#2166 workflow-fix: 5 stale CLAUDE.md prose pins repointed at their rule-file homes — main green (347 passed)

Overview / Motivation

Five test functions across four files pin CLAUDE.md prose that the 2026-08 compaction deliberately relocated into .claude/rules/*. They are red on origin/main right now (reproduced 2026-08-07 at 68fbf9bf3e, repo-root checkout, 5 failed / 10 passed in 1.23s), so every intervening session's Step 9c gate must re-classify them as known-red.

Same class as #2157 → #2159/#2160, but different target files — that pair covered tests/test_issue_skill_trigger_dense_tag_adoption.py only. Dedup key is (target_file, fingerprint), and none of the four files below was touched by that fix.

Surfaced by the #2164 implementer during its gate-scope run, then verified independently by the #2164 orchestrator against current origin/main before filing (the failures are real and are NOT caused by #2164's diff).

Goal

Make the five pins green on origin/main by re-pointing each at the rule file that now carries the relocated prose verbatim, preserving every ordering and window assertion. Do not restore prose to CLAUDE.md — the compaction moved it deliberately to cut always-on context load.

Failing tests (verified red on main 68fbf9bf3e)

Test	Relocated prose now lives in
tests/test_router.py::test_claude_md_compute_backends_section_matches_656_contract	.claude/rules/compute-backends.md
tests/test_issue_skill_file_only_verdict_post.py::test_claude_md_points_at_file_only_path	.claude/rules/codex-ensemble-review.md
tests/test_issue_skill_neutral_gate_vocab_brief.py::test_claude_md_rung_e_neutral_gate_vocab	.claude/rules/context-hygiene.md / trigger-dense-review.md
tests/test_issue_skill_neutral_gate_vocab_brief.py::test_claude_md_rung_e_steering_vocab	.claude/rules/context-hygiene.md / trigger-dense-review.md
tests/test_suffixed_pod_completion_teardown_pin.py::test_claude_md_carries_completion_side_teardown_contract	.claude/rules/pods.md (the count >= 2 two-site assertion — CLAUDE.md now has 1 of the 2 Completion-side teardown sites; the other moved)

Workflow gap

The compaction train landed without re-running the workflow-invariant pin family against the final merged tree, so stale pin targets shipped. #2159 fixed one file; these four were not swept. A post-compaction sweep of the whole pin family would have caught all of them at once.

Acceptance criteria

• All five named tests pass on a pristine origin/main checkout.
• Each re-pointed assertion reads the rule file that verifiably contains the prose verbatim — confirm the string is present there before re-pointing; if a string is genuinely absent everywhere, that is a compaction content loss and must be reported, not papered over by weakening the assertion.
• Ordering / window / count assertions preserved. The teardown pin's count >= 2 needs a decision: either both sites exist across CLAUDE.md + the rule file (assert across both) or the contract genuinely has one site now (adjust the count with a stated reason).
• No prose is moved back into CLAUDE.md.

Notes

Also worth checking in the same sweep, since the compaction is the common cause: whether any other tests/test_*claude_md* / prose-pin test is red on main. uv run python scripts/step9c_baseline.py status reported failing_tests: 1 after a fresh refresh while these 5 were red — the ledger's universe appears not to cover this pin family, which is itself worth a look (a gate that cannot see a main-red test cannot classify it).