[JESUSROYETH]

--model.capture_network true now supports native OBB training, replaying the existing backbone, neck and all six head outputs. Assignment, losses and state updates keep their current arithmetic, and the flag defaults off. Accumulation and changed objectives still run the native path, so the absent-gradient semantics stay the same.

I ran five alternating pairs on an isolated L4, and complete-fit time dropped 12.01%:

Seconds, median (range)	Native	Replay
Complete fit	68.642 (68.213–69.333)	60.397 (59.924–60.784)

Setup: Nano OBB, 104 public DOTA tiles plus 13 validation tiles from disjoint source images, eight epochs, batch 4, 1024 px, BF16, deterministic kernels, on a g2-standard-8 with an L4, driver 580.178.04, torch 2.9.1+cu129, Lightning 2.6.6. Both arms have prepared caching, device HSV, epoch prefetch and optimizer replay enabled, and the timing covers setup, recording, training, EMA, validation and checkpoints.

Model, optimizer, scheduler, EMA and every loss match by byte hash, and validation agrees too. Thirty focused cases passed on torch 2.4 and 2.9 (actual replay, empty/partial batches and AMP recovery/fallback), plus five added CPU cases covering the replaced decode/loss helpers. Full pre-commit and strict mypy pass. Reserved CUDA memory goes up from 3974 to 4300 MiB. A single batch-8 probe only saved 4.10%, so the claim stays at batch 4. Full make gate, accelerator goldens and long-run accuracy are still unverified — more detail in WP-200.

Clean-room acknowledged: unchanged repository expressions, our own CUDA profiles and the registered PyTorch API R43, no external model implementation consulted.

WP: 200
Provenance: R43; native repository stages and measured CUDA profiles
Assumptions: none
Gate: tests/ptl/test_oriented_network_capture.py

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (16 files)

• CHANGELOG.md
• docs/PROVENANCE.md
• docs/ROADMAP.md
• docs/TRAINING.md
• docs/engineering/phase-17-pose-objective-parity.md
• docs/work_packages/wp-193d.md
• docs/work_packages/wp-199.md
• docs/work_packages/wp-200.md
• scripts/lint/audit_docs_present.py
• scripts/_tests/test_audit_docs_present.py
• src/lucid_yolo/ptl/loss_capture.py
• src/lucid_yolo/ptl/module.py
• src/lucid_yolo/ptl/network_capture.py
• tests/ptl/test_network_capture.py
• tests/ptl/test_obb_training.py
• tests/ptl/test_oriented_network_capture.py

Previous Review Summaries (3 snapshots, latest commit ec54fee)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit ec54fee)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (20 files)

• src/lucid_yolo/models/heads/detect.py
• src/lucid_yolo/ptl/loss_capture.py
• src/lucid_yolo/ptl/module.py
• src/lucid_yolo/ptl/network_capture.py
• src/lucid_yolo/ptl/segment_head_capture.py
• scripts/lint/audit_docs_present.py
• scripts/_tests/test_audit_docs_present.py
• tests/ptl/test_network_capture.py
• tests/ptl/test_obb_training.py
• tests/ptl/test_oriented_network_capture.py
• tests/ptl/test_segment_head_capture.py
• tests/ptl/test_segment_head_capture_guards.py
• docs/CHANGELOG.md
• docs/ROADMAP.md
• docs/TRAINING.md
• docs/work_packages/wp-193c.md
• docs/work_packages/wp-193d.md
• docs/work_packages/wp-193e.md
• docs/work_packages/wp-200.md
• docs/engineering/phase-17-pose-objective-parity.md

Previous review (commit 97703ed)

This review did not finish. The model reached its output limit before it
could write the review — a reasoning model can spend the whole budget thinking.
Re-run the review, or lower the model's thinking effort, and it should get
further. Any inline comments below are from an earlier review.

Previous review (commit 97703ed)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (12 files)

• CHANGELOG.md
• docs/PROVENANCE.md
• docs/ROADMAP.md
• docs/TRAINING.md
• docs/engineering/phase-17-pose-objective-parity.md
• docs/work_packages/wp-193d.md
• docs/work_packages/wp-199.md
• docs/work_packages/wp-200.md
• scripts/lint/audit_docs_present.py
• src/lucid_yolo/ptl/module.py
• src/lucid_yolo/ptl/network_capture.py
• tests/ptl/test_network_capture.py
• tests/ptl/test_oriented_network_capture.py

---

_{Reviewed by free · Input: 151.3K · Output: 16.9K · Cached: 823.8K}

[Copilot]

🔵 Needs a closer look

CUDA forward/backward replay needs human review while the full gate, accelerator goldens and long-run accuracy remain unverified.

No open issues 🎉

What changed in this PR

Extends opt-in CUDA network replay to oriented detection training, preserving native loss computation and state updates.

Changes:

• Records all six oriented head outputs, with native fallback for accumulation and modified objectives.
• Adds replay parity, gradient, precision and fallback tests.
• Documents measured performance, memory costs and validation limits.

File	Description
tests/​ptl/​test_oriented_network_capture.py	Adds oriented replay and fallback coverage.
tests/​ptl/​test_network_capture.py	Updates supported-option expectations.
src/​lucid_yolo/​ptl/​network_capture.py	Adds six-output oriented adapter.
src/​lucid_yolo/​ptl/​module.py	Enables guarded oriented replay.
scripts/​lint/​audit_docs_present.py	Raises work-package floor to 200.
scripts/​_tests/​test_audit_docs_present.py	Makes floor fixtures dynamic.
docs/​work_packages/​wp-200.md	Records implementation, measurements and limits.
docs/​work_packages/​wp-199.md	Links dependent work package.
docs/​work_packages/​wp-193d.md	Links dependent work package.
docs/​TRAINING.md	Documents oriented replay requirements.
docs/​ROADMAP.md	Adds WP-200 tracking.
docs/​PROVENANCE.md	Records API provenance.
docs/​engineering/​phase-17-pose-objective-parity.md	Includes WP-200 engineering log.
CHANGELOG.md	Announces oriented replay support.

_{Review effort: Balanced}