perf(rfdetr-seg): skip mask→poly→mask round-trip on workflow path

[aseembits93]

Summary

When the instance-segmentation adapter is invoked from a workflow, the result goes through a pure-overhead encoding round-trip:

adapter.postprocess:
    GPU masks → masks2poly (cv2.findContours, N times)
              → List[Point(x,y)] pydantic validation per vertex
              → InstanceSegmentationPrediction (validated)
v3 block.run_locally:
    predictions → model_dump(by_alias=True) per response
                → sv.Detections.from_inference → polygon_to_mask (rasterize AGAIN)

Nothing between the adapter output and the sv.Detections sink observes the polygon form, yet we pay polygon extraction + pydantic validation + polygon→mask rasterization per frame.

This change short-circuits the round-trip when request.source == "workflow-execution":

• The adapter builds sv.Detections directly from the GPU-derived numpy arrays and attaches it via response.__dict__["_sv_detections_fast"]. Pydantic v2 ignores extra __dict__ keys in model_dump / jsonable_encoder, so HTTP callers are unaffected.
• The v3 block detects the marker and routes through a new _post_process_result_fast, which attaches detection_id/parent_id/image_dimensions/inference_id directly onto the pre-built sv.Detections and skips model_dump + convert_inference_detections_batch_to_sv_detections entirely.

Falls back to the existing polygon path whenever the marker is absent (HTTP responses, RLE responses via response_mask_format=rle, non-tensor masks, mixed-source batches).

Benchmark

rfdetr-seg-nano (TRT) + Triton preproc + full-Triton postproc + CUDA graphs, vehicles_312px.mp4 (538 frames) via InferencePipeline, T4 GPU:

	Run 1	Run 2	Run 3	Run 4	mean
baseline (main)	151.96	151.29	151.39	152.54	151.80
this change	165.63	164.09	163.21	165.09	164.51

+12.7 FPS, ~+8.4%. Same flags, same 538-frame window.

Test plan

• pytest tests/workflows/unit_tests/core_steps/models/roboflow/instance_segmentation/test_v3.py — 23/23 pass
• Manual benchmark: 4 baseline runs vs 4 optimized runs on the same video with the same env flags (RFDETR_USE_TRITON_PREPROC=true RFDETR_TRITON_FULLPOSTPROC=true ENABLE_AUTO_CUDA_GRAPHS_FOR_TRT_BACKEND=true)
• Exercise the HTTP path to confirm serialized payloads are byte-identical to main (no leakage of the _sv_detections_fast private attr through pydantic serialization)
• Exercise RLE response path (response_mask_format=rle) to confirm it still goes through the original polygon/RLE branch

🤖 Generated with Claude Code