Refine SkillSource sub-rule metadata via trigger pointers

Context

PR #11333 completes the immediate #11323 metadata/test gap by emitting isAtlasMonolithSubRule from SkillSource and asserting the current behavior in the focused source-unit test.

Claude's Cycle 1 review on PR #11333 approved the PR, but surfaced a valid semantic-precision follow-up: the current v1 implementation uses pathParts.includes('references') as a coarse approximation. That distinguishes top-level SKILL.md routers from payload files, but it does not distinguish workflow-map files under references/ from extracted sub-rule sibling files.

This ticket captures that precision refinement as a post-#11320 child of Epic #11317, instead of churning an already-approved PR whose current metadata field has no downstream consumer yet.

The Problem

The field name isAtlasMonolithSubRule implies a specific semantic category: a sub-rule extracted from an atlas/workflow monolith. The current v1 heuristic is broader:

• pr-review/SKILL.md -> router map -> false today, correct.
• pr-review/references/pr-review-guide.md -> workflow map under references/ -> true today, too broad for the eventual semantic meaning.
• pr-review/references/audits/loading-runtime-effect.md -> extracted sub-rule sibling -> true today, correct.

The broad behavior is acceptable for #11323 because it is additive, currently unused by downstream routing logic, and still useful as a router-vs-payload discriminator. It should not become the permanent contract once trigger-pointer parsing exists.

The Architectural Reality

Current source evidence:

• ai/services/knowledge-base/source/SkillSource.mjs currently derives isAtlasMonolithSubRule from path membership under references/.
• #11320 owns the trigger-aware workflow lint primitive, including AC4 section-trigger parsing for the canonical <!-- trigger: ... -> read ... --> pointer shape.
• Discussion #11316 / Epic #11317 owns KB semantic recall for skill chunks.
• Discussion #11314 / Epic #11319 owns trigger-aware workflow structure and Map vs World Atlas recursive-application discipline.

Therefore the precise discriminator belongs after #11320 AC4 lands: SkillSource should identify files that are targets of trigger-pointer declarations from workflow maps, not every file under references/.

The Fix

Refine SkillSource metadata extraction to use trigger-pointer knowledge once the #11320 parser contract is available.

Expected shape:

1. Reuse or share the section-trigger parser from ai/scripts/lint-skill-manifest.mjs rather than inventing a second incompatible parser.
2. Build a per-skill set of declared trigger-pointer target paths.
3. Emit isAtlasMonolithSubRule: true only for chunks whose source file is a declared target path.
4. Keep a conservative fallback for skills without trigger-pointer metadata so existing KB ingestion remains backwards-compatible.
5. Update SkillSource.spec.mjs with fixtures covering all three categories: top-level router, workflow map under references/, and extracted sub-rule sibling.

Contract Ledger Matrix

Acceptance Criteria

• AC1: Verify #11320 AC4 has landed or expose a shared parser helper as part of this work if it has not.
• AC2: SkillSource no longer treats every references/** file as an extracted sub-rule when trigger-pointer data is available.
• AC3: SkillSource identifies trigger-pointer target files and emits isAtlasMonolithSubRule: true for those chunks.
• AC4: Workflow-map files under references/ that contain pointers but are not themselves target files emit isAtlasMonolithSubRule: false.
• AC5: Top-level SKILL.md router files continue to emit false.
• AC6: Backwards compatibility is preserved for skills without trigger-pointer declarations.
• AC7: Focused Playwright unit coverage extends test/playwright/unit/ai/services/knowledge-base/source/SkillSource.spec.mjs with router / workflow-map / sub-rule sibling fixtures.
• AC8: PR body cites #11320 dependency state and does not close Epic #11317.

Out of Scope

• Implementing the #11320 trigger-pointer parser itself, unless it still does not exist when this ticket is picked up and the implementation explicitly chooses to share a parser helper.
• Changing the public type: skill MCP query schema from #11326.
• Changing the broader SkillGraph / Memory Core topology.
• Renaming isAtlasMonolithSubRule before downstream consumers exist; this ticket refines semantics, not the field name.

Avoided Traps / Gold Standards Rejected

• Fixing inside PR #11333: rejected because PR #11333 is approved, CI green, and the issue is an Approve-grade follow-up with no downstream consumer yet.
• Permanent path-only classification: rejected because references/ contains both maps and extracted sub-rules.
• Second parser implementation: rejected because duplicate syntax parsing would drift from #11320's lint contract.
• Reactive KB watcher: remains out of scope per Discussion #11316 Option D rejection.

Related

• Parent Epic: #11317
• Current v1 metadata/test PR: #11333
• Source implementation: #11321
• Source-unit coverage: #11323
• Trigger-aware workflow primitive: #11320
• Trigger-aware workflow epic: #11319
• Discussions: #11316, #11314

Origin Session ID: d6d89930-f408-42a0-b60e-ec4487a8cc46 Retrieval Hint: "SkillSource isAtlasMonolithSubRule trigger-pointer precision PR #11333 follow-up"