workflow_lint: flag stale gotchas.md pointers after an entry relocates (live instance in research-project-structure.md)

Mechanize the stale-relocation-pointer class: a lint flagging text that names gotchas.md alongside a #<N> citation where #<N> no longer occurs in .claude/rules/gotchas.md.

Why now

#2189 relocated 27 entries verbatim out of .claude/rules/gotchas.md into six topic-owning rule files (a 24.4 KB reduction), and its ## Out of scope section names generalized residual-relocation as a successor lever — so more entries will move. Every relocation risks leaving behind a pointer that still sends the reader to gotchas.md for a rule that now lives elsewhere.

Two live instances were found by hand during #2189's round-1 code review:

1. .claude/rules/upload-policy.md:247 — pointed at "gotchas.md, #923" for the conditional-.env entry relocated to .claude/rules/pod-side-reporting.md. FIXED in #2189 commit 2ddec80aad (it was in the round's own diff scope).
2. .claude/rules/research-project-structure.md:73 — the same class (`.claude/rules/gotchas.md`; incident #923), left UNFIXED because it was outside #2189's criterion-8 diff scope. This is a live stale pointer on main right now and is the natural first regression fixture.

Both were caught by a reviewer reading prose. Nothing mechanical detects them.

A distinct, related instance worth handling in the same pass

.claude/rules/LESSONS.md's gotchas.md row — the ALWAYS-ON index — still advertises the topic "chained smoke-then-full leg out-root residue" among gotchas.md's contents, though #2189 relocated that rule to .claude/rules/crash-fix-rounds.md. #2189 could not fix it: editing LESSONS.md was a hard prohibition in that task (43 B of headroom against _LESSONS_MAX_BYTES = 9600, and further trimming of that row was ruled out at #1269).

This is the same failure mode one level up — the index misdirects rather than a sibling rule file — and it argues the check should cover LESSONS.md trigger rows too, not just #N co-references. Note the LESSONS.md form has no #N to key on, so it needs a different detector (topic-phrase presence), and fixing any LESSONS.md row still runs into that file's byte budget. Consider whether this half belongs here or with the LESSONS.md budget decision.

Proposed shape (for the implementer to settle)

• Primary detector: for each #<N> that co-occurs with a gotchas.md mention in a tracked .claude/**/*.md file, FAIL (or WARN) when #<N> does not appear in .claude/rules/gotchas.md.
• Expect false positives: a pointer may legitimately cite gotchas.md for CONTEXT while the #<N> belongs to a different claim in the same sentence. Calibrate against the live tree before choosing FAIL vs WARN — a WARN that actually fires beats a FAIL that gets suppressed.
• Baseline is NOT clean: research-project-structure.md:73 must either be fixed in the same change or explicitly grandfathered.

Provenance

Filed by the #2189 orchestrator from the round-1 code-reviewer's prose follow-up (lower-confidence, uncapped per .claude/rules/workflow-fix-on-bug.md). Sibling of #2192 (conflict-marker residue lint), filed from the same review.