fix(vision): /run del worker non blocca più l'event loop, timeout e semaforo (#34)

Decodifica + run_graph girano in asyncio.to_thread sotto asyncio.wait_for
(VISION_RUN_TIMEOUT_S, default 100s, sotto i 120s del backend) -> 504 del
worker allo scadere. /health resta leggero e risponde durante un calcolo
in corso. asyncio.Semaphore(VISION_MAX_CONCURRENCY, default 2): oltre ->
503 "worker busy". vision_service._call_worker inoltra 422/503/504 del
worker con il loro status e motivo invece di appiattirli su 502.

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

src/backend/services/vision_service.py

@@ -21,6 +21,14 @@ 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
 
+# Worker responses whose status and reason are forwarded as-is rather than
+# flattened into a generic 502 - see `_call_worker`.
+_FORWARDED_WORKER_STATUSES = frozenset({
+    status.HTTP_422_UNPROCESSABLE_CONTENT,
+    status.HTTP_503_SERVICE_UNAVAILABLE,
+    status.HTTP_504_GATEWAY_TIMEOUT,
+})
+
 
 def _is_measurable(value: object) -> bool:
     """A worker output counts as a quote only if it is a finite, non-bool number.
@@ -69,17 +77,18 @@ async def _call_worker(image_bytes: bytes, graph: dict) -> dict:
             detail=f"the vision worker is unreachable: {exc}",
         ) from exc
 
-    if response.status_code == status.HTTP_422_UNPROCESSABLE_CONTENT:
-        # The worker already diagnosed a bad request (undecodable image,
-        # malformed graph) - forward its reason instead of flattening it
-        # into a generic 502.
+    # The worker already gave this fault a name - forward its status and
+    # reason instead of flattening every non-200 into a generic 502: 422 is a
+    # bad request (undecodable image, malformed graph), 503 is "busy, try
+    # again" (VISION_MAX_CONCURRENCY exceeded), 504 is its own run timeout
+    # (VISION_RUN_TIMEOUT_S, always tighter than this client's own 120s so
+    # the worker answers first).
+    if response.status_code in _FORWARDED_WORKER_STATUSES:
         try:
             detail = response.json().get("detail", response.text)
         except ValueError:
             detail = response.text
-        raise HTTPException(
-            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT, detail=detail,
-        )
+        raise HTTPException(status_code=response.status_code, detail=detail)
 
     if response.status_code != 200:
         raise HTTPException(