feat: ratchet mode — measure-before-after for autonomous improvements

[stackbilt-admin]

Summary

Add a ratchet mode to the taskrunner that measures outcomes before and after autonomous code changes, automatically reverting changes that don't measurably improve results.

Why

Current flow: self-improvement files issue → taskrunner creates PR → PR gets merged (or not). No verification that the change actually improved anything. We've shipped "improvements" that introduced regressions (synthesis noise, conversation-facts errors).

The ratchet pattern: improve → re-run → measure → keep only if metrics improve, else revert.

How it works

1. Baseline capture — before task runs, snapshot relevant metrics (test pass rate, typecheck, error count from task_runs table)
2. Task execution — normal taskrunner flow, creates branch + PR
3. Validation run — after task completes, re-run tests/typecheck on the branch
4. Comparison — compare post-task metrics against baseline
5. Decision — if metrics improved or held steady: keep. If regressed: auto-close PR with explanation

Scope

• Opt-in per task via ratchet: true in task config or category-level default
• Best fit for refactor and bugfix categories
• docs and tests categories skip ratchet (no regression risk)
• Metrics to compare: typecheck pass/fail, test suite results, error counts

References

• recursive-improve /ratchet command (Apache 2.0) — overnight improvement loop with keep/revert
• Existing: --loop mode, adversarial governance, PR utility scoring