Wire Design Contract Checker lenses into a structural registry

[joshft]

Summary

Closes the DA-001 / AP-036 gap: the eight documented /creview-spec Design Contract Checker lenses (from postmortems PMB-013 to PMB-020) existed only as CLAUDE.md prose and never reached the enforcing agent — a "corrective-action described but not implemented" instance in the contract mechanism itself. This makes the lens set a first-class, structurally-enforced artifact.

• New agents/design-contract-lenses.tsv — the single source of truth (SSOT) for the 8 PMB-derived lenses (DCL-001 cardinality through DCL-008 mechanism-capability-mismatch).
• agents/review-spec-design-contract.md expanded with a ## PMB-derived lenses section — one DCL-tagged, condition-bearing bullet per registry row (keyword + {BLOCKING,flag} directive + {when,if} condition + body floor), plus a template/runbook intro. The generic lens is retained.
• tests/test-design-contract-lens-sync.sh binds registry to agent by set-equality (completeness + no-orphans), validates registry well-formedness (15-case malformed-fixture rejection suite), enforces the anti-gaming body floor over the live registry (not a hard-coded seed), and self-scans to prevent re-coupling to CLAUDE.md.
• skills/cpostmortem/SKILL.md Step 3 convention: a PMB documenting a new Design Contract Checker lens must add a registry row.
• ABS-050 architecture entry — a new primary-SSOT abstraction (opposite authority direction to the ABS-047 sole-writer family: here the registry is the source of truth and the agent is derived from it).

The registry is source-only (sync.sh mirrors only agents/*.md); user installs receive the lens bodies inline in the shipped agent .md.

Incidental fix (separate commit)

A pre-existing, unrelated date-drift in tests/test-cross-feature-intel.sh: a hardcoded PMB-001 fixture date (2026-04-10) crossed the intel script's 90-day recency window on ~2026-07-09, so any CI run after that date silently filtered the "audit" entry and reddened INV-016e. Fixed by generating the fixture dates relative to today (date -d 'N days ago') so they never drift again — fittingly, the exact bound-drift class the new DCL-001/DCL-007 lenses guard against.

Test plan

• tests/test-design-contract-lens-sync.sh — 73 assertions, all pass (set-equality via two shared-regex extractors, 8 anchored seed rows, live-registry substance loop, 15-case malformed-registry rejection suite, PRH-001 self-scan whitelist, mirror parity, source-only find-check, plus test-the-test cases proving the two extractor scopes differ and the rejection branches are live).
• Full test suite green (112 test files).
• sync.sh verify clean; shellcheck --severity=warning clean.

QA findings

• 2 QA rounds — independent Claude review + codex (GPT-5.5) cross-model review. Codex found (and later confirmed resolved): DCL-004/DCL-005 lens-body narrowing vs. their source PMBs, and INV-005 substance checks iterating only the hard-coded seed array. All fixed.
• 3 mini-audit lenses (seed-completeness-vs-postmortems, lens-body-anti-gaming, test-the-test-rejection-coverage) — converged, no CRITICAL/HIGH; independently re-derived the 8-lens set from postmortems and confirmed 1:1 registry mapping.
• 1 verification-phase HIGH (CV-001) from codex — the PRH-001 self-scan whitelist matched grep + the single-quoted needle without pinning the trailing file argument, so a prohibited grep '<needle>' CLAUDE.md (reading CLAUDE.md) would have passed. Tightened to strip the needle and reject any residual guidance-file token, with 3 new test-the-test assertions.
• Surfaced (optional follow-up): MA-001 LOW (the exactly-once uniqueness leg has no dedicated duplicate-token negative fixture — exercised indirectly).

Verification status

PASS — all 13 rules (INV-001..010, PRH-001/002, BND-001) satisfied with real-file evidence and live test enforcement; codex cross-model verification confirmed convergence with no BLOCKING/HIGH outstanding.