refactor(standards): surgical prune, 23 standards to 19
What and why
A value-judgment pass on the standards, the third part of the "we used this greenfield, here is the simplification" cleanup. The test for each standard: does enforcing it catch a real conformity/quality issue, or is it ceremony? Driven by the finding that the registry listed 23 standards but only 17 had an oracle, and several mis-fired.
Stacked on !19 (merged) (which is stacked on !18 (merged)). Targets
andunn-reference-just; retarget up the chain as each merges.
The changes (23 → 19 standards)
| Action | Standard(s) | Rationale |
|---|---|---|
| Cut | examples-dir, patches-dir, templates-gitlab |
Registry-only, no oracle. The first two are niche + conditional; templates-gitlab was a mandatory ghost the scaffolds did not even satisfy. The reference's own .gitlab/ templates are unaffected. |
| Fold | workflow-rules → pipeline |
It is the workflow:/default: CI block, so it belongs in the CI standard. pipeline/check.sh now asserts the workflow: block firing on merge_request_event. |
| Enforce | posture-claims |
Authored check.sh (conditional on .gitlab/posture-claims.yaml): asserts the four canonical claim fields. The last retained ghost now has an oracle. |
| Fix | agents-md |
CLAUDE.md demoted from required to a recommended Claude shim (postern, the canonical exemplar, ships none). |
| Fix | design |
Principle 7 forbids only third-party CDNs (the reference's own Pages origin is first-party); the always-orange band is scoped to web chrome (SVG artifacts use ink #1d1d1f); the no-local-token-redefinition half is noted as review-time-enforced. |
| Verified, not changed | naming |
The strict single-token-slug H1 regex is intentional per its SKILL. The "over-strict" finding was wrong; left as-is. |
| Verified, not changed | release |
Already names pipeline/release-create consistently; no drift to reconcile. |
Validation
just guard / validate-registry (19 standards, declared = on disk) / validate-matrix / validate-versions all pass. just check (self) unchanged: only the pre-existing README posture FAIL, no new failures. A stamped docs project passes just check (0 failed) — no born-broken regression. The site + llms-full.txt rebuild without the cut standards.
Draft for review.