fix(server): load pipeline once at startup and return XML file directly

[haoyu-haoyu]

Summary

Three critical issues in server_pa.py fixed in a single focused PR:

1. Model reloaded on every request (performance)

Pipeline(config) was instantiated inside the /convert handler, causing SAM3 model weights to be loaded into GPU memory on every single upload. For a multi-GB model this means:

• 10-30s latency added to every request just for model loading
• GPU memory fragmentation and potential OOM after repeated requests

Fix: Singleton Pipeline created once during FastAPI lifespan startup. All requests share the same instance.

2. API returned local path instead of file (usability)

The response was {"output_path": "/server/local/path/xxx.drawio"} — a local filesystem path that API callers cannot use. This makes the entire API non-functional for any remote client.

Fix: /convert now returns the actual .drawio XML file as a FileResponse download with proper Content-Type: application/xml and Content-Disposition headers.

3. No upload size limit (stability)

No file size validation meant arbitrarily large uploads could crash the server with OOM errors.

Fix: Added a 20 MB cap (MAX_UPLOAD_BYTES) with a clear HTTP 413 response.

Other improvements

• Proper logging module instead of bare print()
• FastAPI lifespan context manager (modern best practice, replaces deprecated on_event)
• Version bumped to 1.1.0

Before / After

# Before: every request loads the full model
@app.post("/convert")
async def convert(file: UploadFile = File(...)):
    pipeline = Pipeline(config)  # SAM3 loaded here every time!
    ...
    return {"output_path": result_path}  # useless local path

# After: model loaded once, file returned directly
@app.post("/convert")
async def convert(file: UploadFile = File(...)):
    pipeline = get_pipeline()  # singleton, instant
    ...
    return FileResponse(result_path, filename="output.drawio")  # actual file

Test plan

• python server_pa.py starts and logs "Pipeline ready"
• curl -F "file=@test.png" http://localhost:8000/convert -o result.drawio returns valid XML
• Uploading a >20MB file returns HTTP 413
• Second request is fast (no model reload)
• GET /health returns {"status": "ok"}