fix: consequence-based reviews and persistent follow-up context

[vsem-azamat]

Summary

• Make blocking review findings consequence-based: require a concrete defect, governing requirement violation, or realistic material risk. Style preferences alone are advisory.
• Carry the latest substantive changes-requested findings into follow-up reviews. Technical errors preserve context; approval clears it; task resets stay isolated.
• Separate trusted follow-up instructions from untrusted finding data and explicitly describe the supplied diff as cumulative.
• Prevent Deferred labels from automatically waiving defects in retained code; align canonical docs and generated distributions.

Verification

• npm test: 817 tests, 813 passed, 4 skipped, 0 failed.
• npm run check: exit 0 (14 warnings and 1 informational diagnostic; not warning-free).
• npm run selfcheck: stdd check: OK.
• git diff --check: passed.
• Independent read-only review and follow-up: four findings addressed; no remaining material blockers. Review was conducted outside the STDD review runner.
• Regression coverage includes unchanged-snapshot follow-up, changes-requested → error, approved → error without resurrection, task-reset isolation, and prompt trust boundaries.

Documentation evidence

Docs updated first: method/README.md, method/reference-commands.md, playbooks/planning.md

Scope and limitations

• Verdict derivation, schema, gate behavior, round budget, and snapshot freshness contracts are unchanged.
• Budget-exhaustion deadlock and its existing defer/proceed message remain a separate task.
• No live reviewer-runner benchmark: fewer review rounds or token savings are not established by these tests.
• Live runners were stubbed in regression tests. No package publication, deployment, or consumer installation.
• Canonical method is close to its existing size ceiling; an unused import remains as a non-blocking lint warning in the reviewed candidate.