excinfo.value.status_code == 504
+    assert "time" in excinfo.value.detail.lower()
+
+
+@pytest.mark.asyncio
+async def test_a_response_missing_the_expected_keys_is_reported_not_raised_raw(monkeypatch):
+    """A proxy answering 200 with an unexpected shape must not KeyError."""
+    _install(monkeypatch, response=_FakeResponse(200, body={"outputs": {}}))
+
+    with pytest.raises(HTTPException) as excinfo:
+        await _call_worker(b"img", {"schema_version": 1, "tools": []})
+
+    assert excinfo.value.status_code == 502
+    assert "malformed" in excinfo.value.detail.lower()
+
+
+@pytest.mark.asyncio
+async def test_a_response_that_is_not_a_json_object_is_reported(monkeypatch):
+    """200 with HTML or a bare list, say - not a dict at all."""
+    _install(monkeypatch, response=_FakeResponse(200, body=["not", "a", "dict"]))
+
+    with pytest.raises(HTTPException) as excinfo:
+        await _call_worker(b"img", {"schema_version": 1, "tools": []})
+
+    assert excinfo.value.status_code == 502
+    assert "malformed" in excinfo.value.detail.lower()
+
+
+@pytest.mark.asyncio
+async def test_a_well_formed_response_still_passes_through(monkeypatch):
+    body = {"outputs": {"t1.diameter": 10.0}, "failures": [], "engine_version": "a" * 40}
+    _install(monkeypatch, response=_FakeResponse(200, body=body))
+
+    report = await _call_worker(b"img", {"schema_version": 1, "tools": []})
+
+    assert report == body

src/frontend/flask_app/app.py

@@ -1,15 +1,44 @@
 """TieMeasureFlow Client - Flask Entry Point."""
 import json
 import os
+import re
+from datetime import date
+from urllib.parse import urlparse
 
 from flask import Flask, redirect, url_for, session, request
 from flask_babel import Babel
 from flask_wtf.csrf import CSRFProtect
-from markupsafe import Markup
+from markupsafe import Markup, escape
 from werkzeug.middleware.proxy_fix import ProxyFix
 
 from config import Config
 
+# **bold**, the whole of the markup a task description understands besides the
+# line break. DOTALL so a phrase that wraps onto the next line still closes.
+_BOLD_RX = re.compile(r"\*\*(.+?)\*\*", re.DOTALL)
+
+# Punto 12. Nothing loads from outside: every library ships with the install
+# (static/vendor/), because the shop floor network has no way out and a page that
+# waits for a CDN there is a white screen, not a slow one.
+#
+#   unsafe-inline   the templates carry their Alpine components inline
+#   unsafe-eval     Alpine 3 and Plotly both compile expressions at runtime
+#   blob:           PDF.js runs its worker from a blob, Fabric exports canvases
+#   data:           thumbnails and canvas exports are inlined
+CSP = (
+    "default-src 'self'; "
+    "script-src 'self' 'unsafe-inline' 'unsafe-eval'; "
+    "style-src 'self' 'unsafe-inline'; "
+    "font-src 'self'; "
+    "img-src 'self' data: blob:; "
+    "connect-src 'self'; "
+    "worker-src 'self' blob:; "
+    "object-src 'none'; "
+    "base-uri 'self'; "
+    "form-action 'self'; "
+    "frame-ancestors 'none'"
+)
+
 
 def get_locale():
     """Get user's preferred language from session or Accept-Language header."""
@@ -63,7 +92,32 @@ def create_app() -> Flask:
         """Set user's preferred language and store in session."""
         if lang in Config.LANGUAGES:
             session["language"] = lang
-        return redirect(request.referrer or url_for("auth.login"))
+        # Only follow the referrer if it points back to this host
+        # (prevents open redirect via a forged Referer header).
+        referrer = request.referrer
+        if referrer:
+            parsed = urlparse(referrer)
+            if parsed.netloc and parsed.netloc != request.host:
+                referrer = None
+        return redirect(referrer or url_for("auth.login"))
+
+    @app.template_filter("rich_text")
+    def rich_text_filter(value):
+        """Render a task description keeping its line breaks and its bold.
+
+        Whoever writes a recipe pastes from the PDF of the technical sheet, and the
+        text used to arrive flattened. Two conventions carry it: a blank line is a
+        line break, **like this** is bold.
+
+        No HTML is ever stored or trusted - the text is escaped first and the only
+        tags in the result are the ones produced here. That is the sanitisation:
+        there is nothing to sanitise, because nothing is accepted.
+        """
+        if not value:
+            return Markup("")
+        escaped = str(escape(str(value)))
+        bolded = _BOLD_RX.sub(r"<strong>\1</strong>", escaped)
+        return Markup(bolded.replace("\n", "<br>"))
 
     @app.template_filter("tojson_attr")
     def tojson_attr_filter(value):
@@ -83,6 +137,25 @@ def create_app() -> Flask:
         )
         return Markup(rv)
 
+    @app.after_request
+    def security_headers(response):
+        """Same-origin only, on the pages the browser actually loads.
+
+        The policy existed on the backend, which serves the API; the HTML comes
+        from here and carried no policy at all. So the rule that says «nothing
+        from outside» was written where it could not be broken and absent where
+        it could.
+
+        It is also the guard that makes point 12 stay done: a CDN tag added to a
+        template months from now is refused by the browser here, at a desk with a
+        network, instead of on a shop floor that has none.
+        """
+        response.headers.setdefault("Content-Security-Policy", CSP)
+        response.headers.setdefault("X-Content-Type-Options", "nosniff")
+        response.headers.setdefault("X-Frame-Options", "DENY")
+        response.headers.setdefault("Referrer-Policy", "strict-origin-when-cross-origin")
+        return response
+
     @app.context_processor
     def inject_globals():
         """Inject global variables into all templates."""
@@ -92,6 +165,8 @@ def create_app() -> Flask:
             "current_language": get_locale(),
             "languages": Config.LANGUAGES,
             "company_logo": session.get("company_logo"),
+            "auto_logout_minutes": session.get("auto_logout_minutes"),
+            "current_year": date.today().year,
         }
 
     return app

src/frontend/flask_app/blueprints/admin.py

@@ -43,6 +43,28 @@ def user_list():
     return render_template("admin/users.html", users=users)
 
 
+@admin_bp.route("/settings")
+@login_required
+@admin_required
+def settings_page():
+    """System settings page."""
+    resp = api_client.get("/api/settings")
+    settings = resp if not isinstance(resp, dict) or not resp.get("error") else {}
+    return render_template("admin/settings.html", settings=settings)
+
+
+@admin_bp.route("/api/settings", methods=["PUT"])
+@login_required
+@admin_required
+def api_update_settings():
+    """Proxy: Update system settings."""
+    data = request.get_json(silent=True) or {}
+    resp = api_client.put("/api/settings", data=data)
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
 @admin_bp.route("/stations")
 @login_required
 @admin_required
@@ -198,3 +220,14 @@ def api_unassign_recipe(station_id: int, recipe_id: int):
     if isinstance(resp, dict) and resp.get("error"):
         return jsonify(resp), resp.get("status_code", 500)
     return jsonify({"deleted": True}), 200
+
+
+@admin_bp.route("/api/stations/<int:station_id>/recipes", methods=["DELETE"])
+@login_required
+@admin_required
+def api_reset_station_recipes(station_id: int):
+    """Proxy: Clear every recipe assignment of a station."""
+    resp = api_client.delete(f"/api/stations/{station_id}/recipes")
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200

src/frontend/flask_app/blueprints/auth.py

@@ -1,7 +1,7 @@
 """Authentication blueprint - login, logout, profile."""
 from functools import wraps
 
-from flask import Blueprint, abort, flash, redirect, render_template, request, session, url_for
+from flask import Blueprint, abort, flash, jsonify, redirect, render_template, request, session, url_for
 from flask_babel import gettext as _
 
 from services.api_client import api_client
@@ -39,15 +39,32 @@ def role_required(*roles):
     return decorator
 
 
+def _post_login_landing():
+    """Pick the landing endpoint based on the logged-in user's roles.
+
+    A Maker-only user must not be bounced to /measure (MeasurementTec-only),
+    which would raise a 403 right after login. Order of preference follows the
+    primary operator flow first, then editing, analysis, admin.
+    """
+    user = session.get("user", {})
+    roles = user.get("roles", [])
+    if "MeasurementTec" in roles:
+        return url_for("measure.select_recipe")
+    if "Maker" in roles:
+        return url_for("maker.recipe_list")
+    if "Metrologist" in roles:
+        return url_for("statistics.dashboard")
+    if user.get("is_admin"):
+        return url_for("admin.user_list")
+    return url_for("auth.profile")
+
+
 @auth_bp.route("/login", methods=["GET", "POST"])
 def login():
     """Login page and form handler."""
-    # Se già loggato, redirect a home
+    # Se già loggato, redirect alla home in base al ruolo
     if session.get("api_key"):
-        try:
-            return redirect(url_for("measure.select_recipe"))
-        except Exception:
-            return redirect(url_for("auth.profile"))
+        return redirect(_post_login_landing())
 
     if request.method == "POST":
         username = request.form.get("username", "").strip()
@@ -76,20 +93,23 @@ def login():
             if user.get("theme_pref"):
                 session["theme"] = user["theme_pref"]
 
-            # Carica logo aziendale dalle impostazioni di sistema
+            # Carica impostazioni di sistema
             settings_resp = api_client.get("/api/settings")
             if not settings_resp.get("error"):
                 logo_path = settings_resp.get("company_logo_path")
                 if logo_path:
                     session["company_logo"] = logo_path
+                auto_logout = settings_resp.get("auto_logout_minutes")
+                if auto_logout:
+                    try:
+                        session["auto_logout_minutes"] = int(auto_logout)
+                    except (ValueError, TypeError):
+                        pass
 
             flash(_("Benvenuto, %(name)s!", name=user.get("display_name", username)), "success")
 
-            # Redirect dopo login
-            try:
-                return redirect(url_for("measure.select_recipe"))
-            except Exception:
-                return redirect(url_for("auth.profile"))
+            # Redirect dopo login (in base al ruolo)
+            return redirect(_post_login_landing())
 
         except Exception as e:
             flash(_("Errore di connessione al server: %(error)s", error=str(e)), "error")
@@ -113,6 +133,27 @@ def logout():
     return redirect(url_for("auth.login"))
 
 
+@auth_bp.route("/set-theme", methods=["POST"])
+@login_required
+def set_theme():
+    """Persist the user's theme toggle (light/dark) to session and profile.
+
+    Called via AJAX from the navbar theme switch so the choice survives
+    navigation and matches on the next login. Best-effort on the API side:
+    the session is updated regardless so the current browsing stays correct.
+    """
+    theme = (request.get_json(silent=True) or {}).get("theme", "")
+    if theme not in ("light", "dark"):
+        return jsonify({"error": True, "detail": "invalid theme"}), 400
+
+    session["theme"] = theme
+    try:
+        api_client.put("/api/auth/me", data={"theme_pref": theme})
+    except Exception:
+        pass  # Non-fatal: session theme is already updated for this session.
+    return jsonify({"ok": True, "theme": theme})
+
+
 @auth_bp.route("/profile", methods=["GET", "POST"])
 @login_required
 def profile():

src/frontend/flask_app/blueprints/maker.py

@@ -1,12 +1,13 @@
 """Maker blueprint - recipe creation and editing."""
 import requests as http_requests
 
-from flask import Blueprint, Response, flash, jsonify, redirect, render_template, request, session, url_for
+from flask import Blueprint, flash, jsonify, redirect, render_template, request, session, url_for
 from flask_babel import gettext as _
 
 from blueprints.auth import login_required, role_required
 from config import Config
 from services.api_client import api_client
+from services.file_proxy import proxy_file
 
 maker_bp = Blueprint("maker", __name__, url_prefix="/maker")
 
@@ -380,19 +381,7 @@ def api_upload_file():
 @login_required
 def api_get_file(file_path: str):
     """Proxy: Serve file from API server (browser can't send X-API-Key)."""
-    api_key = session.get("api_key", "")
-    base_url = Config.API_SERVER_URL.rstrip("/")
-    resp = http_requests.get(
-        f"{base_url}/api/files/{file_path}",
-        headers={"X-API-Key": api_key},
-        timeout=30,
-    )
-    if resp.status_code != 200:
-        return Response(resp.text, status=resp.status_code)
-    return Response(
-        resp.content,
-        content_type=resp.headers.get("content-type", "application/octet-stream"),
-    )
+    return proxy_file(file_path)
 
 
 @maker_bp.route("/api/files/<path:file_path>", methods=["DELETE"])
@@ -424,3 +413,31 @@ def api_get_measurement_count(recipe_id: int, version_number: int):
         return jsonify(resp), resp.get("status_code", 500)
 
     return jsonify(resp), 200
+
+
+@maker_bp.route("/api/recipes/<int:recipe_id>/parse-technical-sheet", methods=["POST"])
+@login_required
+@role_required("Maker")
+def api_parse_technical_sheet(recipe_id: int):
+    """Proxy: Parse PDF technical sheet with AI."""
+    file = request.files.get("file")
+    if not file:
+        return jsonify({"error": True, "detail": "Nessun file caricato"}), 400
+
+    api_key = session.get("api_key", "")
+    base_url = Config.API_SERVER_URL.rstrip("/")
+    resp = http_requests.post(
+        f"{base_url}/api/recipes/{recipe_id}/parse-technical-sheet",
+        headers={"X-API-Key": api_key},
+        files={"file": (file.filename, file.stream, file.content_type)},
+        timeout=90,
+    )
+
+    try:
+        payload = resp.json()
+    except ValueError:
+        payload = {
+            "error": True,
+            "detail": resp.text or f"HTTP {resp.status_code}",
+        }
+    return jsonify(payload), resp.status_code

src/frontend/flask_app/blueprints/measure.py

@@ -1,8 +1,6 @@
 """MeasurementTec blueprint - recipe selection and measurement execution."""
-import requests as http_requests
-
 from flask import (
-    Blueprint, Response, flash, jsonify, redirect, render_template,
+    Blueprint, flash, jsonify, redirect, render_template,
     request, session, url_for,
 )
 from flask_babel import gettext as _
@@ -10,9 +8,65 @@ from flask_babel import gettext as _
 from blueprints.auth import login_required, role_required
 from config import Config
 from services.api_client import api_client
+from services.file_proxy import proxy_file
 
 measure_bp = Blueprint("measure", __name__)
 
+# Session key holding a station chosen at runtime, overriding Config.STATION_CODE.
+STATION_OVERRIDE_KEY = "station_override"
+
+
+def _current_station() -> tuple[str | None, bool]:
+    """Return the station this client acts as, and whether it is an override."""
+    override = session.get(STATION_OVERRIDE_KEY)
+    if override:
+        return override, True
+    return Config.STATION_CODE, False
+
+
+def _apply_station_switch() -> None:
+    """Honour ?station=CODE, used to exercise several stations from one PC.
+
+    Only active when STATION_SWITCH_ENABLED is set: in production the station
+    identity comes from the local install, and silently measuring against another
+    station's recipes would break traceability. An empty value clears the override
+    and falls back to the configured station.
+
+    The requested code is validated against the server before being stored, so a
+    typo leaves the operator on the current station with an explanation rather
+    than stranded on a station that does not exist.
+    """
+    if "station" not in request.args:
+        return
+
+    requested = (request.args.get("station") or "").strip().upper()
+
+    if not Config.STATION_SWITCH_ENABLED:
+        flash(_("Il cambio stazione da URL non è abilitato su questa postazione."), "error")
+        return
+
+    if not requested:
+        if session.pop(STATION_OVERRIDE_KEY, None):
+            flash(_("Stazione riportata a quella configurata."), "info")
+        return
+
+    try:
+        probe = api_client.get_station_recipes(requested)
+    except Exception:
+        flash(_("Impossibile contattare il server per cambiare stazione."), "error")
+        return
+
+    if isinstance(probe, dict) and probe.get("error"):
+        flash(
+            _("Stazione '%(code)s' inesistente o disattivata: stazione invariata.",
+              code=requested),
+            "error",
+        )
+        return
+
+    session[STATION_OVERRIDE_KEY] = requested
+    flash(_("Stazione corrente: %(code)s", code=requested), "info")
+
 
 # ---------------------------------------------------------------------------
 # Route: Recipe selection
@@ -22,13 +76,16 @@ measure_bp = Blueprint("measure", __name__)
 @role_required("MeasurementTec")
 def select_recipe():
     """Recipe selection page with search and barcode support."""
-    # Fail-fast if STATION_CODE is not configured
-    if not Config.STATION_CODE:
+    _apply_station_switch()
+    station_code, station_overridden = _current_station()
+
+    # Fail-fast if no station is configured and none was chosen
+    if not station_code:
         return render_template("errors/station_not_configured.html"), 503
 
     # Load recipes filtered by station
     try:
-        resp = api_client.get_station_recipes(Config.STATION_CODE)
+        resp = api_client.get_station_recipes(station_code)
     except Exception as e:
         return render_template(
             "errors/station_not_configured.html", error=str(e),
@@ -53,13 +110,121 @@ def select_recipe():
     return render_template(
         "measure/select_recipe.html",
         recipes=recipes,
-        station_code=Config.STATION_CODE,
+        station_code=station_code,
+        station_overridden=station_overridden,
+        station_configured=Config.STATION_CODE,
         auto_recipe_code=auto_recipe_code,
         auto_lot=auto_lot,
         auto_serial=auto_serial,
     )
 
 
+def _remember_traceability() -> tuple[str, str]:
+    """Carry lot and serial from the query string into the session."""
+    lot_number = request.args.get("lot_number", session.get("lot_number", ""))
+    serial_number = request.args.get(
+        "serial_number", session.get("serial_number", ""),
+    )
+    if lot_number:
+        session["lot_number"] = lot_number
+    if serial_number:
+        session["serial_number"] = serial_number
+    return lot_number, serial_number
+
+
+def _missing_traceability(recipe: dict, lot: str, serial: str) -> list[str]:
+    """The compulsory fields this recipe is still owed."""
+    missing = []
+    if recipe.get("requires_lot") and not (lot or "").strip():
+        missing.append(_("lotto"))
+    if recipe.get("requires_serial") and not (serial or "").strip():
+        missing.append(_("seriale"))
+    return missing
+
+
+def _ordered_tasks(recipe_id: int) -> list[dict] | None:
+    """Tasks of a recipe's current version in execution order, or None on error."""
+    resp = api_client.get(f"/api/recipes/{recipe_id}/tasks")
+    if isinstance(resp, dict) and resp.get("error"):
+        return None
+    tasks = resp if isinstance(resp, list) else resp.get("items", [])
+    return sorted(tasks, key=lambda t: (t.get("order_index", 0), t.get("id", 0)))
+
+
+def _task_progress(version_id, run_id) -> dict[int, dict]:
+    """Quotes already taken per measurement task, keyed by task id.
+
+    Never fatal: a list that cannot say which task was left half done is still a
+    usable list, and losing the whole page over a badge would be the worse trade.
+    """
+    if not version_id:
+        return {}
+    params = {"version_id": version_id}
+    if run_id:
+        params["production_run_id"] = run_id
+    resp = api_client.get("/api/measurements/task-progress", params=params)
+    if not isinstance(resp, dict) or resp.get("error"):
+        return {}
+    return {row["task_id"]: row for row in resp.get("tasks", [])}
+
+
+def _current_run_id() -> int | None:
+    """The production open at this station, if there is one."""
+    station_code, _overridden = _current_station()
+    if not station_code:
+        return None
+    # The endpoint answers with the run itself, or null when nothing is running.
+    resp = api_client.get(
+        "/api/production-runs/current", params={"station_code": station_code},
+    )
+    if not isinstance(resp, dict) or resp.get("error"):
+        return None
+    return resp.get("id")
+
+
+# ---------------------------------------------------------------------------
+# Route: start the guided sequence
+# ---------------------------------------------------------------------------
+@measure_bp.route("/start/<int:recipe_id>")
+@login_required
+@role_required("MeasurementTec")
+def start_sequence(recipe_id: int):
+    """Open the first task of a recipe and let the operator walk the sequence.
+
+    Choosing a recipe used to land on the full task list, which asked the operator
+    to pick a starting point before they had done anything. The sequence is the
+    normal way through; the list is still there, one level down, for when they need
+    to see the whole of it.
+    """
+    lot_number, serial_number = _remember_traceability()
+
+    recipe_resp = api_client.get(f"/api/recipes/{recipe_id}")
+    if recipe_resp.get("error"):
+        flash(
+            _("Ricetta non trovata: %(detail)s",
+              detail=recipe_resp.get("detail", "")),
+            "error",
+        )
+        return redirect(url_for("measure.select_recipe"))
+
+    missing = _missing_traceability(recipe_resp, lot_number, serial_number)
+    if missing:
+        flash(
+            _("Compila prima: %(fields)s", fields=", ".join(missing)),
+            "error",
+        )
+        return redirect(url_for(
+            "measure.select_recipe", recipe=recipe_resp.get("code", ""),
+        ))
+
+    tasks = _ordered_tasks(recipe_id)
+    if not tasks:
+        flash(_("Questa ricetta non ha ancora task definiti."), "error")
+        return redirect(url_for("measure.task_list", recipe_id=recipe_id))
+
+    return redirect(url_for("measure.task_execute", task_id=tasks[0]["id"]))
+
+
 # ---------------------------------------------------------------------------
 # Route: Task list for a recipe
 # ---------------------------------------------------------------------------
@@ -68,17 +233,7 @@ def select_recipe():
 @role_required("MeasurementTec")
 def task_list(recipe_id: int):
     """Task list for selected recipe."""
-    # Persist lot/serial from query params into session
-    lot_number = request.args.get(
-        "lot_number", session.get("lot_number", ""),
-    )
-    serial_number = request.args.get(
-        "serial_number", session.get("serial_number", ""),
-    )
-    if lot_number:
-        session["lot_number"] = lot_number
-    if serial_number:
-        session["serial_number"] = serial_number
+    lot_number, serial_number = _remember_traceability()
 
     # Load recipe details
     recipe_resp = api_client.get(f"/api/recipes/{recipe_id}")
@@ -90,22 +245,19 @@ def task_list(recipe_id: int):
         )
         return redirect(url_for("measure.select_recipe"))
 
-    # Load tasks for this recipe
-    tasks_resp = api_client.get(f"/api/recipes/{recipe_id}/tasks")
-    if isinstance(tasks_resp, dict) and tasks_resp.get("error"):
-        flash(
-            _("Errore nel caricamento dei task: %(detail)s",
-              detail=tasks_resp.get("detail", "")),
-            "error",
-        )
+    tasks = _ordered_tasks(recipe_id)
+    if tasks is None:
+        flash(_("Errore nel caricamento dei task."), "error")
         tasks = []
-    else:
-        tasks = tasks_resp if isinstance(tasks_resp, list) else tasks_resp.get("items", [])
+
+    version = recipe_resp.get("current_version") or {}
+    progress = _task_progress(version.get("id"), _current_run_id())
 
     return render_template(
         "measure/task_list.html",
         recipe=recipe_resp,
         tasks=tasks,
+        progress=progress,
         lot_number=lot_number,
         serial_number=serial_number,
     )
@@ -135,11 +287,19 @@ def task_execute(task_id: int):
     # Load all task IDs for this recipe (ordered) for auto-advance
     recipe_id = task_resp.get("recipe_id")
     all_task_ids = []
+    measurement_interval_minutes = None
+    # Allowed until the recipe says otherwise: a recipe that could not be read must
+    # not silently take the keypad away from the operator.
+    allow_manual_input = True
     if recipe_id:
         tasks_resp = api_client.get(f"/api/recipes/{recipe_id}/tasks")
         if isinstance(tasks_resp, list):
             sorted_tasks = sorted(tasks_resp, key=lambda t: t.get("order_index", 0))
             all_task_ids = [t["id"] for t in sorted_tasks]
+        recipe_resp = api_client.get(f"/api/recipes/{recipe_id}")
+        if not recipe_resp.get("error"):
+            measurement_interval_minutes = recipe_resp.get("measurement_interval_minutes")
+            allow_manual_input = bool(recipe_resp.get("allow_manual_input", True))
 
     return render_template(
         "measure/task_execute.html",
@@ -147,6 +307,8 @@ def task_execute(task_id: int):
         lot_number=lot_number,
         serial_number=serial_number,
         all_task_ids=all_task_ids,
+        measurement_interval_minutes=measurement_interval_minutes,
+        allow_manual_input=allow_manual_input,
     )
 
 
@@ -307,9 +469,15 @@ def save_measurement():
         "lot_number": data.get("lot_number", session.get("lot_number", "")),
         "serial_number": data.get("serial_number", session.get("serial_number", "")),
         "input_method": data.get("input_method", "manual"),
+        "input_duration_ms": data.get("input_duration_ms"),
+        # Attaches the measurement to the production under way, so fine produzione
+        # can hand the whole run to the statistics file.
+        "production_run_id": data.get("production_run_id"),
     }
 
-    resp = api_client.post("/api/measurements", data=payload)
+    # Trailing slash matters: the route is declared as "/", so omitting it costs
+    # a 307 redirect on every single measurement saved.
+    resp = api_client.post("/api/measurements/", data=payload)
 
     if resp.get("error"):
         status_code = resp.get("status_code", 500)
@@ -321,6 +489,189 @@ def save_measurement():
     return jsonify(resp), 201
 
 
+# ---------------------------------------------------------------------------
+# Routes: Out-of-tolerance authorisation (point 5)
+# ---------------------------------------------------------------------------
+@measure_bp.route("/api/measurements/pending-authorisation", methods=["GET"])
+@login_required
+@role_required("MeasurementTec")
+def api_pending_authorisation():
+    """Proxy: the out-of-tolerance measurement holding this operator up, or null.
+
+    Asked when the measurement screen loads, so reloading the page finds the same
+    block the server enforces rather than a clean slate.
+    """
+    params = {"version_id": request.args.get("version_id")}
+    run_id = request.args.get("production_run_id")
+    if run_id:
+        params["production_run_id"] = run_id
+
+    resp = api_client.get("/api/measurements/pending-authorisation", params=params)
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+@measure_bp.route("/api/measurements/<int:measurement_id>/authorise", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_authorise_measurement(measurement_id: int):
+    """Proxy: the capoturno lets an out-of-tolerance value stand.
+
+    The credentials go to the endpoint that records the approval on the
+    measurement. Checking them separately, as this used to, left the approval
+    written down nowhere and nothing depending on it.
+    """
+    data = request.get_json(silent=True) or {}
+    username = (data.get("supervisor_username") or "").strip()
+    password = data.get("supervisor_password") or ""
+    if not username or not password:
+        return jsonify({
+            "error": True, "detail": _("Username e password richiesti"),
+        }), 400
+
+    resp = api_client.post(
+        f"/api/measurements/{measurement_id}/authorise",
+        data={"supervisor_username": username, "supervisor_password": password},
+    )
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+# ---------------------------------------------------------------------------
+# Routes: Production run (state that must outlive the page)
+# ---------------------------------------------------------------------------
+@measure_bp.route("/api/production/current", methods=["GET"])
+@login_required
+@role_required("MeasurementTec")
+def api_current_production():
+    """Proxy: the production open at this station, or null.
+
+    Every page asks this on load. Before, the timer and the cycle count lived in the
+    Alpine component and a change of task - a full page load - wiped them.
+    """
+    station_code, _overridden = _current_station()
+    if not station_code:
+        return jsonify({"error": True, "detail": _("Stazione non configurata")}), 503
+
+    resp = api_client.get(
+        "/api/production-runs/current", params={"station_code": station_code},
+    )
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+@measure_bp.route("/api/production/start", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_start_production():
+    """Proxy: open a production at this station."""
+    station_code, _overridden = _current_station()
+    if not station_code:
+        return jsonify({"error": True, "detail": _("Stazione non configurata")}), 503
+
+    data = request.get_json(silent=True) or {}
+    payload = {
+        "station_code": station_code,
+        "recipe_id": data.get("recipe_id"),
+        "version_id": data.get("version_id"),
+        "lot_number": data.get("lot_number") or session.get("lot_number") or None,
+        "serial_number": data.get("serial_number") or session.get("serial_number") or None,
+    }
+    resp = api_client.post("/api/production-runs", data=payload)
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 201
+
+
+@measure_bp.route("/api/production/<int:run_id>/cycle", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_complete_cycle(run_id: int):
+    """Proxy: record a finished measurement task.
+
+    The task travels with the call: only the last measurement task of the recipe
+    closes the cycle and restarts the interval, and the server is what decides.
+    """
+    data = request.get_json(silent=True) or {}
+    resp = api_client.post(
+        f"/api/production-runs/{run_id}/cycle",
+        data={"task_id": data.get("task_id"), "note": data.get("note")},
+    )
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+@measure_bp.route("/api/production/<int:run_id>/remeasure", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_remeasure(run_id: int):
+    """Proxy: the piece was turned over - measure again inside the same cycle."""
+    data = request.get_json(silent=True) or {}
+    resp = api_client.post(
+        f"/api/production-runs/{run_id}/remeasure",
+        data={"task_id": data.get("task_id"), "note": data.get("note")},
+    )
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+def _supervised_action(run_id: int, action: str):
+    """Forward an action that needs the capoturno's authorisation.
+
+    The credentials go straight to the API, which checks them and records who
+    authorised what on the run's trace. Validating them separately first would
+    leave no such record, and is an extra round trip besides.
+    """
+    data = request.get_json(silent=True) or {}
+    username = (data.get("supervisor_username") or "").strip()
+    password = data.get("supervisor_password") or ""
+    if not username or not password:
+        return jsonify({
+            "error": True, "detail": _("Username e password richiesti"),
+        }), 400
+
+    resp = api_client.post(
+        f"/api/production-runs/{run_id}/{action}",
+        data={
+            "supervisor_username": username,
+            "supervisor_password": password,
+            "note": data.get("note"),
+        },
+    )
+    if isinstance(resp, dict) and resp.get("error"):
+        return jsonify(resp), resp.get("status_code", 500)
+    return jsonify(resp), 200
+
+
+@measure_bp.route("/api/production/<int:run_id>/pause", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_pause_production(run_id: int):
+    """Proxy: fermo linea - suspends the run and freezes the countdown."""
+    return _supervised_action(run_id, "pause")
+
+
+@measure_bp.route("/api/production/<int:run_id>/resume", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_resume_production(run_id: int):
+    """Proxy: restart a stopped line, giving back the time the stop took."""
+    return _supervised_action(run_id, "resume")
+
+
+@measure_bp.route("/api/production/<int:run_id>/close", methods=["POST"])
+@login_required
+@role_required("MeasurementTec")
+def api_close_production(run_id: int):
+    """Proxy: fine produzione - closes the run and emits the statistics file."""
+    return _supervised_action(run_id, "close")
+
+
 # ---------------------------------------------------------------------------
 # Route: File proxy (browser can't send X-API-Key directly)
 # ---------------------------------------------------------------------------
@@ -328,16 +679,4 @@ def save_measurement():
 @login_required
 def api_get_file(file_path: str):
     """Proxy: Serve file from API server (browser can't send X-API-Key)."""
-    api_key = session.get("api_key", "")
-    base_url = Config.API_SERVER_URL.rstrip("/")
-    resp = http_requests.get(
-        f"{base_url}/api/files/{file_path}",
-        headers={"X-API-Key": api_key},
-        timeout=30,
-    )
-    if resp.status_code != 200:
-        return Response(resp.text, status=resp.status_code)
-    return Response(
-        resp.content,
-        content_type=resp.headers.get("content-type", "application/octet-stream"),
-    )
+    return proxy_file(file_path)

src/frontend/flask_app/config.py

@@ -26,6 +26,14 @@ class Config:
     # code it belongs to. Empty/None means "not configured".
     STATION_CODE: str | None = os.getenv("STATION_CODE") or None
 
+    # Allow switching the current station from the URL (?station=CODE), keeping the
+    # override in the session. Meant for commissioning: it lets one PC exercise
+    # several stations instead of needing one machine per station.
+    # Off by default - on the shop floor the station identity comes from the local
+    # install, and measuring against the wrong station's recipes would silently
+    # break traceability.
+    STATION_SWITCH_ENABLED: bool = os.getenv("STATION_SWITCH_ENABLED", "0") == "1"
+
     # Babel i18n
     BABEL_DEFAULT_LOCALE = "it"
     BABEL_DEFAULT_TIMEZONE = "Europe/Rome"

src/frontend/flask_app/services/api_client.py

@@ -56,11 +56,21 @@ class APIClient:
         try:
             error_body = response.json()
             detail = error_body.get("detail", str(error_body))
-            # FastAPI 422 returns detail as a list of validation errors
+            # FastAPI 422 returns detail as a list of validation errors:
+            # [{"type": ..., "loc": ["body", "password"], "msg": "...", ...}, ...]
             if isinstance(detail, list):
-                detail = "; ".join(
-                    e.get("msg", str(e)) for e in detail if isinstance(e, dict)
-                ) or str(detail)
+                parts = []
+                for e in detail:
+                    if not isinstance(e, dict):
+                        parts.append(str(e))
+                        continue
+                    msg = e.get("msg", str(e))
+                    loc = e.get("loc") or []
+                    # Strip leading "body"/"query"/"path" segments; keep field path.
+                    field_path = [str(p) for p in loc if p not in ("body", "query", "path")]
+                    field = ".".join(field_path) if field_path else None
+                    parts.append(f"{field}: {msg}" if field else msg)
+                detail = "; ".join(parts) or str(detail)
         except Exception:
             detail = response.text or f"HTTP {response.status_code}"
 

src/frontend/flask_app/services/file_proxy.py

@@ -0,0 +1,25 @@
+"""Shared file proxy: relay uploads from the FastAPI server to the browser.
+
+The browser can't send the X-API-Key header directly, so blueprints expose
+a proxy route and delegate here.
+"""
+import requests as http_requests
+from flask import Response, session
+
+from config import Config
+
+
+def proxy_file(file_path: str) -> Response:
+    """Fetch a file from the API server and relay it with its content type."""
+    base_url = Config.API_SERVER_URL.rstrip("/")
+    resp = http_requests.get(
+        f"{base_url}/api/files/{file_path}",
+        headers={"X-API-Key": session.get("api_key", "")},
+        timeout=30,
+    )
+    if resp.status_code != 200:
+        return Response(resp.text, status=resp.status_code)
+    return Response(
+        resp.content,
+        content_type=resp.headers.get("content-type", "application/octet-stream"),
+    )

src/frontend/flask_app/static/css/themes.css

@@ -69,6 +69,23 @@
 }
 
 
+/* ============================================================
+   Brand details
+   ============================================================ */
+
+:root {
+  accent-color: var(--color-primary);
+}
+
+::selection {
+  background: rgba(37, 99, 235, 0.2);
+}
+
+.dark ::selection {
+  background: rgba(96, 165, 250, 0.3);
+}
+
+
 /* ============================================================
    Global Transitions
    ============================================================ */
@@ -107,10 +124,66 @@ textarea:focus-visible {
 }
 
 
+/* ============================================================
+   Page shell — one frame for every view
+
+   Punto 14. Ogni schermata si disegnava la propria larghezza: sette valori
+   diversi fra le viste, nessuno allineato alla navbar, che sta a 80rem su
+   tutte. Passando da un menu all'altro il blocco di contenuto cambiava
+   larghezza e si spostava rispetto alla barra sopra di sé.
+
+   Qui la cornice è una sola e coincide con quella della navbar. Le pagine
+   che hanno bisogno di una colonna stretta - i moduli - la stringono al loro
+   interno: il bordo della pagina resta dov'è, cambia solo il contenuto.
+   ============================================================ */
+
+.tmf-page {
+  width: 100%;
+  max-width: 80rem;          /* == max-w-7xl, la stessa della navbar */
+  margin-inline: auto;
+  padding-inline: 1rem;      /* == px-4  */
+  padding-block: 1.5rem;     /* == py-6  */
+}
+
+@media (min-width: 640px) {
+  .tmf-page { padding-inline: 1.5rem; }   /* == sm:px-6 */
+}
+
+@media (min-width: 1024px) {
+  .tmf-page { padding-inline: 2rem; }     /* == lg:px-8 */
+}
+
+/* I moduli restano su una colonna leggibile: un campo di testo largo 1280px non
+   si compila meglio, si compila peggio. Due larghezze in tutto il prodotto -
+   questa e quella sopra - al posto delle sette di prima. */
+.tmf-page.tmf-page-narrow {
+  max-width: 56rem;          /* == max-w-4xl */
+}
+
+
 /* ============================================================
    Custom Scrollbar
    ============================================================ */
 
+/* Lo spazio della barra di scorrimento è riservato sempre, anche quando la
+   pagina non scorre. Senza, ogni passaggio fra una vista lunga e una corta -
+   e ogni ingresso nella schermata di misura, che blocca lo scorrimento -
+   spostava di 8px tutto il contenuto centrato, navbar compresa. */
+html {
+  scrollbar-gutter: stable;
+}
+
+/* La schermata di misura è alta quanto la finestra vera, non quanto quella
+   teorica: su Android e iOS `100vh` è l'altezza che la pagina avrebbe con la
+   barra dell'indirizzo nascosta, quindi su un tablet il piede - dove stanno
+   «Fine ciclo misura» e il tastierino - finisce sotto il bordo dello schermo, e
+   ricompare quando la barra si ritrae. `100dvh` segue la finestra reale. */
+@supports (height: 100dvh) {
+  body.h-screen {
+    height: 100dvh;
+  }
+}
+
 /* Webkit (Chrome, Safari, Edge) */
 ::-webkit-scrollbar {
   width: 8px;
@@ -151,6 +224,7 @@ textarea:focus-visible {
 
 .tmf-card:hover {
   box-shadow: var(--shadow-md);
+  border-color: var(--scrollbar-thumb);
 }
 
 .tmf-card-header {
@@ -250,11 +324,21 @@ textarea:focus-visible {
   border: 1px solid transparent;
 }
 
+.btn:active:not(:disabled) {
+  transform: scale(0.97);
+}
+
 .btn:disabled {
   opacity: 0.5;
   cursor: not-allowed;
 }
 
+.btn-lg {
+  padding: 0.75rem 1.5rem;
+  font-size: 1rem;
+  border-radius: 0.625rem;
+}
+
 .btn-primary {
   background: var(--color-primary);
   color: #FFFFFF;

src/frontend/flask_app/static/img/tmflow-icon.svg

@@ -1,27 +1,21 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32" fill="none">
-  <!-- Stylized Caliper Icon for Favicon -->
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" fill="none">
+  <!-- TieMeasureFlow brand mark: measurement span glyph on brand-blue badge -->
+  <defs>
+    <linearGradient id="tmfBadge" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0" stop-color="#3B82F6"/>
+      <stop offset="1" stop-color="#1D4ED8"/>
+    </linearGradient>
+  </defs>
 
-  <!-- Caliper outer frame -->
-  <rect x="2" y="3" width="5" height="26" rx="1.5" fill="#2563EB"/>
-  <rect x="2" y="3" width="19" height="5" rx="1.5" fill="#2563EB"/>
-  <rect x="2" y="24" width="19" height="5" rx="1.5" fill="#2563EB"/>
+  <!-- Badge -->
+  <rect x="2" y="2" width="60" height="60" rx="15" fill="url(#tmfBadge)"/>
 
-  <!-- Sliding jaw -->
-  <rect x="12" y="8" width="4.5" height="6.5" rx="1" fill="#1E40AF"/>
-  <rect x="12" y="17.5" width="4.5" height="6.5" rx="1" fill="#1E40AF"/>
+  <!-- Measurement end bars -->
+  <rect x="14" y="16" width="5" height="32" rx="2.5" fill="#FFFFFF"/>
+  <rect x="45" y="16" width="5" height="32" rx="2.5" fill="#FFFFFF"/>
 
-  <!-- Depth rod -->
-  <rect x="4" y="14" width="12" height="3.5" rx="1" fill="#3B82F6" opacity="0.55"/>
-
-  <!-- Scale markings -->
-  <rect x="8" y="5" width="1" height="2.5" rx="0.5" fill="#FFFFFF" opacity="0.65"/>
-  <rect x="11" y="5" width="1" height="2.5" rx="0.5" fill="#FFFFFF" opacity="0.65"/>
-  <rect x="14" y="5" width="1" height="2.5" rx="0.5" fill="#FFFFFF" opacity="0.65"/>
-  <rect x="17" y="5" width="1" height="2.5" rx="0.5" fill="#FFFFFF" opacity="0.65"/>
-
-  <!-- Flow arrow -->
-  <path d="M22 16 C25 16, 26.5 11, 29 11"
-        stroke="#3B82F6" stroke-width="2.2" stroke-linecap="round" fill="none"/>
-  <path d="M27 8.5 L29.5 11 L27 13.5"
-        stroke="#3B82F6" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round" fill="none"/>
+  <!-- Double-headed arrow (measured span) -->
+  <path d="M24 32 H40" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round"/>
+  <path d="M28 26.5 L22.5 32 L28 37.5" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round" stroke-linejoin="round"/>
+  <path d="M36 26.5 L41.5 32 L36 37.5" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round" stroke-linejoin="round"/>
 </svg>

src/frontend/flask_app/static/img/tmflow-logo.svg

@@ -1,41 +1,23 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 220 44" fill="none">
-  <!-- Stylized Caliper Icon -->
-  <g transform="translate(2, 2)">
-    <!-- Caliper outer frame -->
-    <rect x="0" y="4" width="6" height="32" rx="1.5" fill="#2563EB"/>
-    <rect x="0" y="4" width="24" height="6" rx="1.5" fill="#2563EB"/>
-    <rect x="0" y="30" width="24" height="6" rx="1.5" fill="#2563EB"/>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 340 64" fill="none">
+  <!-- TieMeasureFlow full logo: brand badge + wordmark (standalone asset) -->
+  <defs>
+    <linearGradient id="tmfBadgeFull" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0" stop-color="#3B82F6"/>
+      <stop offset="1" stop-color="#1D4ED8"/>
+    </linearGradient>
+  </defs>
 
-    <!-- Caliper sliding jaw -->
-    <rect x="14" y="10" width="5" height="8" rx="1" fill="#1E40AF"/>
-    <rect x="14" y="22" width="5" height="8" rx="1" fill="#1E40AF"/>
+  <!-- Badge -->
+  <rect x="2" y="2" width="60" height="60" rx="15" fill="url(#tmfBadgeFull)"/>
+  <rect x="14" y="16" width="5" height="32" rx="2.5" fill="#FFFFFF"/>
+  <rect x="45" y="16" width="5" height="32" rx="2.5" fill="#FFFFFF"/>
+  <path d="M24 32 H40" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round"/>
+  <path d="M28 26.5 L22.5 32 L28 37.5" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round" stroke-linejoin="round"/>
+  <path d="M36 26.5 L41.5 32 L36 37.5" stroke="#FFFFFF" stroke-width="4" stroke-linecap="round" stroke-linejoin="round"/>
 
-    <!-- Caliper depth rod -->
-    <rect x="2" y="18" width="16" height="4" rx="1" fill="#3B82F6" opacity="0.6"/>
-
-    <!-- Scale markings -->
-    <rect x="7" y="7" width="1" height="3" rx="0.5" fill="#FFFFFF" opacity="0.7"/>
-    <rect x="10" y="7" width="1" height="3" rx="0.5" fill="#FFFFFF" opacity="0.7"/>
-    <rect x="13" y="7" width="1" height="3" rx="0.5" fill="#FFFFFF" opacity="0.7"/>
-    <rect x="16" y="7" width="1" height="3" rx="0.5" fill="#FFFFFF" opacity="0.7"/>
-    <rect x="19" y="7" width="1" height="3" rx="0.5" fill="#FFFFFF" opacity="0.7"/>
-
-    <!-- Flow arrow (smooth curve) -->
-    <path d="M26 20 C30 20, 32 14, 36 14 C40 14, 40 20, 36 20"
-          stroke="#2563EB" stroke-width="2.5" stroke-linecap="round" fill="none"/>
-    <path d="M34 17 L37 20 L34 23" stroke="#2563EB" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" fill="none"/>
-  </g>
-
-  <!-- Text: TieMeasureFlow -->
-  <g transform="translate(48, 0)">
-    <!-- "Tie" in bold primary -->
-    <text x="0" y="28" font-family="Inter, system-ui, sans-serif" font-size="20" font-weight="700" fill="#2563EB"
-          letter-spacing="-0.5">Tie</text>
-    <!-- "Measure" in medium dark -->
-    <text x="32" y="28" font-family="Inter, system-ui, sans-serif" font-size="20" font-weight="500" fill="#1E40AF"
-          letter-spacing="-0.5">Measure</text>
-    <!-- "Flow" in bold primary -->
-    <text x="120" y="28" font-family="Inter, system-ui, sans-serif" font-size="20" font-weight="700" fill="#2563EB"
-          letter-spacing="-0.5">Flow</text>
-  </g>
+  <!-- Wordmark: single text element with flowing tspans (cannot be clipped/overlap) -->
+  <text x="76" y="42" font-family="Inter, system-ui, -apple-system, sans-serif"
+        font-size="30" letter-spacing="-0.5" fill="#0F172A">
+    <tspan font-weight="700">Tie</tspan><tspan font-weight="600" fill="#2563EB">Measure</tspan><tspan font-weight="700">Flow</tspan>
+  </text>
 </svg>

src/frontend/flask_app/static/js/alpine-init.js

@@ -16,9 +16,21 @@
 
   /**
    * Resolve the initial dark-mode state.
-   * Priority: localStorage > system preference > light (default).
+   * Priority: server (per-user theme_pref) > localStorage > system preference > light.
+   *
+   * The server value (window.__SERVER_THEME) is authoritative for logged-in
+   * users: it reflects the theme_pref of whoever is currently authenticated,
+   * so switching user actually switches theme instead of inheriting the
+   * previous user's localStorage value. It is empty for anonymous pages.
    */
   function getInitialDark() {
+    var server = (typeof window !== 'undefined' && window.__SERVER_THEME) || '';
+    if (server === 'dark' || server === 'light') {
+      // Keep localStorage in sync so the rest of the app stays consistent.
+      try { localStorage.setItem(STORAGE_KEY, server); } catch (_) {}
+      return server === 'dark';
+    }
+
     var stored = null;
     try {
       stored = localStorage.getItem(STORAGE_KEY);
@@ -70,6 +82,7 @@
       toggle: function () {
         this.dark = !this.dark;
         this._apply();
+        this._persist();
       },
 
       set: function (dark) {
@@ -77,6 +90,24 @@
         this._apply();
       },
 
+      // Persist an explicit user choice to the server (session + theme_pref),
+      // so it survives navigation and matches on the next login. Best-effort.
+      _persist: function () {
+        try {
+          var csrf = document.querySelector('meta[name=csrf-token]');
+          fetch('/auth/set-theme', {
+            method: 'POST',
+            headers: {
+              'Content-Type': 'application/json',
+              'X-CSRFToken': (csrf && csrf.content) || ''
+            },
+            body: JSON.stringify({ theme: this.dark ? 'dark' : 'light' })
+          }).catch(function () {});
+        } catch (_) {
+          // Non-fatal: localStorage already holds the choice for this browser.
+        }
+      },
+
       _apply: function () {
         if (this.dark) {
           document.documentElement.classList.add('dark');

src/frontend/flask_app/static/js/annotation-editor.js

@@ -671,6 +671,53 @@ function annotationEditor() {
       return rect;
     },
 
+    // ----------------------------------------------------------------
+    // Object creation - Text
+    // ----------------------------------------------------------------
+
+    addText(x, y, textContent, fontSize) {
+      var size = fontSize || 18;
+      var itext = new fabric.IText(textContent || 'Testo', {
+        left: x,
+        top: y,
+        fontSize: size,
+        fill: this.currentColor,
+        fontFamily: 'Inter, sans-serif',
+        fontWeight: 'bold',
+        selectable: true,
+        evented: true,
+        hasControls: false,
+        hasBorders: true,
+        lockRotation: true,
+        lockScalingX: true,
+        lockScalingY: true,
+        objectType: 'text',
+        textColor: this.currentColor,
+        textFontSize: size,
+      });
+
+      var self = this;
+      itext.on('editing:exited', function () {
+        if (!itext.text || itext.text.trim() === '') {
+          self.canvas.remove(itext);
+        }
+        itext.textColor = itext.fill;
+        self.isDirty = true;
+        window.dispatchEvent(new CustomEvent('annotations-changed', { detail: { json: self.getAnnotationsJson() } }));
+      });
+
+      itext.on('changed', function () {
+        self.isDirty = true;
+      });
+
+      this.canvas.add(itext);
+      itext.setCoords();
+      this.canvas.setActiveObject(itext);
+      this.isDirty = true;
+      window.dispatchEvent(new CustomEvent('annotations-changed', { detail: { json: this.getAnnotationsJson() } }));
+      return itext;
+    },
+
     // ----------------------------------------------------------------
     // Toolbar / mode switching
     // ----------------------------------------------------------------
@@ -678,7 +725,7 @@ function annotationEditor() {
     /**
      * Set the current interaction mode.
      *
-     * @param {'select'|'marker'|'arrow'|'rect'} mode
+     * @param {'select'|'marker'|'arrow'|'rect'|'text'} mode
      */
     setMode(mode) {
       this.activeMode = mode;
@@ -731,6 +778,13 @@ function annotationEditor() {
           self.addMarker(pointer.x, pointer.y);
           self.setMode('select');
           window.dispatchEvent(new CustomEvent('anno-mode-changed', { detail: { mode: 'select' } }));
+        } else if (self.activeMode === 'text') {
+          var textObj = self.addText(pointer.x, pointer.y);
+          self.setMode('select');
+          window.dispatchEvent(new CustomEvent('anno-mode-changed', { detail: { mode: 'select' } }));
+          self.canvas.setActiveObject(textObj);
+          textObj.enterEditing();
+          textObj.selectAll();
         } else if (self.activeMode === 'arrow') {
           isDrawing = true;
           drawStartX = pointer.x;
@@ -902,7 +956,7 @@ function annotationEditor() {
     setupKeyboard() {
       var self = this;
 
-      document.addEventListener('keydown', function (e) {
+      this._onKeyDown = function (e) {
         // Skip when typing in form fields
         var tag = e.target.tagName;
         if (tag === 'INPUT' || tag === 'TEXTAREA' || tag === 'SELECT') return;
@@ -916,7 +970,8 @@ function annotationEditor() {
           self.canvas.discardActiveObject();
           self.canvas.renderAll();
         }
-      });
+      };
+      document.addEventListener('keydown', this._onKeyDown);
     },
 
     // ----------------------------------------------------------------
@@ -1023,6 +1078,13 @@ function annotationEditor() {
         if (obj.arrowLineDash) data.lineDash = obj.arrowLineDash;
         if (obj.areaLineDash) data.lineDash = obj.areaLineDash;
 
+        // Text-specific
+        if (obj.objectType === 'text') {
+          data.textContent = obj.text;
+          data.textColor = obj.textColor || obj.fill;
+          data.fontSize = obj.textFontSize || obj.fontSize;
+        }
+
         // Arrow-specific: store original endpoints
         if (obj.objectType === 'arrow') {
           data.x1 = obj.arrowX1 != null ? Math.round(obj.arrowX1 * 100) / 100 : null;
@@ -1120,6 +1182,11 @@ function annotationEditor() {
             this.currentColor = savedColor3;
             this.currentStrokeWidth = savedWidth2;
             this.currentLineDash = savedDash2;
+          } else if (obj.type === 'text') {
+            var savedColor4 = this.currentColor;
+            if (obj.textColor) this.currentColor = obj.textColor;
+            this.addText(obj.left * cs, obj.top * cs, obj.textContent, (obj.fontSize || 18) * cs);
+            this.currentColor = savedColor4;
           }
         }
 
@@ -1223,6 +1290,9 @@ function annotationEditor() {
       if (this._onImageLoaded) {
         window.removeEventListener('image-loaded', this._onImageLoaded);
       }
+      if (this._onKeyDown) {
+        document.removeEventListener('keydown', this._onKeyDown);
+      }
       if (this.canvas) {
         this.canvas.dispose();
         this.canvas = null;

src/frontend/flask_app/static/js/annotation-viewer.js

@@ -436,6 +436,8 @@ function _drawEditorAnnotations(ctx, data, scale) {
       _drawPreviewArrow(ctx, obj, annoScale);
     } else if (obj.type === 'area') {
       _drawPreviewArea(ctx, obj, annoScale);
+    } else if (obj.type === 'text') {
+      _drawPreviewText(ctx, obj, annoScale);
     }
   }
 }
@@ -514,6 +516,24 @@ function _drawPreviewArea(ctx, obj, scale) {
   ctx.setLineDash([]);
 }
 
+function _drawPreviewText(ctx, obj, scale) {
+  var x = obj.left * scale;
+  var y = obj.top * scale;
+  var color = obj.textColor || '#000000';
+  var fontSize = Math.max(10, Math.round((obj.fontSize || 18) * scale));
+
+  ctx.fillStyle = color;
+  ctx.font = 'bold ' + fontSize + 'px Inter, sans-serif';
+  ctx.textAlign = 'left';
+  ctx.textBaseline = 'top';
+
+  var lines = (obj.textContent || '').split('\n');
+  var lineHeight = fontSize * 1.2;
+  for (var j = 0; j < lines.length; j++) {
+    ctx.fillText(lines[j], x, y + j * lineHeight);
+  }
+}
+
 /**
  * Draw annotations in the legacy format (markers: point/rectangle).
  */

src/frontend/flask_app/static/js/numpad.js

@@ -3,7 +3,9 @@
  * Used for measurement data entry in task_execute.html
  */
 
-function numpad() {
+function numpad(options) {
+  var opts = options || {};
+
   return {
     // State
     value: '',           // String representation of the current value
@@ -13,9 +15,16 @@ function numpad() {
     maxIntDigits: 6,     // Maximum integer digits
     maxDecDigits: 6,     // Maximum decimal digits
 
+    /* Whether a value may be typed at all, as the recipe declares it. When false
+       the keypad shows no digits and a value that looks typed is refused here as
+       well as by the server: the caliper is the instrument. Defaults to allowed,
+       so a caller that says nothing gets the behaviour that came before. */
+    allowManual: opts.allowManual !== false,
+
     // HID burst detection (USB caliper vs manual typing)
     _lastKeyTime: 0,     // Timestamp of last keystroke
     _burstCount: 0,      // Consecutive fast keystrokes
+    _enterWasFast: false, // Enter arrived in the same burst as the digits
 
     /**
      * Get the display value with sign
@@ -109,6 +118,20 @@ function numpad() {
       this.value = '';
       this.negative = false;
       this.hasDecimal = false;
+      this._enterWasFast = false;
+    },
+
+    /**
+     * How the value in the display got there.
+     *
+     * A wedge caliper sends its digits and the Enter that follows them as one
+     * burst; a person is slower on both. The Enter is the stronger of the two
+     * signals: a short reading like "9.5" is only three keystrokes, too few to
+     * judge by count alone, and used to be filed as typed by hand.
+     */
+    _classifyInput() {
+      if (this._enterWasFast) return 'usb_caliper';
+      return this._burstCount >= 3 ? 'usb_caliper' : 'manual';
     },
 
     /**
@@ -118,9 +141,18 @@ function numpad() {
       if (!this.hasValue) return;
 
       const val = this.numericValue;
+      const inputMethod = this._classifyInput();
 
-      // Determine input method: 3+ fast keystrokes = USB caliper burst
-      const inputMethod = this._burstCount >= 3 ? 'usb_caliper' : 'manual';
+      // The recipe forbids typing: say so and keep the value on screen rather than
+      // clearing it, so the operator sees what was refused. The server refuses the
+      // same request anyway - this is only the earlier, kinder of the two answers.
+      if (inputMethod === 'manual' && !this.allowManual) {
+        this.$dispatch('numpad-rejected', { reason: 'manual_not_allowed', value: val });
+        this._enterWasFast = false;
+        this._burstCount = 0;
+        this._lastKeyTime = 0;
+        return;
+      }
 
       // Dispatch custom event for parent component to handle
       this.$dispatch('numpad-confirm', { value: val, inputMethod: inputMethod });
@@ -163,6 +195,17 @@ function numpad() {
      * @param {KeyboardEvent} e - The keyboard event
      */
     handleKeydown(e) {
+      // Ignore keystrokes aimed at an editable field (e.g. the supervisor
+      // login modal). The numpad has no text inputs of its own, so this
+      // window-level capture is only meant for the USB caliper / keyboard
+      // wedge when no field is focused. Without this guard the numpad would
+      // swallow digits and Backspace inside those inputs.
+      const t = e.target;
+      if (t && (t.tagName === 'INPUT' || t.tagName === 'TEXTAREA' ||
+                t.tagName === 'SELECT' || t.isContentEditable)) {
+        return;
+      }
+
       // Number keys
       if (e.key >= '0' && e.key <= '9') {
         e.preventDefault();
@@ -188,6 +231,12 @@ function numpad() {
       // Enter - confirm
       else if (e.key === 'Enter') {
         e.preventDefault();
+        // Measured before confirming: an Enter that lands within the burst is the
+        // caliper closing its own transmission, not a person reaching for a key.
+        const gap = this._lastKeyTime > 0
+          ? performance.now() - this._lastKeyTime
+          : Infinity;
+        this._enterWasFast = gap < 80;
         this.confirm();
       }
       // Minus sign - toggle sign

src/frontend/flask_app/static/js/production-clock.js

@@ -0,0 +1,176 @@
+/**
+ * The measurement clock, shared by every screen an operator can be on.
+ *
+ * Two things the shop floor asked for live here. First, the countdown does not stop
+ * at zero: it keeps going the other way, so how late a measurement is can be read
+ * off the screen instead of worked out. Second, when the interval expires the
+ * measurement comes back on its own, wherever the operator happens to be - the task
+ * list, the summary, the recipe picker.
+ *
+ * The number always comes from the server (seconds_to_next_measurement, already
+ * signed). Ticking locally only keeps the display smooth between round trips, and
+ * every resync overwrites it: a tab that was asleep, or a machine whose clock is
+ * off, still shows the same figure as the station next to it.
+ *
+ * ProductionClock holds the pure parts, used here and by the measurement screen,
+ * which has its own component but must not grow a second copy of these rules.
+ */
+(function (global) {
+  'use strict';
+
+  var ProductionClock = {
+    /* mm:ss of the distance from the deadline, in whichever direction. The sign is
+       carried by the wording around it ("tra" / "in ritardo di"), not by a minus. */
+    format: function (seconds) {
+      var total = Math.abs(Math.round(seconds || 0));
+      var m = Math.floor(total / 60);
+      var s = total % 60;
+      return (m < 10 ? '0' : '') + m + ':' + (s < 10 ? '0' : '') + s;
+    },
+
+    hasClock: function (seconds) {
+      return seconds !== null && seconds !== undefined;
+    },
+
+    isOverdue: function (seconds) {
+      return ProductionClock.hasClock(seconds) && seconds <= 0;
+    },
+
+    /* Three short beeps. Wrapped because a browser refuses to make noise before the
+       page has been touched, and a silent buzzer must not take the screen with it.
+       Whether a light column replaces this is question D-5. */
+    playBuzzer: function () {
+      try {
+        var ctx = new (global.AudioContext || global.webkitAudioContext)();
+        [0, 0.25, 0.5].forEach(function (delay) {
+          var osc = ctx.createOscillator();
+          var gain = ctx.createGain();
+          osc.connect(gain);
+          gain.connect(ctx.destination);
+          osc.frequency.value = 880;
+          osc.type = 'square';
+          gain.gain.value = 0.3;
+          osc.start(ctx.currentTime + delay);
+          osc.stop(ctx.currentTime + delay + 0.15);
+        });
+      } catch (_) {}
+    },
+
+    /* The task the operator has to be taken back to: the first measurement task of
+       the recipe. Null when the recipe has none, which is the whole reason this is
+       a lookup and not an assumption. */
+    measurementUrl: function (run, executeUrlTemplate) {
+      var ids = (run && run.measurement_task_ids) || [];
+      if (!ids.length) return null;
+      return executeUrlTemplate.replace('/0', '/' + ids[0]);
+    },
+  };
+
+  global.ProductionClock = ProductionClock;
+
+  /* How long the operator gets to see why the screen is about to change. Long
+     enough to read it, short enough that it is not a way of putting the
+     measurement off. */
+  var ANNOUNCE_SECONDS = 5;
+  /* The server is asked again on this cadence: the run may have been stopped or
+     closed from another screen, and a sleeping tab drifts. */
+  var RESYNC_MS = 60000;
+
+  /**
+   * Banner component for the screens that only watch the clock.
+   *
+   * options.currentUrl  - endpoint returning the run open at this station, or null
+   * options.executeUrl  - url of a task execution page with 0 as the task id
+   */
+  global.productionClock = function (options) {
+    var opts = options || {};
+
+    return {
+      run: null,
+      seconds: null,
+      announced: false,
+      redirectIn: 0,
+      _tick: null,
+      _resync: null,
+
+      init: function () {
+        var self = this;
+        this.load();
+        this._tick = setInterval(function () { self.onSecond(); }, 1000);
+        this._resync = setInterval(function () { self.load(); }, RESYNC_MS);
+      },
+
+      destroy: function () {
+        if (this._tick) clearInterval(this._tick);
+        if (this._resync) clearInterval(this._resync);
+      },
+
+      load: async function () {
+        try {
+          var resp = await fetch(opts.currentUrl);
+          if (!resp.ok) return;
+          var run = await resp.json();
+          this.adopt(run && run.id ? run : null);
+        } catch (e) {
+          // Offline or server down: the page stays usable, just without the clock.
+        }
+      },
+
+      adopt: function (run) {
+        this.run = run;
+        this.seconds = run ? run.seconds_to_next_measurement : null;
+        if (!run || run.status !== 'running') {
+          // Closed or stopped: nothing is due, and an announcement left on screen
+          // would be telling the operator to go and measure a stopped line.
+          this.announced = false;
+          this.redirectIn = 0;
+          return;
+        }
+        // Arriving on a page that is already late is the same event as going late
+        // while sitting on it.
+        if (this.isDue) this.announce();
+      },
+
+      get running() { return !!this.run && this.run.status === 'running'; },
+      get paused() { return !!this.run && this.run.status === 'paused'; },
+      get visible() {
+        return !!this.run && ProductionClock.hasClock(this.seconds);
+      },
+      get isDue() { return this.running && ProductionClock.isOverdue(this.seconds); },
+      get display() { return ProductionClock.format(this.seconds); },
+      get measurementUrl() {
+        return ProductionClock.measurementUrl(this.run, opts.executeUrl || '');
+      },
+
+      onSecond: function () {
+        var wasAnnounced = this.announced;
+
+        if (this.running && ProductionClock.hasClock(this.seconds)) {
+          // Past zero it keeps counting, into the negative: that is the figure the
+          // operator reads as "how long I am late by".
+          this.seconds--;
+          if (this.seconds <= 0) this.announce();
+        }
+
+        if (wasAnnounced && this.redirectIn > 0) {
+          this.redirectIn--;
+          if (this.redirectIn === 0) this.goToMeasurement();
+        }
+      },
+
+      announce: function () {
+        if (this.announced) return;
+        this.announced = true;
+        ProductionClock.playBuzzer();
+        // Without a measurement task to go to there is nothing to announce beyond
+        // the banner going red, and a redirect to nowhere would be a broken link.
+        this.redirectIn = this.measurementUrl ? ANNOUNCE_SECONDS : 0;
+      },
+
+      goToMeasurement: function () {
+        var url = this.measurementUrl;
+        if (url) global.location.href = url;
+      },
+    };
+  };
+})(window);

src/frontend/flask_app/static/js/rich-text.js

@@ -0,0 +1,54 @@
+/**
+ * Task descriptions: line breaks and bold, and nothing else.
+ *
+ * Whoever writes a recipe pastes from the PDF of the technical sheet, and the text
+ * used to arrive flattened - every line run into the next, every emphasis lost, to
+ * be put back by hand each time. Two conventions carry it now: a newline is a line
+ * break, **like this** is bold.
+ *
+ * No HTML is stored or trusted. The text is escaped first and the only tags in the
+ * result are the ones produced here, so a description that contains <script> shows
+ * those characters and does nothing. The server-side filter (rich_text in app.py)
+ * follows exactly these rules, so a description reads the same whether the page
+ * was rendered by Flask or by Alpine.
+ */
+(function (global) {
+  'use strict';
+
+  function escapeHtml(text) {
+    return String(text)
+      .replace(/&/g, '&amp;')
+      .replace(/</g, '&lt;')
+      .replace(/>/g, '&gt;')
+      .replace(/"/g, '&quot;')
+      .replace(/'/g, '&#39;');
+  }
+
+  /* Escaped text, then the two tags we add ourselves. Order matters: escaping
+     after the substitution would turn our own <strong> into visible text. */
+  function richText(value) {
+    if (!value) return '';
+    return escapeHtml(value)
+      .replace(/\*\*([\s\S]+?)\*\*/g, '<strong>$1</strong>')
+      .replace(/\n/g, '<br>');
+  }
+
+  /* Wrap whatever is selected in a textarea in ** **, the way a B button does.
+     Dispatching 'input' is what keeps x-model in step: assigning to value alone
+     changes the DOM and leaves Alpine with the old string. */
+  function wrapSelectionBold(el) {
+    if (!el) return;
+    var start = el.selectionStart;
+    var end = el.selectionEnd;
+    var text = el.value || '';
+    el.value = text.slice(0, start) + '**' + text.slice(start, end) + '**' + text.slice(end);
+    el.dispatchEvent(new Event('input', { bubbles: true }));
+    el.focus();
+    // Leave the cursor around the same words, now inside the markers.
+    el.selectionStart = start + 2;
+    el.selectionEnd = end + 2;
+  }
+
+  global.richText = richText;
+  global.wrapSelectionBold = wrapSelectionBold;
+})(window);