fix(aidd-dev): review leaves out-of-diff phases and the verified percent undefined #858

Problem

05-review does not say what happens to plan phases outside the reviewed diff, nor how the verified percent is computed. Reviewing one phase of a multi-phase plan, three models on the same skill produced three incompatible review.md files.

The skill contradicts itself:

• SKILL.md:30 says review.md is overwritten each run: "a snapshot of the current diff, not a history".
• 02-review-functional.md:23 says the Phases section holds "one block per plan phase".
• Nothing says whether the previous review.md is read before writing.
• review-template.md:33 renders {{pct}}% ({{n_checked}}/{{n_total}}), but nothing says which phases enter n_total. not-applicable is only defined as "neither a finding nor a blocker" (review-rubric.md:19), never for the percent, and never when to use it.

Observed on a 7-phase plan, phase 4 reviewed in three worktrees, aidd-dev 2.4.1 (05-review is unchanged in 2.5.0):

• Mistral Medium and Qwen have word-for-word identical blocks for phases 1, 2, 3, 5, 6 and 7: both copied the previous review.md. Mistral's phase 2 still says "ajoutés par ce diff" while its own Files checked lists only phase-4 files. Copied evidence reads as fresh verification.
• GLM dropped the six other phases.
• Qwen tagged the 15 criteria of out-of-diff phases not-applicable, Mistral used the same tag for a partially verifiable phase-4 criterion: one tag, two meanings.

Scope

• Functional review reads only plan.md, its phase files, and the current diff. It never reads a previous review.md, and says so.
• Every run renders one block per plan phase, re-derived from the plan.
• The phase projection decides scope. A phase whose projected files are untouched remains present, marked out of scope in its heading. Every criterion stays unchecked, with no evidence and no status tag. Nothing is copied from an earlier review.
• A phase in scope uses only current-diff evidence: [x] requires a file:line citation; [ ] fix means the current diff does not meet the criterion; [ ] not-applicable means the criterion cannot be assessed in this review by nature.
• Remove fixed. It represents historical state and contradicts a snapshot.
• not-applicable applies only to a criterion of a phase in scope. It never marks an entire phase out of scope and adds no finding or blocker.
• Verified is checked / (all criteria in scope - not-applicable), rounded to the nearest whole percent, and must match the rendered boxes.
• When no criterion remains after exclusions, render — (0/0), never 100%. This adds no finding or blocker by itself; the overall verdict still follows the shared rubric.
• Update the Test section of 02-review-functional.md and add regression coverage for snapshot isolation, phase scope, status tags, and the verified calculation.

Acceptance criteria

• Reviewing phase 4 of a 7-phase plan renders 7 phase blocks, 6 of them marked out of scope in their headings.
• An out-of-scope block holds only unchecked criteria, with no status tag, gap, evidence, [x], or file:line.
• Every [x] in an in-scope phase cites evidence from the current diff.
• Re-running the review on the same diff gives the same Phases and Verification, whatever the content of the previous review.md.
• With 6 criteria in scope, 5 checked and 1 not-applicable, the report shows 100% (5/5).
• With 34 countable criteria and 16 checked, the report shows 47% (16/34).
• With no countable criterion, the report shows — (0/0) and does not create a finding or blocker solely for that result.
• The fixed status is absent from 02-review-functional.md, review-template.md, and review-rubric.md.
• SKILL.md, 02-review-functional.md, review-template.md, and review-rubric.md agree on snapshot inputs, phase scope, statuses, and counting.
• Regression coverage proves an old contradictory review.md cannot affect the generated Phases or Verification sections.

Prior art in this repository

• SKILL.md:30: the snapshot rule, kept.
• 02-review-functional.md:17: "the percent and count of checked criteria", to make precise.
• review-rubric.md:19: the only definition of not-applicable, to extend.

Out of scope

• How "the diff against the repository default branch" resolves (5 occurrences in 05-review, never defined). A clone whose origin/HEAD points at a feature branch diffs against it. Separate issue.
• Enforcing review-validator.yml: a section outside the list survived in the three reviews above. Separate issue.

Reported on Discord: https://discord.com/channels/1173363373115723796/1547134531188822047