Refocus ko-skill, simplify supporting skills, and curb documentation sprawl

[mosobande]

Stack: ori → #25 → #98 → #99 → #100. This layer targets #99 (simplify/pr-workflows). Review its diff against that parent; merge bottom-up. #95 is excluded. Existing commits are preserved with a merge commit; no force-push or PR merge performed.

Outcome

Refocus ko-skill on creating, improving, and verifying useful skills. Keep consequential boundaries and expert depth; remove duplicated governance and report ceremony.

Changes

• ko-skill: one ordinary authoring path; canonical identity/admission rule; conditional guidance for resource changes, expert references, scripts, knowledge corpora, maturity decisions, and explicit portfolio audits. Routine edits do not reopen retained resources or Experimental maturity.
• Preserve read-only validation, separate publication/install authority, historical-evidence reuse without recursive investigation, Experimental fallback for unproved new identities, and proportionate exact-candidate proof.
• Permit selectively loaded domain expertise inside execution skills when it improves recurring decisions. No compulsory separate catalogue owner.
• ro-wo: evidence-led premise check with preserved verdicts and uncertainty.
• handoff: coverage requirements instead of a fixed seven-section form, preserving exact state and missing-source recovery.
• se-triage: streamlined investigation and reporting, preserving bounded access, classification/action identifiers, pagination, and ambiguous-write recovery.
• AGENTS.md: admit a new document only for a recurring reader/task or enduring decision without an existing home. Put implementation rationale, comparisons, measurements, and review results in PR discussions/CI by default.
• Remove our duplicate simplification report and three superseded Akowe Java/Spring/catalogue research documents. Checked for inbound filename references and unique current knowledge; applicable runtime expertise remains in Alaga references. Removed files remain recoverable in Git history.

Lessons

Specify the outcome, completion evidence, and consequential exceptions. Trust capable execution of routine mechanics. Use one canonical rule per loading path, meaningful conditional references, and result-sized reporting. Preserve deep expertise, attribution, safety/authority, maintained regression cases, and machine-consumed identifiers. Do not replace deleted ceremony with another mandatory report or prompt harness.

Primary guidance consulted on 2026-09-05:

• Anthropic skill authoring: capable-model assumptions, progressive loading, specificity matched to fragility, and real-use testing.
• OpenAI model guidance: consistent instructions, scoped autonomy, and proportionate verification.

These motivate the approach, not a claim of measured cross-model reliability or performance.

Proof and limits

Package validator passes for 39 skills; plugin-agent validator passes; all 21 repository contract tests pass; git diff --check passes.
Five earlier fresh-context simulations covered deployment authority, incomplete handoff recovery, pagination, ambiguous comment-write recovery, and non-reproduction uncertainty.
This cleanup adds independent baseline/candidate source simulations for a routine Experimental link fix, duplicate-owner admission, embedded Spring expertise, historical packet reuse, proof placement, and validate-only defects. Review caught the missing Experimental fallback for credible but unproved new identities; it was restored before publication. No additional concrete finding in the six cases.

These are source-grounded simulated decisions in ChatGPT Work, not live provider mutations, an Anthropic run, or a latency/reliability benchmark. No maintained behavioral eval packs were removed.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}