Hunt and fix real bugs with proof

Recurring correctness sweeps where bugs must be proven by test before shipping (not style/design audits)

• · codebase sections per lens (concurrency, error-paths, boundaries, state-integrity, security)
• · test suite (pytest, cargo test, vitest)
• · optional maxFixes limit

• · proven bugs (title, file, line, severity, trigger)
• · worktree branches with regression tests
• · PR URLs with fixes ready to merge

• · agent (fable model for finding, sonnet for proof)
• · git worktree (isolation)
• · test runners (pytest, cargo, vitest)
• · gh CLI

• · pytest environment configured (agent/)
• · cargo test working (cli/, vestad/)
• · vitest working (web/)
• · origin/master fetchable and clean

• · unproven claim without failing test (rejected)
• · test determinism fails (flaky tests)
• · quality metric regression on fix
• · concurrency/integration issues hard to prove
• · bug claim lacks concrete trigger sequence

• · North Star: failed test is gate; no plausible-only claims
• · 5 diverse finder lenses per round (concurrency, error-paths, boundaries, state, security)
• · Max 6 fixes per run; rounds until dry
• · Fable (deep reasoning) for finding, Sonnet for proof
• · Regression test co-commits with fix
• · Worktree isolation: never touches main branch