Skill reference lint skips §1d and backticked targets

Context

PR #15997 corrected three live references from ticket-create-workflow.md §1c to §1d. The source target is unambiguous: §1c is the Agent OS Structure Map Gate, while §1d is the Ungraduated-Discussion Cross-Check. During exact-head review, removing inline-code backticks from the filename was expected to make those corrected references mechanically validated.

A direct positive/negative probe falsified that expectation. checkSkillReferenceIntegrity() returned [] for all four of these forms:

ticket-create-workflow.md §1d
ticket-create-workflow.md §1x
`ticket-create-workflow.md` §1d
`ticket-create-workflow.md` §1x

The valid and deliberately invalid digit-leading anchors are therefore indistinguishable to the lint.

The Problem

The skill-reference lint has two interacting silent-green paths:

1. SECTION_REF_SOURCE accepts dotted numeric identifiers such as §5.1 or letter-first named identifiers such as §template-block, but not digit-leading alphanumeric identifiers such as §1d.
2. Inline-code-wrapped filename targets are not associated with following prose refs. A backticked numeric ref can then fall back to the containing file and emit a misleading dangling-reference error, while an unqualified nonnumeric ref is skipped.

Heading extraction compounds the first gap: ### 1d. The Ungraduated-Discussion Cross-Check is indexed as numeric anchor 1, not 1d. Removing backticks alone therefore does not restore validation.

This is a false proof surface. A green Skill Manifest Lint currently cannot establish that a changed target.md §1d reference resolves, even when the target file exists.

The Architectural Reality

• ai/scripts/lint/lint-skill-manifest.mjs:300-308 owns the accepted section-reference and target grammar.
• ai/scripts/lint/lint-skill-manifest.mjs:314-329 owns heading-anchor extraction.
• ai/scripts/lint/lint-skill-manifest.mjs:644-662 owns target resolution and dangling-reference reporting.
• ai/scripts/lint/lint-skill-manifest.mjs:699-725 owns changed-line prose-reference validation and currently skips unqualified nonnumeric matches.
• test/playwright/unit/ai/scripts/lint/lintSkillManifest.spec.mjs:619-741 covers dotted numeric and letter-first named references, but has no inline-code-target or digit-leading-alphanumeric controls.
• #12493 introduced numeric reference-integrity coverage; #12582 extended it to named refs. Neither ticket covered §1d-shaped headings or inline-code target equivalence.
• #11599 is not the owner: it proposes a broad semantic-heading migration and explicitly leaves historical positional refs out of scope.

Structure-map result: the existing owner is ai/scripts/lint/lint-skill-manifest.mjs, with the existing focused sibling spec under test/playwright/unit/ai/scripts/lint/. No new module or directory is needed.

The Fix

Repair the existing parser and focused spec in one PR:

1. Give prose refs and heading extraction one shared accepted section-ID grammar that covers the live digit-leading alphanumeric form (§1d) without regressing dotted numeric (§5.1) or letter-first named (§template-block) forms.
2. Normalize inline-code filename targets before target association so `target.md` §N and target.md §N have the same validation semantics.
3. Never silently skip a target-qualified section token. A target that resolves but whose section token is unsupported or missing must produce an actionable error.
4. Preserve --base changed-line ownership so newly recognized historical refs do not block unrelated PRs.
5. Add positive and negative controls for each supported target/anchor family.

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0008 — this repairs the existing skill-manifest enforcement layer and does not change Progressive Disclosure or skill-reference policy.

Acceptance Criteria

• A bare ticket-create-workflow.md §1d reference resolves to the live ### 1d. heading.
• A bare ticket-create-workflow.md §1x reference fails with an actionable dangling/unsupported section error.
• The equivalent backticked target forms produce the same pass/fail results as the bare forms.
• A backticked numeric target is validated against its named target file, never the containing source file.
• Existing dotted-numeric and letter-first-named reference controls remain green.
• Heading extraction and reference parsing share compatible grammar; a heading accepted as a target can be referenced, and a target-qualified token cannot be silently skipped.
• --base changed-line ownership remains intact.
• Focused tests cover valid and invalid bare/backticked numeric, named, and digit-leading alphanumeric references.
• The PR records a focused unit invocation and node ai/scripts/lint/lint-skill-manifest.mjs --base origin/dev; do not use the default unscoped npx playwright test.

Out of Scope

• Repo-wide heading migration or reopening #11599.
• Changing the divergence-window rule delivered by #15997.
• Rewriting every historical skill reference in the same PR.
• General-purpose Markdown link validation outside .agents/skills/**.

Avoided Traps

• Removing backticks only: falsified; §1d remains outside the accepted grammar.
• Treating green lint as evidence that the ref was checked: the direct negative control proves the current green is vacuous.
• Silently banning a live heading form: §1c/§1d are current workflow anchors. If the long-term heading convention migrates them, that belongs to #11599; the current enforcement must still be truthful.
• Absorbing the repair into #15997: that PR has already exhausted two ordinary Request Changes cycles and its graduated rule is independent of this parser defect.

Related

Follow-up from #15996 / PR #15997. Extends #12493 and #12582. Related but not owned by #11599.

Retrieval Hint: lint-skill-manifest backticked target digit-leading section ref 1d silently skipped

Live latest-open sweep and A2A in-flight claim sweep: recorded immediately before creation.