fix(ai): risposta AI malformata -> 502, non 400 (#36 punto 7)

json.JSONDecodeError e' un ValueError: KeyError/JSONDecodeError sulla
risposta del modello finivano nell'except ValueError del router (pensato
per i nostri errori di configurazione, es. API key mancante) e uscivano
come 400 invece di 502. Nuova AIResponseError dedicata in ai_service.py
per contenuto malformato o JSON non valido; recipes.py la traduce in 502,
toccato solo per questo except.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/backend/services/ai_service.py

@@ -11,6 +11,17 @@ from src.backend.config import settings
 
 logger = logging.getLogger(__name__)
 
+
+class AIResponseError(Exception):
+    """The AI model's response could not be used: missing fields or invalid JSON.
+
+    Distinct from ValueError (raised for our own misconfiguration, e.g. a
+    missing API key) so the router can tell "we are broken" (400) apart from
+    "the upstream AI service is broken" (502) instead of both landing on the
+    same except clause.
+    """
+
+
 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:
@@ -65,7 +76,10 @@ async def parse_technical_sheet(pdf_bytes: bytes) -> list[dict]:
         resp.raise_for_status()
 
     data = resp.json()
-    content = data["choices"][0]["message"]["content"].strip()
+    try:
+        content = data["choices"][0]["message"]["content"].strip()
+    except (KeyError, IndexError, TypeError, AttributeError) as e:
+        raise AIResponseError(f"Risposta del modello AI malformata: {e}") from e
 
     # Strip markdown fences if present
     if content.startswith("```"):
@@ -74,4 +88,10 @@ async def parse_technical_sheet(pdf_bytes: bytes) -> list[dict]:
             content = content[:-3]
         content = content.strip()
 
-    return json.loads(content)
+    try:
+        return json.loads(content)
+    except json.JSONDecodeError as e:
+        # json.JSONDecodeError is itself a ValueError: without this, the
+        # router's `except ValueError` (meant for our own config errors)
+        # would misreport a broken upstream response as our fault (400).
+        raise AIResponseError(f"Risposta del modello AI non e' JSON valido: {e}") from e