feat: add governed agent approval loop

[fly1d]

Outcome

Deliver the first framework-neutral governance loop:

• register and inspect agents
• activate or pause an agent with optimistic revision checks
• allow only active agents to request approval
• decide human approvals exactly once
• query append-only audit events
• keep persistence behind an owned store protocol

The bundled adapter is explicitly development-only: state is process-local and actor identity is not yet authenticated.

Evidence

• Focused tests added or updated
• make check passes: 22 tests, 100% coverage
• Compatibility impact reviewed: additive v1 endpoints and schemas
• Monitoring or audit impact reviewed: every mutation emits an audit event
• Python 3.11 container HTTP smoke: inactive denied -> active -> approve -> audit
• OpenAPI contract generation verified: 10 paths
• Review finding fixed in 3354684: inactive agents can no longer request approval

Risk

• Level: medium
• Failure mode: process restart loses in-memory state; unauthenticated callers can self-assert actor identity; duplicate client retries are not yet idempotent
• Rollback plan: revert this PR; existing health and AgentSpec validation endpoints remain unchanged

Release notes

Added the first agent registration, lifecycle, human approval, and audit API loop.

[fly1d]

Second-pass code review complete: no remaining findings. The initial review identified that inactive agents could request approvals; commit 3354684 closes that gap with a fail-closed 409 response and unit, API, and container-level evidence. The store boundary, optimistic revisions, single-decision approvals, structured failures, and explicit non-production limitations are coherent. Both CI checks pass.