[Borda]

_SegmentHeadCapture.clear() now resets the holder before draining the recorded streams, so a failing synchronize still leaves no recordings behind. _SegmentHeadCaptureLifecycle.on_exception reports that drain failure as a warning, so the trainer's own exception reaches the caller and its teardown runs; on_fit_end still raises it. _holder returns None for a module without a segmentation holder and every hook is then a no-op, as in network_capture.py.

Guard test names drop filler words, and the stage-helper certificate test lists proto_fusion among the certified stages.

WP: 193e
Provenance: none (review remediation of replay lifecycle error handling; repository code only)
Assumptions: none
Gate: tests/ptl/test_segment_head_capture_guards.py::TestRecordingLifetime, tests/ptl/test_segment_head_capture.py::test_stage_helpers_and_shared_storage_are_not_certified

[Copilot]

🟡 Changes recommended

Promoted warnings can still mask the original training exception and interrupt teardown.

1 open issue

• Prevent UserWarning promotion from masking training exceptions · New

What changed in this PR

This PR hardens segmentation replay cleanup after failed training while preserving drain errors at normal fit completion.

Changes:

• Reset capture state before draining streams and warn on failed-fit drain errors.
• Make lifecycle hooks no-ops for modules without a segmentation holder.
• Add lifecycle regression tests, simplify test names, and include proto_fusion in certification coverage.

File	Description
tests/​ptl/​test_segment_head_capture.py	Include proto_fusion in stage certification.
tests/​ptl/​test_segment_head_capture_guards.py	Cover drain failures and missing holders; simplify names.
src/​lucid_yolo/​ptl/​segment_head_capture.py	Harden capture reset and lifecycle cleanup.

_{Review effort: Balanced}