fix(vision): file non immagine e grafo malformato non escono più come 500 (#15)

Worker: Image.open().convert() in try/except (422 "image is not
decodable"); il grafo deve essere un dict (422 altrimenti); run_graph
distingue costruzione (GraphBuildError -> 422) da esecuzione (500 con
motivo). Router: verifica dei magic bytes (PNG/JPEG/BMP/TIFF/WebP/GIF)
prima di chiamare il worker.

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

src/vision/runner.py

@@ -22,6 +22,14 @@ from visionsuite.task import Task
 _SUBMODULE = Path(__file__).resolve().parents[2] / "vendor" / "visionsuite"
 
 
+class GraphBuildError(ValueError):
+    """`Task.from_dict` refused the graph itself - a bad request, not a
+    server fault (a missing "kind", an unknown schema version, a graph that
+    is not even a dict). Kept apart from an error raised once the task is
+    actually running: the worker maps this one to 422, everything else that
+    can go wrong inside `task.run()` to 500."""
+
+
 @dataclass
 class ToolFailure:
     tool_id: str
@@ -103,8 +111,21 @@ def run_graph(
     `image` is whatever the caller already has in hand: a plain numpy array is
     wrapped into the `visionsuite.core.image.Image` that vs-measure's tools
     actually read via `.pixels`; an `Image` passed in is used as-is.
+
+    Two phases, on purpose: building the task from `graph` can fail on
+    nothing but the caller's input (a bad schema version, a tool with no
+    "kind", a graph that is not a dict at all) - that is `GraphBuildError`,
+    a 422 for the worker. Once the task exists, `task.run()` is the engine's
+    own business; an exception from there is not the caller's mistake, and
+    the worker must not call it one.
     """
-    task = Task.from_dict(graph)
+    try:
+        task = Task.from_dict(graph)
+    except (KeyError, TypeError, AttributeError, ValueError) as exc:
+        raise GraphBuildError(
+            f"malformed graph: {type(exc).__name__}: {exc}"
+        ) from exc
+
     wrapped_image = image if isinstance(image, Image) else Image(pixels=np.asarray(image))
     context: dict[str, Any] = {"image": wrapped_image}
     if calibration is not None: