Patch release on top of v0.2.0. Full series notes: the 0.2.x line ships the long-running coding-session vertical slice (task contracts, user-owned acceptance checks, verified/failed/not-verified states, bounded repair, cache-aware context accounting) plus the evaluation layer (real-engine suite runner, scripted mock transport, verifier-assisted opt-in policy, campaign audit).
What v0.2.1 fixes (from the post-release advisor review):
- A failed loop-breaker model request is now recorded as a provider error and graded infrastructure_error, never a fabricated pass/fail.
- A repair response that used tools no longer ends the turn silently: one-shot mode gets a deterministic, clearly-labeled completion notice.
- Acknowledged host-mode evaluation runs can pass the audit (containment requirement now matches the mode actually executed; execution_mode recorded per attempt).
- /resume no longer restores transient context-prefix state; cumulative counters still survive.
- /cwd persists the acceptance-contract rebind, and resuming in a different working directory invalidates stale verified status.
- One-shot mode exits nonzero when a provider failure ends the turn.
- Acceptance evidence carries a checked_at timestamp; README claims tightened.
136 tests pass; self-test passes; the existing 40-attempt evaluation records still audit VALID.