Swallowed errors: serve.py 500s on bad input, capture proxy kills connections, has_cuda silent fallback

[thedataengineer]

Places where a failure is swallowed or misreported:

1. serve.py:1473 blanket except turns validation errors into 500s. The whole do_POST route table is wrapped in one except Exception that returns 500 {type}: {e}. Unchecked field reads like b["model"] (serve.py:1440), b["prompt"] (:1447), b["messages"] (:1459) surface as 500 KeyError: 'model' instead of a 4xx with a usable message.

2. capture.py proxy dies instead of returning an error body. capture.py:113 does json.loads(self.rfile.read(int(self.headers["Content-Length"]))) with no handling for a missing header or bad JSON, and _serve() (capture.py:148) has no try/except around model.chat(). Any failure kills the connection, so the agent under capture sees a transport error rather than an OpenAI-shaped error response. That undercuts the drop-in-proxy contract, since real agent frameworks retry on API errors but often crash on transport errors.

3. backends/torch.py:74 has_cuda() is a bare except Exception: return False with no annotation or log. A broken CUDA install silently degrades select_backend("auto") to CPU, which contradicts the accel.py property of logging exactly what was enabled.

4. serve.py:788-789 and serve.py:824-825 swallow torch.cuda.empty_cache() failures with unannotated except Exception: pass. In the VRAM-pressure retry path the subsequent _load() then fails opaquely, and clear_vram() reports {"unloaded": n} success even when the cache clear threw.

5. Frontend: 23 swallowed promise rejections (.catch(() => {}) across App.tsx, pages/Models.tsx, pages/Dashboard.tsx, pages/Datasets.tsx, pages/Runs.tsx, pages/Playground.tsx, pages/Train.tsx, pages/Machines.tsx). Polling loops recover, but a first-load failure renders as an empty state with no error surface, indistinguishable from "no data yet".

[thedataengineer]

Fixed in 204ee33: the capture proxy returns OpenAI-shaped 400/500 error bodies instead of dropping the connection (and now records the OS-assigned port when built with port=0); /v1/generate, /v1/chat and /v1/prewarm 422 naming the missing field instead of 500 KeyError; torch.has_cuda() logs a failed CUDA probe before falling back to CPU; and the two cuda.empty_cache() swallows log what threw, with clear_vram reporting the error rather than a clean sweep it didn't achieve. 12 new tests.

Not addressed: the 23 swallowed promise rejections in the frontend — that's a UI error-surface design call, better done as one pass with a toast/error-state component than 23 scattered catch blocks.