docs-hygiene(rename-references): pattern library misses command-argument and document-title forms (6 stale refs survived 3 sweeps in #1276)

[kyle-sexton]

Observation

Renaming the re-anchor plugin to discipline (#1276) left six stale references that three sweep passes missed. All six are the same two syntactic shapes, and context/patterns.md's twelve forms cover neither:

File	Line	Shape
plugins/discipline/README.md	1	# re-anchor — document title
plugins/discipline/README.md	340	/plugin install re-anchor@melodic-software — command argument
plugins/discipline/README.md	363	/plugin configure re-anchor — command argument
plugins/discipline/skills/setup/SKILL.md	73	/plugin configure re-anchor — command argument
plugins/discipline/skills/setup/SKILL.md	81	/plugin configure re-anchor — command argument
plugins/discipline/skills/setup/evals/evals.json	8, 12	/plugin configure re-anchor inside expected_output / expectations

Four are functional: a consumer following the README's own install block, or the setup skill's own reconfiguration step, hits plugin-not-found. The eval pair is worse than cosmetic — expected_output asserted the assistant routes to the stale command, so a correctly-behaving assistant would have failed the eval.

patterns.md is explicit that this is the moment to act:

Phase 6 — pattern library evolution

When the skill's re-sweep finds a NEW syntactic form not covered above:

1. STOP — do not silently mangle. Report the new form to user.
2. Document the pattern in this file with all 5 fields …

Why the existing twelve forms miss both

Form 1 (slash-prefixed token, \B/<old>\b) does not fire on /plugin configure re-anchor. The slash anchors plugin, not <old> — the renamed token sits in argument position, several words downstream of the only slash on the line. Form 1 is built for /<old> as a skill invocation; the plugin name as an argument to a different command is a distinct position.

Form 3 (path references) does not fire — nothing here is a path.

Form 2 (bare token, \b<old>\b) does fire, and this is where the failure is instructive rather than merely a coverage hole. re-anchor is used as an English verb 210 times in that plugin ("each skill re-anchors ONE discipline"). Form 2 hits all 210 plus the 6 real defects with no way to tell them apart:

• re-anchor is not in triage.md's English-verb blocklist (confirm, test, review, fix, clean, …), so triage rates every one of those 210 matches Certain.
• If it were added to the blocklist, all 216 land in the ambiguous bucket, and triage.md mandates per-match AskUserQuestion confirmation — 216 prompts to find 6 defects.

Neither branch is usable. The static blocklist cannot fix this, because the token's part of speech is a property of the consuming codebase, not of the token.

Direction

Add high-precision position-anchored forms, so the container-sense uses are isolated on syntax rather than sorted out of verb-sense noise by a human:

• Command-argument form — <old> as a bare argument to a tool or CLI verb, not as the slash-command itself: /plugin (install|configure|enable|disable|uninstall) <old>, <old>@<marketplace>, and the same token appearing as a key or value in an enabledPlugins / pluginConfigs style config example. Triage default Certain: the enclosing command word supplies the disambiguation that bare-token position lacks.
• Document-title form — ^#{1,2}\s+<old>\s*$, plus frontmatter name: / title:. Triage default Certain: a heading whose entire content is the renamed token names the thing, never uses it as a verb.

Both are container-position patterns. Worth checking during authoring whether a single generalization — "positions where <old> can only be a proper name" — subsumes them and anything else in that family, rather than accreting two more numbered forms.

Also note in triage.md that a token which is an English verb in the consuming codebase but absent from the static blocklist is the case the blocklist cannot serve, and that position-anchored forms are the answer rather than blocklist extension.

Acceptance

• context/patterns.md carries the new form(s) with all five documented fields (form name, regex, triage default, example match, known false-positives), consistent with Forms 1–12.
• Each new regex is validated against the real six references above — the commit that fixed them is 930c97a4, so git show 930c97a4 is the fixture.
• Each new regex is validated for precision against the same tree: it must NOT match the verb uses in plugins/discipline/README.md or context/re-anchor-audit-correct.md.
• triage.md records the blocklist limitation.
• /skill-quality:check docs-hygiene:rename-references passes.

Collision risk

Searched all issue states for rename-references, and open issues for patterns, form, sweep, rename. No existing issue covers the pattern library. #253 (docs-hygiene proactive repo-scan) is a different skill and a different detection target; #280 (shadowed-skill-renames contract) is closed and concerned naming grammar, not sweep coverage.

[kyle-sexton]

Built and in review as PR #1386.

Forms 13–15 added, validated on both axes against the real fixture rather than asserted — recall from 930c97a4's removed lines (the defect set), precision from the whole pre-fix tree:

Pattern	Hits on the pre-fix tree	Real defects among them
Form 2 (bare token)	134	8
Forms 13–15	9	8

Two things this issue did not anticipate, both surfaced by review and both real:

Position-anchoring alone is not the remedy. Adding high-precision forms only ADDS hits — every line they catch, Form 2 catches too, so the prompt flood this issue describes survives untouched. Two mechanisms were needed on top:

• Precedence — dedup by (file, line) after the sweep and before triage, so a container-position match supersedes its bare-token duplicate rather than raising a second finding.
• Container-rename mode — the bare-token residue (the 126 lines no container form touches) excluded from Certain entirely, reported as one aggregate count. Mode is decided by WHAT is renamed, which is why it works where the static blocklist cannot: it does not depend on anyone having listed the token in advance.

Final: 8 Certain, 126 reported-not-proposed, 0 confirmation prompts.

The title form over-reaches on ordinary-word container names. This issue's framing assumed a coined token like re-anchor, where "a heading that IS the token can only be naming it" holds. It fails for a common word — verified in this repo: renaming the testing plugin matches README.md:86 (### Testing, a marketplace category heading), and architecture matches plugins/miro/README.md:39 (an unrelated design section). Under precedence a false Certain there is worse than a plain Form 2 hit, because it discards the safer classification. Form 14 is now scoped to container-owned documents and always Ambiguous when the token is a common English word.

Acceptance criteria

All five met:

• context/patterns.md carries Forms 13–15 with the five documented fields each.
• Both-axis validation against 930c97a4 as the fixture, figures above.
• Precision verified against the verb uses — the forms do not match each skill re-anchors ONE discipline.
• triage.md records the blocklist limitation and routes to position instead.
• Static gates pass: claude plugin validate ., markdownlint-cli2 0 errors, changelog parity.

Five eval cases (7–11) rather than the single case anticipated — precedence, mode, and title-scoping each needed their own.