[feature] Detector self-evaluation corpus + measured FPR/FNR

[yuyu04]

What problem does this solve

cladding gates everything on ~37 drift detectors, and SECURITY.md itself elevates a detector false-negative to "security-adjacent, because it erodes the falsifiability claim." Yet there is no measured precision / recall / false-positive rate for the detectors, and no regression suite that proves they don't drift over releases toward either noise (false-positives train users to ignore findings) or blindness (the documented security risk). clad benchmark today measures token/compression behavior, not gate accuracy. This is the meta-gap: the thing that judges everything is itself unjudged.

Proposed shape

A labeled detector-accuracy corpus + a scorer, run as a contributor self-audit (sibling to npm run conformance):

• Fixtures tagged {detector, expect: drift|clean} — extend the existing conformance-fixture idea (GOVERNANCE.md §4.1 explicitly welcomes new fail-case fixtures).
• A scorer computes per-detector TP/FP/FN → precision/recall/FPR, aggregated across the corpus.
• A tracked baseline (like perf/baseline.json) so a PR that pushes a detector's FPR up, or recall down, is visible in review.

This turns "we have 37 detectors" into "here is each detector's measured accuracy, and it isn't regressing."

Versioning scope (GOVERNANCE.md §2)

• Minor — new fixtures + a self-audit scorer (no change to shipped detector behavior)

In-scope check (GOVERNANCE.md §4.1 / §4.2)

• Not regressing Iron Law conformance — strengthens its evidentiary basis
• Not bypassing the anti-self-cert guard
• Not forking the Ironclad spec
• Not cosmetic-only — the deliverable is fixtures + a scorer with tests

Alternatives considered

• Rely on the existing 26 conformance fixtures — those verify the contract (does the suite match), not per-detector statistical accuracy on adversarial near-miss cases.

Willing to implement?

• Yes — I'd open a PR (would value scope guidance on the corpus size / labeling convention first)

---

Part of a competitive-gap analysis; flagged as the highest-leverage "integrity of the integrity layer" item.

[Inkayuyu04]

The benefits of adding (Future Ratchet and Dashboard) are minor and redundant, while the benefits of removing (maintaining integrity and eliminating debt) are clearly greater, so it is removed.