perf(is-seg): swap pydantic response for dataclass twins on workflow path (stacked on #22)

[aseembits93]

Summary

Stacked on #22. Non-stacked version against main is #28. Open this PR once #22 lands; then its diff against main will be identical to #28's (modulo target-branch differences).

InferenceModelsInstanceSegmentationAdapter.postprocess built a full pydantic tree per frame — Point × V per polygon vertex, InstanceSegmentationPrediction × N, then InstanceSegmentationInferenceResponse. The workflow block then called response.model_dump(by_alias=True, exclude_none=True) to get a plain dict for sv.Detections.from_inference. Neither the validation nor the serializer is needed on that path — the block only consumes the dict.

This change adds slotted dataclass twins (PointDC, InferenceResponseImageDC, InstanceSegmentationPredictionDC, InstanceSegmentationInferenceResponseDC) plus _is_pred_dc_to_dict / _is_response_dc_to_dict helpers that emit the exact dict model_dump(by_alias=True, exclude_none=True) produces.

The adapter gates on kwargs.get("source") == "workflow-execution" and returns the dataclass response on that path. Every other caller — HTTP response_model at http_api.py:1640, isinstance-based cache dispatch at cache/serializers.py:71, draw_predictions visualization — keeps the pydantic path untouched.

The v3 workflow block detects the dataclass via isinstance and calls _is_response_dc_to_dict; falls back to model_dump for any other response type.

Numbers

Microbench (4 dets × 6-vertex polygon, construct + dump):

	µs/frame
pydantic	~81
dataclass	~34

Δ ≈ 2.4× faster.

End-to-end on top of #22's full stack (rfdetr-seg-nano TRT + Triton preproc + Triton fullpost + CUDA graphs, vehicles_312px.mp4, 538 frames, 4 runs each):

	run 1	run 2	run 3	run 4	mean
baseline (optimize-rfdetr-seg HEAD)	152.55	152.62	153.23	153.33	152.93
this change	156.33	156.88	157.10	155.86	156.54

Δ ≈ +3.6 FPS (+2.4%).

Correctness

Bit-exact parity: _is_response_dc_to_dict(dc) == pyd.model_dump(by_alias=True, exclude_none=True) on mixed inputs including varying polygon lengths, empty polygons, and post-construction mutation of .time / .inference_id (assigned by Model.infer_from_request at inference/core/models/base.py:154-157). @dataclass(slots=True) permits the reassignment because those fields are declared.

Test plan

• pytest tests/workflows/unit_tests/core_steps/models/roboflow/instance_segmentation/test_v3.py — 23/23 pass
• Parity test (_is_response_dc_to_dict(dc) == pyd.model_dump(by_alias=True, exclude_none=True))
• Microbench
• End-to-end FPS benchmark on optimize-rfdetr-seg's full Triton stack
• HTTP regression: hit /infer/instance_segmentation locally and confirm JSON is byte-identical to pre-change (expected — adapter falls through to pydantic when source != "workflow-execution")

🤖 Generated with Claude Code