feat(codebase-readiness, agent-ready): score and install regression-aware quality gates

[dgalarza]

Summary

Extends the two existing skills so that regression-aware quality gates are scored by codebase-readiness and installed by agent-ready. No new plugin. (The earlier quality-ratchet plugin on this branch has been removed; nothing from it was kept.)

A regression-aware gate: the project's native complexity / duplication / dead-code tools, a baseline that lets legacy debt stay, PR CI that blocks new or worsened debt relative to the merge-base, the same commands locally and in CI, a human-reviewed baseline, and tests of the gate itself.

codebase-readiness → 1.8.0

• references/quality-gates.md (new, shared): defines the gate, a single Gate Maturity Level (L0 none → L1 report-only → L2 threshold → L3 regression-aware blocking → L4 governed) determined once in Phase 1 and recorded in the snapshot, native tooling per language, anti-patterns, and a credit-ownership table so the four dimensions never double-count:
  • Code Clarity — which structural properties are covered (complexity, duplication, reliable dead-code), more credit when each is blocked at L3+
  • Consistency — lint-debt treatment (pruned/shrinking todo or baseline vs growing baseline / inline suppressions)
  • Feedback Loops — actionability, reproducibility, gate tests (PR annotations / changed-files scoping, local == CI command, self-test)
  • Change Safety — blocking semantics and baseline governance (L3 blocking, reasons + review + pruning + CODEOWNERS)
• Top-band credit in every dimension requires CI that blocks new/worsened debt; report-only tooling is explicitly capped.
• recon.sh gains a === QUALITY GATES === section (tooling, baseline files with grow/shrink history, soft-fail and merge-base markers in CI, local commands, gate tests, CODEOWNERS).
• Four dimension guides updated (examine bullets, evidence commands, band notes, modifiers, output lines); SKILL.md wires the shared reference into the three relevant agent prompts and adds the snapshot field; report template gains a Gate Maturity section; README dimension table updated.

agent-ready → 1.5.0

• New quality-gates mode (SKILL.md), integrated into scaffold (Step 8) and audit (Quality Gate section: governed / ungoverned / report-only / absent, plus auto-fix offers), with the check command and three directives added to the AGENTS.md template's Definition of Done / conventions.
• references/quality-gates-pattern.md: the cross-language contract (report / check / baseline), two routes — Option A native baseline/diff modes (golangci-lint new-from-merge-base, detekt --baseline, PHPStan baseline + reportUnmatchedIgnoredErrors, RuboCop todo regenerated) and Option B a stdlib engine — with adapter recipes for TS/JS, Ruby, Python, Go, JVM, Rust, PHP, the baseline review protocol, CI shape, and debt-fixture snippets.
• assets/quality-gate.py: fingerprint baseline engine (identity ignores line numbers and numeric values; first integer in a message is tracked as a metric so a higher complexity value counts as worsened); merge-base-aware --changed-only and changed-file attribution; GitHub annotations; tool failure ≠ pass.
• Baseline governance: baseline requires --reason, prints a review summary (by rule / top files), supports --dry-run, is refused under CI, and writes the baseline unreviewed — check fails until a human runs --approve --reviewed-by NAME. Stale entries fail check until baseline --prune removes them (prune only tightens, so agents may run it). CODEOWNERS on the baseline file is required by the mode.
• assets/quality-gate-ci-template.yml (fetch-depth 0, base fetch, no continue-on-error, runs the self-test), assets/quality-gates-guide-template.md (docs/guides/quality-gates.md), assets/quality-gate-test-template.sh (proves clean tree passes / synthetic debt fails / stale entry pruned, on a temp copy of the baseline).
• tests/test_quality_gate.py: 18 unit tests for the engine — legacy passes after review, line moves don't create new findings, new / worsened-metric / worsened-count fail, stale → prune → pass, prune tightens metrics and the ratchet holds, --reason required, refused in CI, unreviewed fails, extend resets review, dry-run writes nothing, --changed-only scoping via merge-base, tool failure is an error, jsonl format. Wired into CI as a test-templates job.

Validation

• quick_validate.py passes for both skills; all plugin.json + marketplace.json parse; both workflow YAMLs parse
• python3 -W error -m unittest discover -s plugins/agent-ready/tests → 18/18
• recon.sh syntax-checked and run (new section produces output on this repo)
• Test template exercised end-to-end against the engine in a scratch git repo with a fake analyzer → 5/5 assertions
• Versions bumped in both plugin.json files and marketplace.json

Not done

• Do not merge; no release tag cut