Warm predictor: torch.compile + GPU-resident export

[aperson30]

Two speed additions to the persistent LesionSegmenter warm service (lesionseg_warm_server.py), both of which only make sense for a warm/persistent process.

torch.compile (env-gated, LESIONSEG_COMPILE, default off)

Compiles the patch-level network once at startup (max-autotune-no-cudagraphs). The sliding-window patch is a constant [128,224,224], so one compile serves every patch of every case.

• 1.18× on inference, accuracy-safe (voxel agreement 0.99996, lesion Dice 0.999 vs eager — measured over 20 runs on an idle node with zero latency stalls).
• The one-time compile (~20s–3min) is paid at startup behind the health gate: while it runs, the health check is unreachable and lesionseg_predict.py uses the cold path, so the site keeps working.
• Env-gated so it can be dropped without a redeploy if a node shows compile-latency variance (seen on some shared GB10 nodes). The warm predictor’s main win — cold-start elimination — is independent of compile.

GPU-resident export (default for softmax-argmax models)

predict_logits returns logits on the CPU, so convert_logits moves the ~1.5 GB, 43-channel volume CPU→GPU for the (GPU-patched) resample and back. This runs the same resample, spacings, and argmax GPU-resident and returns only the uint8 label map.

• Bit-identical to convert_logits (0 differing voxels across a size-stratified sample incl. anisotropic cases); 3.6–25× on the export stage. Same logic as the already-merged LesionSegmenter: GPU-resident export (no CPU round trip) #121.
• Region-based models fall back to convert_logits.

Deploy notes (not automatic)

Going live also requires, on the server: starting the warm service, and setting LESIONSEG_WARM_URL in .env so lesionseg_predict.py routes to it (with automatic cold-path fallback if it is down/busy/mismatched). No frontend/DB change.

Testing

Verified the service starts, compiles (17.7s on an idle node), and listens; the export logic is identical to #121 which was validated bit-exact end-to-end.