task: drop default-valued skill sidecar fields and keep case digests in one record #175

Goal and acceptance

• Problem / desired outcome:

  Two kinds of redundancy in the knowledge base make records harder to read and
  let copies drift:

  1. Default-valued sidecar fields. Of the 55 local-skill-v1 sidecars under
     src/inferrsi/data/skill_library/, 32 store "sources": [], 49 store
     "references": [] and 40 store the same placeholder
     "applicability": {"scope": "See procedure prerequisites and counter-signals"}.
     The readers already default these
     (skill_catalog._references, KnowledgeReader._skill), so the stored values
     add no information. The placeholder also feeds lexical Collect scoring with
     identical text for every skill, and is copied into every generated
     repo_packs/inferrsi/knowledge/**/skills/*.json projection.
  2. Case provenance duplicated in Markdown. admission-drain-occupancy.md
     repeats the raw trace SHA256 that its JSON record already holds, and carries a
     request-metadata SHA256 that exists only in the Markdown. Nothing checks that
     the two files agree. The other three casebook documents do not link their
     typed record at all.

• Completion criteria:

  • Sidecars omit sources/references when empty and applicability when it is
    the shared placeholder; the effective reader output is unchanged apart from
    applicability becoming {} for those skills.
  • Case source identities and digests live only in the case JSON record; each case
    document links its record and cites the record's source URLs.
  • Contract tests fail if either redundancy returns.
  • The packaged skill index is regenerated and build_skill_index.py --check
    passes; the knowledge-base contract documents both rules.

Scope and constraints

• In scope: skill sidecars and casebook cases under skill_library/, generated
  repo_packs/inferrsi/knowledge/ projections, tests/test_knowledge_contract.py,
  docs/knowledge-base.md.
• Out of scope: reader/schema code changes; catalog JSON records (they are a
  structured index and stay as-is per docs/knowledge-base.md); folding sidecars
  back into frontmatter; normalizing the legacy untyped source inventories;
  provenance_readsets: [], which the contract test reads directly.
• Relevant dependencies, shared interfaces or limits: follows the field
  ownership defined by RFC [RFC] Knowledge Base v1: Skills, Knowledge, and Internal Experience #17. Paths are protected by the knowledge-model policy
  ([Tracking] Knowledge PR admission: model and effort policy #50). Bundle content revisions of the touched skills and cases change, as they
  do for any content edit; no stored revision is rewritten.

Ownership and authorization

• Proposed by / actual requester: @david6666666 (agent-drafted with Claude Code)
• Status: Proposed
• Accountable parent (Issue-level owner, if assigned): TBD
• Slice implementer(s) / acting contributor (if assigned): Claude Code (claude-opus-5-5, effort max), with the requester
• Goal-acceptance owner: TBD (knowledge maintainers)
• Accepted decision or scoped delegation: Not yet granted

Inputs and deliverables

• Relevant sources, revisions or linked issues: RFC [RFC] Knowledge Base v1: Skills, Knowledge, and Internal Experience #17, policy [Tracking] Knowledge PR admission: model and effort policy #50, base 6b13cf0.
• Expected artifacts: one PR with the sidecar/case edits, regenerated index,
  contract tests and documentation.

Completion and closure

• Who accepts completion and may close: the goal-acceptance owner above.
• Remaining uncertainty / next decision: whether maintainers also want the
  remaining placeholder limitations text and empty provenance_readsets
  normalized; not proposed here.