workflow-fix: cited-body gate labels Word:-prefixed content corrections 'frontmatter-only' (fm_line over-matches ordinary prose)

Goal

Stop check_cited_body_currency.py from labelling a genuine content correction frontmatter-only. The frontmatter heuristic matches any Word:-shaped changed line, so a real correction written as RESULT: ... -> CORRECTED: ... is classified as a metadata-only edit.

The defect

In scripts/check_cited_body_currency.py:

fm_line = ^[+-](?:---\s*$|[A-Za-z_][A-Za-z0-9_-]*:)

Any changed line beginning with a word followed by a colon satisfies it. That shape is common in ordinary body prose — RESULT:, CORRECTED:, Note:, Takeaway: — so a conclusive-window diff whose only content changes carry such prefixes earns the advisory frontmatter-only label.

Verified by execution during #2384's round-5 review: a -RESULT:.../+CORRECTED:... content diff under conclusive=True returned frontmatter-only, while genuine prose returned None.

Why it matters, and why it is not urgent

It matters because this label feeds an operator decision. .claude/skills/adversarial-planner/SKILL.md tells the operator on exit 3 to re-read the changed sections and then either record that the plan text is unaffected, or bounce. A frontmatter-only label nudges toward "unaffected". The round-2 reconciler's whole argument for why a mislabel was blocking rested on exactly that: the label drives the disposition.

And the shape is plausibly the real-world one. #2384's founding incident (#2378) was a plan quoting #825's superseded linear NULL after the parent's body was corrected to the opposite sign — and a correction commit that rewrites a RESULT:-prefixed line to a CORRECTED:-prefixed one is a natural way for that edit to look.

It is NOT urgent because the harm is bounded in a way the rename-only case was not: unlike rename-only, the frontmatter-only label never suppresses the diff. The operator still sees every changed line, so the evidence is present and only the summary label is wrong. The verdict is STALE regardless.

Provenance

Introduced in #2384's round-1 commit 92e537779ef and unchanged since, surviving four review rounds and two binding reconcilers. #2384's round-5 diff strictly NARROWS label emission and does not touch this regex, so this is not a round-5 regression. Both the round-5 reviewer and the round-4 reconciler recorded it as out of ruled scope with the fix deferred here, rather than expanding a tightly-scoped round.

Fix sketch (from the round-5 reviewer)

Anchor the heuristic to known frontmatter keys, or to the leading --- delimited region, instead of matching any Word: shape.

Pin: build a CONCLUSIVE window whose only content change is a Word:-shaped body line and assert label is None (or a content label), demonstrated failing before the change and passing after. A negative control should confirm a real frontmatter-only edit still earns the label — the point is to stop over-matching, not to retire the label.

Acceptance

1. A conclusive-window content correction whose changed lines carry Word: prefixes is no longer labelled frontmatter-only, proven by a committed fixture that fails pre-fix.
2. A genuine frontmatter-only edit still earns the label (do not fix by deletion).
3. The conclusive-gating behaviour that #2384 round 5 landed is preserved — an inconclusive window still yields label is None on every path.