v0.6.0: compaction-aware refill turns recovered plate into fewer sheets

[Steel-tech]

Implements docs/plans/2026-08-24-001-feat-compaction-aware-refill-plan.md (included in this PR). v0.5.0's verified true-shape compaction (#10) recovered plate area, but it ran only after every part was already assigned to a plate — so the recovered space enlarged remnant candidates while the sheet count and stranded parts stayed exactly what bounding-box packing decided. This PR closes that loop: recovered plate now becomes fewer purchased sheets.

What changed

Refill inside the packing loop — when a part fails to insert on every compatible open plate, eligible plates (all irregular parts outlined) that received new placements since their last attempt are compacted behind the same independent verification gate as v0.5.0, and the insertion is retried against the rebuilt free rectangles — before a new sheet is opened, and before the part is declared stock_exhausted. Any successful insertion re-marks the plate dirty, and the v0.5.0 end-of-run pass still tightens final layouts.

One shared gate helper — the compact → verify → accept/restore/disqualify block is now _attempt_plate_compaction, used by both the mid-pack and end-of-run paths. compaction.recovered_in and compaction.passes accumulate across attempts on the same plate; accepted reflects whether the current layout comes from an accepted compaction. No schema changes.

Safety argument

• Refilled parts land only in free rectangles rebuilt by subtracting every compacted bounding box plus spacing, so a refilled part keeps at least kerf+gap clearance to every bounding box — and therefore to every true profile — on the plate. Mixed compacted/refilled plates still pass the full true-shape pairwise verification.
• A rejected compaction restores the plate, records the non-blocking COMPACTION_REJECTED finding, permanently disqualifies that plate for the run, and packing proceeds exactly as the bounding-box flow would have (proved by the mid-pack corruption test).
• Never worse: with compaction enabled a run uses at most as many plates and strands at most as many parts as with it disabled — compacting before opening a sheet can only add placement options (asserted across fixtures).
• Bounded work: a plate is re-attempted only when dirty (new placements since last attempt), and each attempt keeps v0.5.0's pass/step bounds.
• Burn-DXF eligibility, review posture, metrics definitions, remnant non-certification: unchanged. Algorithm version bumped (maxrects-bssf-u3-tsc2), which reaches the configuration hash as before; normalized input hashes untouched.

Tests (5 new; 211 total green)

• Sheet reduction: the L/Z interlock plus a 4×6 filler needs 2 sheets boxed, 1 sheet with refill, with the filler landing exactly one clearance off the compacted Z's right face (x = 11.5).
• Stranded-part rescue: with qty 1 stock the boxed run is blocked (stock_exhausted); with refill it completes as review_required with nothing unplaced.
• Never-worse guarantee across fixtures, determinism, and the mid-pack rejection fallback (corrupt compactor → plate disqualified → identical-to-boxed layout on 2 sheets, one warning, verification still verified).

npm run release:check fully green (tests, ruff, privacy, pack, provenance); full-render smoke green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MquyTzVAPKpV1V8Vt3XxnJ

---

Generated by Claude Code

[coderabbitai]

Warning

Review limit reached

Next included review available in 42 minutes.

View limit details

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 490b95b4-7730-4cf6-ac53-9693e1f16ad1

📥 Commits

Reviewing files that changed from the base of the PR and between 3007214 and 15d4094.

📒 Files selected for processing (8)

• CHANGELOG.md
• README.md
• docs/plans/2026-08-24-001-feat-compaction-aware-refill-plan.md
• package.json
• pyproject.toml
• skills/steel-nest/SKILL.md
• skills/steel-nest/scripts/nest.py
• tests/test_compaction.py

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Steel-tech]

🤖 Lab Code Review (draft opinion)

• a/skills/steel-nest/scripts/nest.py:1034: plate["compaction_dirty"] = True should be set only on successful placement to avoid unnecessary compaction attempts; move inside the if placement: block.
• a/skills/steel-nest/scripts/nest.py:1056-1083: The refill loop retries insertion on the same plate after compaction without re-checking compatibility; add if not compatible(plate["stock"], unit): continue after compaction to prevent invalid attempts.
• a/skills/steel-nest/scripts/nest.py:993: plate["compaction"] initialization in open_plate duplicates later reset in the end-of-run loop; remove the redundant block to avoid confusion.
• a/skills/steel-nest/scripts/nest.py:1094-1104: End-of-run compaction uses _attempt_plate_compaction but ignores its return value; the helper already handles disqualification and findings, so the call is correct but the prior manual compaction block removal left dead code—verify no remnants remain (already addressed in diff).
• a/skills/steel-nest/scripts/nest.py:339: _compaction_eligible returns bool(irregular) and all(...); if irregular is empty, it returns False, which is correct, but the function name implies eligibility for compaction—consider renaming to _has_outlined_irregular_parts for clarity (non-critical).

Only one critical finding is required; the rest are minor. The most severe issue is the missing compatibility check post-compaction in the refill loop, which could lead to invalid placements. However, since the diff shows the loop skips incompatible plates before compaction and the compaction helper does not alter stock compatibility, this is actually safe. Re-evaluating: the refill loop iterates over plates, checks compatible(plate["stock"], unit) at the start, and only proceeds if true. Compaction does not change plate["stock"], so re-checking is unnecessary. Thus, no correctness bug exists.

The only actionable issue is the premature compaction_dirty flag: setting it on every commit (line 1034) marks the plate dirty even if no placement occurred, wasting compaction cycles. This violates R5 (bounded work) by potentially increasing attempts beyond necessity.

File: a/skills/steel-nest/scripts/nest.py, line 1034
Fix: Move plate["compaction_dirty"] = True inside the if placement: block, directly after commit(plate, unit, placement).

All other points are either safe or non-critical. The diff is otherwise correct.

Final answer: one finding.

a/skills/steel-nest/scripts/nest.py:1034: Move plate["compaction_dirty"] = True inside the if placement: block to avoid unnecessary compaction attempts.

[Steel-tech]

Verified the lab reviewer's draft opinion against the code — its single actionable finding does not hold, and no change is needed:

"Move plate["compaction_dirty"] = True inside the if placement: block" — the flag is already set exactly and only when a placement is committed. It is the last line of commit() (nest.py:1037), and commit(plate, unit, placement) is invoked at three call sites (lines 1056, 1079, 1089), each guarded by if placement:. There is no code path that sets the dirty flag without a successful placement, so no extra compaction attempts occur and the R5 bounded-work property is intact — a plate is attempted only after receiving a new placement, as the plan specifies.

The comment's remaining points it withdrew itself (correctly: the compatibility check before compaction is sufficient since compaction never changes plate["stock"]). One correction to its third point: open_plate's compaction-dict initialization is not duplicated — the end-of-run loop no longer resets that dict; it only calls the shared _attempt_plate_compaction, which accumulates into it.

---

Generated by Claude Code