Redesign reviewing code skill

[QuentinTorg]

Intent

Redesign reviewing-code around high-signal, proportional review judgment while preserving independent analytical passes that catch material defects.

Delivered change

• Replaces the mandatory nineteen-check audit, phase holds, tracking document, and fixed line-count gate with adaptive independent passes.
• Preserves Phase 0 as the high-level orientation and architectural pass while requiring later passes to catch integration, correctness, safety, maintainability, and verification issues it misses.
• Requires a focused behavior-and-safety minimum for every behavior-changing review.
• Separates broad internal investigation from a strict output threshold so low-impact preferences and automated nits are omitted without narrowing bug discovery.
• Adds a finding policy with severity, current-change relevance, and resolution-risk classifications.
• Keeps current-change findings separate from valuable follow-up candidates.
• Makes reviewer context and Hunk notes the normal scratch state, with temporary head-bound recovery snapshots only when needed.
• Adds a proportional medium-independent review-result template and behavioral evaluation scenarios.
• Removes GitHub command mechanics from the analytical skill; Hunk and GitHub remain adapter responsibilities.

Stack dependency

This pull request is stacked on #2 and should be reviewed as the reviewing-code-specific layer. After #2 merges, its base should be changed to main.

Verification

• Checked skill metadata length and core instruction size.
• Checked all progressive-disclosure links and shallow resource paths.
• Checked the staged diff and Markdown whitespace.
• Audited the package for stale hard-hold, fixed-size, tracking-document, and GitHub CLI instructions.
• Behavioral evaluation scenarios are specified but not automated or executed.

Review guidance

Focus on whether the design preserves both review recall and precision: broad phase-independent investigation should still find subtle bugs, while only material evidence-backed findings should reach the author. Also examine the correctness minimum, finding classifications, rereview contract, Hunk recovery policy, and adapter boundary.