[sunwrobert]

Summary

The skill used to review gh pr diff output. Large PRs got cut to a sample, and the saved review depth still said full. In one run, 182 of 266 reviews came from incomplete patches, but the state labeled them full-net-diff.

• Adds skills/pr-merge-tracker/scripts/collect_diffs.py. It diffs the merge base against the head in a blob-filtered bare git cache, so there are no API file or line caps. It writes the full patch, a file inventory, and 80,000-character chunks, with hunks split and never cut off.
• Sorts files into required (code, config, tests, docs, export changes), mechanical (import-path-only) and generated chunks.
• With a saved reviewed head, it reviews only files whose own +/- lines changed, so changes merged in from main drop out.
• depth computes full or sampled N% from the chunk IDs a reviewer actually read. SKILL.md now forbids writing the depth by hand and requires sampled N% in report bullets.

Testing

Ran against OpenRouterTeam/openrouter-web:

• #55990 (4 files): 1 chunk, full.
• #56021 (893 files): 3 required chunks plus 661 import-only files; reading 1 of 3 chunks reports sampled 36%.
• #56069 (17,376 files): reading 2 chunks reports sampled 1%.
• Delta mode on #56021 cut the review from 893 files to 54.

bun run check reports a format issue in executor/executor.jsonc, which was already on main; the changed files pass oxfmt --check.