[lbacik]

Fixes #385

Makes recovery-retry and errors-reset pass in the producer suite. The issue's Agent Brief narrows #385 to these two scenarios.

Changes

• recovery retry on a finalized attempt completes as a no-op. It commits a recovery_retry command as completed with a detail saying nothing remained to retry, and changes nothing else. A resend with the same request ID replays the stored record.
• An unknown attempt ID answers not_found. recovery retry and recovery release with an ID that matches neither the retained attempt nor any recorded attempt answer not_found (exit 5) on the contract path. Nothing is stored. The legacy path still exits 1.
• recovery release on a finalized attempt stays rejected.
• errors reset on a clear guard is rejected. With a readable count of 0 nothing is stored, no consecutive_errors_reset event is logged and the request ID stays reusable. This applies with and without --contract and to the consecutive-errors reset alias. An unreadable count still resets.
• Docs. operator-onboarding.md and agentctl-command-replies.md describe the new replies.

Deviations from the brief

• An accepted retry record whose attempt finalized in the meantime (a crash after acceptance) is now resolved to completed. Before, it was rejected on every resend.
• The completion ledger is read before the unknown-ID check, because the finalized case has to be known first. An unknown ID with an unreadable ledger now answers the ledger error (rejected), not not_found.
• agentctl-command-replies.md gains a section for the two legacy replies that change (errors reset at 0 now exits 1; recovery retry on a finalized attempt now exits 0).

Verification

• Producer suite in Docker, both variants (--require-all with all capabilities, and --capabilities ""): 39 PASS, 2 FAIL. recovery-retry and errors-reset pass. The two failures (recovery-release, offline-reads) are out of scope (recovery-release scenario releases an active and an already finished attempt, not a retained one agent-operator-contract#36, Producer suite v1.0.1: offline-reads fails on resume during startup reconciliation #388), so the producer job stays red.
• uv run pytest locally on macOS: 1984 passed, 10 failed. The 10 (test_attempt_processes.py, test_control_offline.py) fail identically on a clean main on that machine; CI's test jobs pass on main.
• uv run ruff check: clean.

Open points

• The docs mark the legacy reply changes "since v0.9.1", assuming that is the next release after v0.9.0.
• A finalized attempt whose checkpoint was left behind (a crash between finalization and checkpoint removal): recovery retry now answers completed, but the checkpoint still holds intake until the next restart removes it. The command's detail says so.

🤖 Generated with Claude Code