rule(verify-reviewer-findings): extend blocked-green-ci rule with verify-before-fix discipline

.claude/rules/blocked-green-ci-investigate-threads.md

@@ -23,6 +23,68 @@ When `bun tools/github/poll-pr-gate.ts <PR>` reports
 than once without having run the threads query. If this happens,
 stop and run it.
 
+## Composes with: verify-before-fix on reviewer findings
+
+Once threads are surfaced, **verify each finding against the
+source before applying a fix**. Reviewer tools (Copilot, etc.)
+produce real findings AND hallucinations; treating both as
+equally actionable wastes commits and can introduce regressions.
+
+Verification anchors:
+
+- **Direct line-level inspection**: substitute the literal line number for
+  `<N>` via `-v` — `awk -v N=22 'NR==N { print NR": ["$0"]" }' <file>` —
+  to read the exact bytes the reviewer is commenting on. (Without
+  `-v N=…`, awk treats `N` as an uninitialized variable equal to 0 and
+  prints nothing.)
+- **`gh api repos/<owner>/<repo>/pulls/<N>`** to fetch PR metadata
+  (state, merge_commit_sha, head SHA); **`gh pr view <N> --json
+  commits,mergeCommit`** to inspect commits; plus
+  `git log --grep '#<N>'` to find the merge commit by PR-number in the
+  local repo: confirm historical / cross-reference claims, not memory.
+  (Plain `git log <N>` does NOT work — git log expects
+  refs/commits/paths, not PR numbers.)
+- **Local lint/build**: re-run the relevant check locally to see
+  if the reviewer's failure-mode reproduces.
+
+### Suspect-by-default Copilot finding classes
+
+Empirical patterns observed multiple times during a single
+autonomous session (2026-05-16) where Copilot reported a
+finding that direct inspection confirmed as false:
+
+| Class | Reviewer claim | Reality |
+|-------|----------------|---------|
+| Table double-pipe (`\|\|`) | "extra leading `\|` in tables creates empty first column" | Direct `awk` shows single `\|` rows; 4 confirmed FPs in one session (PR #3685, #3690, #3699-era, #3709) |
+
+When a thread matches a known-FP class:
+
+1. Verify via direct inspection (don't trust the reviewer's quote)
+2. If confirmed FP, resolve no-op with a brief comment OR just resolve
+3. Do NOT push a fix that would change correct content
+
+This list is empirical; new FP classes get appended as they
+accumulate evidence. A single FP doesn't justify entry —
+two-or-more across distinct PRs is the threshold for
+"suspect-by-default."
+
+### Verify-also-on-stale-but-fresh-looking findings
+
+A second class: findings that were TRUE at thread-filing time
+but became STALE by review-resolution time. Common cases:
+
+- Parent-tick links to shard files in sibling PRs — the link
+  is broken at file-tree time but resolves once the sibling
+  PR merges. The thread is filed correctly but resolution-time
+  state has self-healed.
+- "Inconsistency between X status and Y status" prose
+  observations — true at write-time, accurate then,
+  but the underlying state has moved.
+
+Resolve these no-op (the prose was correct at write-time;
+substrate edits would be retroactive rewriting). Stale ≠
+false; it just means the action window closed.
+
 ## Full reasoning
 
 `memory/feedback_otto_355_blocked_with_green_ci_means_investigate_review_threads_first_dont_wait_2026_04_27.md`