[JESUSROYETH]

--model.capture_network true now replays the backbone and the neck for segmentation too, and it can be combined with capture_segment_heads — the prediction groups keep independent gradients and the native loss arithmetic. The flag is off by default, and accumulation or an altered head/forward dispatch falls back to the native path.

I ran five alternating pairs on an isolated L4, against main (b293970), with the existing training options on: nano segmentation, batch 4, 640 px, BF16, five epochs, 512 COCO val2017 images for training and 16 disjoint images for validation. Complete-fit time dropped 18.42%:

Seconds, median (range)	Native features	Replay
Complete fit	107.412 (106.764–109.816)	87.629 (87.043–88.462)
Whole process	112.721 (111.918–115.208)	92.937 (92.471–93.827)

Model, optimizer, scheduler, EMA and per-step-loss bytes match across the ten fits. Peak reserved CUDA memory rises from 5,496 to 5,702 MiB. Focused CPU/CUDA regressions and the full pre-commit, strict mypy included, pass.

I also screened batch 16 once and it only saved 2.90%, so the claim above stays at batch 4. Full make gate, the accelerator goldens and long-run accuracy are still unverified; the measured scope is written up in WP-193g.

Clean-room acknowledged: this only touches native stage composition, our own profiles and the registered PyTorch API R43 — no new numerical choice.

WP: 193g
Provenance: R43; native shared stages and measured CUDA profiles
Assumptions: none
Gate: tests/ptl/test_segment_network_capture.py

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (12 files)

• CHANGELOG.md — added WP-193g entry; consistent with existing entries
• docs/ROADMAP.md — added WP-193g row; dependency links updated in 193d/193e/194
• docs/TRAINING.md — segmentation capture_network paragraph added; task list extended to include task: segment
• docs/engineering/phase-17-pose-objective-parity.md — WP-193g include added in phase order
• docs/work_packages/wp-193d.md — "Needed by" updated to WP-193g
• docs/work_packages/wp-193e.md — "Needed by" and navigation updated to WP-193g
• docs/work_packages/wp-193g.md — new work-package log
• docs/work_packages/wp-194.md — navigation updated
• src/lucid_yolo/ptl/module.py — _segmentation_features, _EAGER_SEGMENT_FORWARD, topology capture, validation and docstrings
• src/lucid_yolo/ptl/network_capture.py — docstring and run_features retirement message generalised from "pose" to "shared network"
• tests/ptl/test_network_capture.py — parametrised cases and error regex updated
• tests/ptl/test_pose_network_capture.py — retirement warning text updated
• tests/ptl/test_segment_network_capture.py — new test module

---

_{Reviewed by free · Input: 314.4K · Output: 46.8K · Cached: 779.8K}

[Copilot]

🔵 Needs a closer look

CUDA backward replay needs human review while full-gate, accelerator-golden and long-run accuracy validation remain unverified.

No open issues 🎉

What changed in this PR

Extends opt-in CUDA network replay to segmentation’s backbone and neck while preserving independent prediction gradients and native loss computation.

Changes:

• Adds segmentation replay routing, compatibility checks and native fallbacks.
• Adds parity, lifecycle, fallback and CLI regression tests.
• Documents usage, measured performance and remaining validation gaps.

File	Description
tests/​ptl/​test_segment_network_capture.py	Adds segmentation replay regression coverage.
tests/​ptl/​test_pose_network_capture.py	Updates the frozen-feature warning assertion.
tests/​ptl/​test_network_capture.py	Updates option compatibility expectations.
src/​lucid_yolo/​ptl/​network_capture.py	Generalizes shared-feature documentation and warnings.
src/​lucid_yolo/​ptl/​module.py	Enables guarded segmentation feature replay.
docs/​work_packages/​wp-194.md	Updates work-package navigation.
docs/​work_packages/​wp-193g.md	Records implementation, benchmarks and validation limits.
docs/​work_packages/​wp-193e.md	Links the new dependent work package.
docs/​work_packages/​wp-193d.md	Links the new dependent work package.
docs/​TRAINING.md	Documents segmentation replay usage and constraints.
docs/​ROADMAP.md	Adds WP-193g and its validation status.
docs/​engineering/​phase-17-pose-objective-parity.md	Includes the new engineering record.
CHANGELOG.md	Announces segmentation feature replay.

_{Review effort: Balanced}