chore(main): release 7.181.0

[github-actions]

feat(harness-opt): governed HarnessOpt capstone with held-out validation

Adds /swarm harness-opt plan|run|compare|status|stop|history (issue #2503 — Workstream F PR 11 of 21): the executing, governed optimization capstone over the declarative harness_evolution surface (#1825). Rounds run one governed candidate evaluation at a time through the production evaluation substrate in disposable worktrees (the substrate fingerprints the active checkout before and after every execution, so the running checkout is never mutated); the task set is content-frozen before any round and the materialized input is integrity-verified against its content-addressed hash; a test split consumes the held-out set exactly once, enforced by the substrate's claim ledger; every round records durable lineage with task-cost accounting (tokens_input/tokens_cache/tokens_output; missing host data stays unknown, never zero) and replays by identity against the immutable substrate run.

• Governed round controller (src/services/harness-optimizer/controller.ts): persisted monotonic round counter salts each round's substrate seed so held-out consumption surfaces typed from the substrate itself; enforced budget/circuit stops (transient_retry_budget_exhausted at the exact retry bound, wall_clock_budget_exhausted, spend_budget_exhausted on host-reported spend, round_budget_exhausted from the command path) — every stop-reason member has a producing code path, pinned by producer-discipline tests; pilot-graduation evaluation retains negative evidence verbatim; a proper-lockfile serializes all mutations.
• Comparative evaluation protocol (comparative.ts): baseline, ablation, and simple-agent control arms on the same frozen task population, per-arm denominators, retained negative results, a stream-snapshot duplicate guard (STREAM_SNAPSHOT_DUPLICATE), and arm toggles honored from harness_opt.run_ablation_arm / run_simple_agent_arm.
• Independent oracle (oracle.ts): scores accepted task outcomes, artifact validity, verification evidence, and completion quality separately from the optimizer; rejects a token-improving candidate whose artifact quality or verification evidence falls.
• Frozen manifest validation (manifest.ts): rejects empty required fields and improvement claims without a measured result.
• Store hardening (src/harness/store.ts): activateHarnessCandidate and rollbackHarnessVersion now accept the live harness_evolution config and re-validate a recorded candidate's approved paths against the CURRENT source allowlist, refusing typed allowlist_revoked.
• Wiring: harness_opt config block (disabled by default; run is human-only and requires --confirm), swarm-harness-opt TUI shortcut, config-doctor case, regenerated schema/commands/configuration docs, README section, new docs/harness-optimizer.md, and a harness-opt-store retention-registry row for the new .swarm/evolution/harness-opt/ state.

🤖 I have created a release beep boop

7.181.0 (2026-09-14)

Features

• harness-opt: governed HarnessOpt capstone with held-out validation (#2503) (fc9876b)

Bug Fixes

• harness-opt: bound wall-clock via cleared timer instead of native AbortSignal.timeout (#2503 CI) (f78fba6)

---

This PR was generated with Release Please. See documentation.

[github-actions]

Drift check report

Found 2 drift finding(s): 0 error, 0 warning, 2 notice.

required-check-contract (2)

• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not yet required by the captured ruleset
• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not present for every expected event in captured external workflow evidence

[github-actions]

🤖 Created releases:

• v7.181.0

🌻