fix(aidd-dev): a review appends a round and scores the whole plan

[blafourcade]

🎯 What & why

One phase of a 7-phase plan, one diff, three models: 100% (6/6) on 7 blocks, 100% (6/6) on 1 block, 53% (16/34) where 16/34 is 47%, and boxes copied from the previous run. 05-review never said what a phase outside the diff becomes, how the percent is computed, or where a past round lives.

🛠️ How it works

A review is appended to, one section per round, and never rewritten. A round carries its date, its author from git config user.name, its diff, the axes that ran, a verdict and a score. Its Criteria list holds the plan's criteria, its Findings list the defects no criterion covers.

The score replaces the percent that had no definition: it counts over the plan's whole criteria total, in three numbers that sum to it — 1/6 met, 1 unmet, 4 out of the diff. One denominator, so a plan covered at a third can no longer read as finished. A phase the diff leaves untouched gets one Out of the diff: line instead of a block of empty boxes.

The skill now follows the authoring contract aidd-context:04-skill-generate applies to itself: the router holds the flow, the action table and one transversal rule, and nothing else. 01-prepare resolves the diff once for the three axes and opens the round; 05-finalize judges it and checks its shape. references/report-contract.md holds where the report lives, what diff is reviewed and how a round joins it. SKILL.md, the only always-on text, goes from 446 words to 189.

🧪 How to verify

node scripts/check-tests-leave-git-alone.js -- node --test 'scripts/__tests__/**/*.test.js'   # 580 pass, 0 fail
claude -p --plugin-dir "$PWD/plugins/aidd-dev" "Run the aidd-dev:05-review skill, functional axis only, on <plan.md>, for the diff HEAD~1...HEAD."

• Guard: 26 assertions, 34 mutations, each reddening one assertion alone.
• Volume: 55 runs over 11 fixtures, every fixture unanimous. A plan of 10 phases and 34 criteria whose diff touches 2 phases reads 3/34 met, 5 unmet, 26 out of the diff, 5 runs of 5.
• Against next, same fixtures, same prompts, counting checked boxes against the truth: 6 wrong runs of 18 for next, 0 of 45 here. next both over-credits (6 boxes where 5 hold, crediting a stock: "untouched" label) and under-credits (a criterion evidenced by a context line, a not-applicable one).
• Callers: aidd-orchestrator:01-sdlc, 00-async-dev and the ship-a-feature recipe name the skill and its findings, never a section or a percent. Nothing downstream reads the artifact's shape.

⚠️ Heads-up

• The artifact changes shape. review.md becomes a log of rounds. That is not what fix(aidd-dev): review leaves out-of-diff phases and the verified percent undefined #858 asked for, and the issue names "the verified percent" that the Score line replaces — to settle before merge.
• Four rules come from the volume campaign, each turning a split between runs into unanimity: a context line of the diff is evidence; a criterion whose subject is absent is a gap, not not-applicable; a phase projecting no file is traced, not out of the diff; a label echoing the criterion's words is not evidence. They cost 03-review-functional 97 words, and they are the only words in it whose effect is measured.
• The skill is 1983 words against 1785 on next. The always-on router shrank by 58%; the total grew because 01-prepare and 05-finalize own work that was implicit, and because the contract requires a | Case | Pass | table per action.
• Cross-model is unresolved. On haiku, both designs produce a report about half the time and neither is clearly better on the landed runs. haiku did expose a real defect, now fixed: nothing in the skill named review.md any more, and a reference named review-report.md sat one hyphen away, so haiku wrote its report there.
• references/review-rubric.md keeps 🟡 warning in a scale whose neighbours are critical and minor. major would be consistent; it touches 04-audit, 03-assert and 07-refactor, so it belongs to its own change.

🔗 Linked issue

Refs #858

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VE16AxC1V9wAMGmGgefe4z