Adopt security hardening from PR #17

[imsatoshi]

Extracts the reviewable, self-contained security hardening from #17 (contributed by @jehonathanBao) and reconciles it with the fixes already on main from #18.

The rest of #17 — bollinger_evolver/ (212 files), frontend/ (85 files), and 66 stage_*.md process documents — is not included here. See the review notes on #17 for why.

Agent API

• Reject API keys passed via query string; headers only. Query keys leak into access logs, browser history, and reverse-proxy diagnostics.
• Replace wildcard CORS with an explicit origin allowlist.
• Refuse to start with a missing, weak, or short (<16 char) master key. default-key and the placeholder keys are rejected by name.
• Bind 127.0.0.1 by default instead of 0.0.0.0; publish the container port on 127.0.0.1 only and pass --api-host explicitly in compose.
• Stop echoing internal exception text in 500 responses.
• Leave /api/v1/health unauthenticated (returns only liveness + timestamp) so container healthchecks work; every other endpoint still requires auth.
• Close the listening socket on stop().

Deployment safety — fail closed instead of pretending

• StrategyDeployer no longer auto-approves when no approval callback is set; it fails the deployment and says why.
• Deployment notes now state plainly that shadow trading and gradual rollout are not executed, instead of implying a pipeline that never runs. total_phases reflects reality.
• AdaptiveOptimizer no longer auto-approves purely on improvement percentage.
• RollbackManager defaults to enabled=False / require_confirmation=True, and refuses to run without a deploy callback rather than reporting a rollback that never restored the strategy file.
• New rollback_require_confirmation setting so operators who explicitly opt into auto_rollback_enabled still get unattended rollback; the dataclass default stays fail-closed for anyone constructing RollbackConfig() bare.

Reconciliation

PR #17 predates #18, so applying it verbatim would have reverted four things now on main: the read-permission gate on GET endpoints, the rate limiter, the real rollback endpoint, and the deployer/rollback-manager injection into AdaptiveOptimizer. Those are kept.

tests/test_api_hardening.py is new and pins those invariants — a write-only key cannot read, read endpoints are rate limited, the rollback endpoint delegates to the deployer and reports failure honestly, and the rollback manager never claims success without a deploy callback. A future merge cannot silently drop them again.

Test plan

• python3 -m pytest tests/ -q → 232 passed (210 on main; +14 from Add mock-first GA pipeline, safety gates, and custom strategy abstraction #17's tests, +8 new regression tests)
• Weak-key rejection verified end to end against ga.json.example
• docker-compose.adaptive.yml and ga.json.example parse cleanly

🤖 Generated with Claude Code