chore: make PR review findings evidence-gated

[bluestreak01]

Summary

• align the Pi and Claude review-pr skills around a private HYPOTHESIS → FALSIFYING → PROVEN → ADMITTED workflow
• require independent, neutral-payload falsification plus producer, reachability, base/head, impact, and artifact evidence before publishing behavioral findings
• classify missing coverage proportionally instead of treating missing tests, a bug-fix label, or a missing tandem PR as automatically Critical
• preserve the Java client’s Java 8 floor, zero-GC, QWP, committed-binary, blackbox, and cross-repo e2e/tandem policies

Why

A review candidate is not a verified finding. The previous wording could turn plausible source-reading narratives—or test absence by itself—into blocking findings before the load-bearing premise had been disproved or executed. This change makes omission the required outcome for unsupported behavioral claims and reserves Critical coverage gaps for material, reachable user/operator risk.

Validation

• git diff --check
• balanced Markdown fences and valid skill frontmatter
• Pi/Claude semantic-parity audit for coverage and admission sections
• stale automatic-Critical directive scan
• independent read-only review: PASS

Documentation/agent-policy only; no runtime product tests are applicable.

[RaphDal]

	Fails
🚫	Please update the PR title to match this format: type(subType): description Where type is one of: feat, fix, chore, docs, style, refactor, perf, test, ci, revert And: subType is one of: build, log, core, ilp, qwp, http, conf, utils For Example: perf(sql): improve pattern matching performance for SELECT sub-queries

Generated by 🚫 dangerJS against cf66e18