Reliability: gate real loop regressions and token efficiency

[Wibias]

Stack

PR 6 of the v0.5.0 reliability/convergence rewrite. Stacked on PR #227.

Scope

• replay the real Baseline-is-green/tool-emission incident
• replay malformed <atool> protocol generation
• replay cross-channel narration stalls
• replay the repeated GitHub Actions evidence spiral
• enforce hard production bounds for unique no-progress generation and cumulative output tokens
• add false-positive corpus for legitimate tool-rich investigation and long final verdicts
• calibrate final-response allowance separately from active-workflow budgets without disabling protocol/tool-emission detection
• make these quantitative regression gates release-blocking

TDD status

This PR begins RED where the current production defaults would interrupt a legitimate long completed-plan final verdict. The implementation must distinguish finalization from active no-progress work without weakening the real-incident bounds.

Stack contract

v0.5.0 release remains blocked until this PR and all earlier stack PRs are green.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.