[lbacik]

Fixes #390

The engine now records an attempt's verdict when publication decides it, and no recovery path recomputes it afterwards (contract §14). With the suite pinned at v1.0.4, the producer job passes recovery-release.

Changes

• Recording at publication. As soon as the publisher returns, the engine writes the outcome and a short reason into the attempt checkpoint (new optional verdict field) and the outcome into the archive record. This happens before the result-comment check, the issue release and cleanup. Only the first recording counts. Old checkpoints without the field read as "no recorded verdict".
• outcome_recorded once per attempt. It is emitted at the recording moment. _commit_finalization emits it only when no verdict was recorded, so an attempt interrupted before publication returned still announces its outcome once. The Contract Event order is unchanged.
• recovery release finalizes with the recorded verdict and lifecycle_state: released. The confirmed release comment names the verdict and --saved-at. It is followed by a plain release (never round-finished), and the branch is retained.
• recovery retry republishes towards the recorded verdict and finalizes with it. It answers not_fulfilled and keeps the hold when the verdict is unreachable or the publisher answers another outcome. Two cases are unreachable: a handoff whose note is gone, and a complete the checkpoint no longer shows as publishing, which would push the branch first.
• Startup reconciliation after a restart steers the same way. Before this change it re-inferred the verdict and could finalize a recorded infrastructure_error as complete. When the verdict cannot be reached, the attempt stays retained for the operator and the engine emits recorded_verdict_held.
• No recorded verdict → today's inference, unchanged, including Recover and release retained implementation attempts #84's rule that an unpublished branch is never a success.
• Docs. ADR 0010 records the decision and its relation to Recover and release retained implementation attempts #84. CONTEXT.md gains "Recorded verdict" and "Retained Attempt". docs/conformance-harness.md now says the after-outcome hold takes effect before the result comment is confirmed.
• Pin. pyproject.toml names agent-operator-contract v1.0.4. That release fixes recovery-release scenario expects a claim to open without a release verb agent-operator-contract#39: the last step of recovery-release used to expect a Claim without a release verb, and it now opens the later Claim with set-issue followed by release. Harness protocol stays 1.1. uv.lock is gitignored, so only the local lockfile was regenerated. It names rev=v1.0.4#5e17e8c.
• Tests. The new tests/test_recorded_verdict.py covers release, retry, restart and once-only emission. The checkpoint tests cover the new field. The xfail is removed from test_recovery_release_finalizes_a_retained_attempt. One event-order test now expects outcome_recorded before the error-limit halt. Two retry tests now advance the checkpoint to publishing, as the production publisher does, because the fake publisher does not.

Verification

• Producer suite v1.0.4 in Docker, --require-all with all capabilities: 41 PASS, exit 0, including recovery-retry and recovery-release.
• The same with --capabilities "": 41 PASS, exit 0.
• At v1.0.3, before the pin bump, both runs were 40 PASS and 1 FAIL. The only failure was the last step of recovery-release (a later eligible issue must open a claim after the release). Every verdict check already passed.
• tests/test_conformance_harness.py and the contract tests on v1.0.4: 128 passed.
• uv run pytest locally on macOS: 2017 passed, 10 failed. The 10 (test_attempt_processes.py, test_control_offline.py) fail identically on a clean main on that machine.

🤖 Generated with Claude Code