feat: harness evolution (SafeEvolve-inspired)

[chinkan]

Summary

Evidence collector after tool runs → failure pattern threshold → bounded harness proposal (SOUL/SKILL/config) → human accept/reject gate → versioned apply + rollback.

Why deferred

Research-shaped; static harness is fine until repeated failure patterns justify the loop. No core agent-loop change required for A/B memory work.

Spec source

/tmp/rustfox/PLAN-2026-09-01-memory-tooltimeout-cancel.md Part 6.

Acceptance (when picked up)

• harness_evidence + harness_mutations tables
• record_evidence after tool execute (opt-in)
• threshold → proposal flag; Kan approves via Telegram
• one component per mutation; rollback by hash

[chinkan]

/oc please review the pull request