workflow-fix: test (10) FAMILY_OF parity is one-way + comment-blind — enforce bidirectional dict equality between the Step 5a and Step 10d copies

kind: infra

Goal

Harden tests/test_issue_skill_lint_family_sync.py's test (10) so it enforces exact bidirectional equality of the FAMILY_OF tables between the Step 5a copy (.claude/skills/issue/steps/09-step-5.md) and the Step 10d copy (.claude/skills/issue/steps/18-step-10d.md), closing the two divergence channels the current comparator cannot see.

The gap (surfaced, adjudicated non-blocking, still real)

Test (10) today iterates only the Step-5a span's lines and asserts each selected line also occurs in the Step-10d span. Two consequences:

1. One-way. A FAMILY_OF assignment present ONLY in the Step-10d copy passes — nothing iterates the Step-10d span looking for Step-5a peers.
2. Suffix-filtered. The selection predicate is stripped.startswith("FAMILY_OF[") AND stripped.endswith(('="workflow"', '="lint"', '="guard"', '="agents"')), so ANY assignment carrying a trailing comment is silently skipped. Meanwhile the guard-(20) parser _family_of_map (~:2194-2201) is deliberately comment-TOLERANT, so a commented assignment is real to guard (20) and invisible to test (10).

Test (14) compares SPECS/SPECS_10D path TOKENS only and carries no family VALUES; test (1) pins the Step-5a SPECS literal only. Neither can close the gap.

Measured current state (2026-08-21, branch issue-2260 at 89c1f003b6): both copies carry 49 FAMILY_OF assignments and the parsed maps are dict-equal in both directions (agents 31, workflow 7, lint 7, guard 4); SPECS and SPECS_10D are both 53 tokens and set-equal. So this is a GUARD-COVERAGE gap, not a live divergence — nothing is broken on trunk today.

Why it was not fixed in #2260

#2260's registered change set touched test (10) with exactly one edit: extending the suffix tuple by the single '="agents"' token. The one-way + comment-blind comparator shape is pre-existing origin/main behavior that approved plan v3 explicitly recorded as a caveat and mitigated with an all-BARE-entries convention (every one of the 31 new agents entries is bare, verified). The epm:review-reconcile v1 verdict (2026-08-21T19:30:00Z) ruled the demanded comparator redesign to be hardening beyond the registered change set, and downgraded the Codex blocker family-parity-check-lossy to a deferred concern — with the standing recommendation that it be filed as its own follow-up. This task is that filing.

Note also why the headline impact claim fell: guard (20)'s collector _agents_uncovered_readers refuses to count a bare SPECS token as a disposition, so the specific "Step 5a singleton / Step 10d agents member with every guard green" state is unreachable for agents-family readers. The surviving channel is narrower — comment-then-diverge, and non-agents families.

Fix direction (implementer to confirm the right shape)

1. Parse BOTH spans with _family_of_map (the comment-tolerant parser already in the file) and assert exact dict equality in both directions — not a subset/membership relation.
2. Add two negative controls that mutate copied spans and assert the parity helper reports each divergence:
   • a Step-10d-only assignment;
   • a Step-5a assignment carrying a trailing comment whose Step-10d peer is absent.
3. If the all-bare-entry convention is to remain a contract, pin it separately: assert every agents-family assignment matches a full-line bare-assignment regex (do NOT fold this into the equality check — a bare-shape violation and a divergence are different failures and should report differently).

Keep _family_of_map's comment tolerance — it is the correct semantic parser; the defect is that test (10) does not use it.

Acceptance criteria

1. A Step-10d-only FAMILY_OF assignment fails test (10).
2. A commented Step-5a assignment with no Step-10d peer fails test (10).
3. Test (10)'s docstring matches what it proves (today it claims the copies carry the "SAME FAMILY_OF entries" while proving only a filtered Step-5a subset relation).
4. uv run pytest tests/test_issue_skill_lint_family_sync.py -q stays green on the unmutated tree.
5. No change to the Step 5a / Step 10d sync fail-safe direction (dirty ⇒ whole-family skip).

Boundary with sibling tasks

Scoped to test (10)'s COMPARATOR only. #2260 owns the agents-axis family coupling (landed). #2420 owns tests/test_ensemble_review_cap.py's workflow-prose axis. #2374 owns its own per-surface axis. Do not widen into those.

Provenance

Surfaced by the reconciler in /issue 2260 round 1 as a standing non-blocking recommendation on a PASS verdict (marker epm:review-reconcile v1, 2026-08-21T19:30:00Z), grounded on the codex-code-reviewer Major at tests/test_issue_skill_lint_family_sync.py:618. Deferred concern id: family-parity-check-lossy. Fingerprint: (family-of-parity-comparator-lossy, tests/test_issue_skill_lint_family_sync.py).