fix: keep completed reviewed actions terminal on replay

[kvnloo]

What changed

Fixes #36.

The action layer already treats an idempotency replay as the same logical action and can return an existing succeeded proposal. The task/model layer previously treated that returned proposal as though it were a fresh review, reinstalling actionId and setting the task back to waiting_approval.

This patch preserves the terminal receipt:

• AgentService.prepare() returns an already-succeeded proposal without checkpointing it back into the task.
• terminal non-success review states remain explicit errors rather than becoming a new review.
• prepare_email / prepare_event treat a succeeded replay as an existing receipt, checkpoint approvalResult, keep actionId clear, and allow the model run to continue.
• ordinary awaiting_review behavior is unchanged.

The regression forces the model to call the exact same prepare_event after the reviewed action has succeeded, then verifies there is still exactly one action, the task never ends in a fresh approval state, and the existing receipt is visible to the continuing model run.

Verification

• Reviewed the branch against current main at 82ff35b.
• Existing tests/actions.test.ts already pins that an idempotent proposal replay returns the completed action without a second provider preparation.
• Added a model-worker regression for the missing task/model boundary.
• Repository tests were not run locally because this execution environment cannot clone/install the repository; CI on this PR is the executable verification.

Integration limits

Fixture-only model/action regression; no live Google write was performed.

AI-use note: I used an AI assistant to trace the reviewed-action state machine, draft the regression and patch, and review the final diff. I verified the replay behavior against the existing action-level idempotency regression before opening this PR.

[davidmckayv]

Needs one change before merge. The new test in tests/model-worker.test.ts builds createApp without intelligenceApiKey, so on current main it fails with "OpenMuse requires CPK_INTELLIGENCE_API_KEY". Add intelligenceApiKey: "test-project-key-never-sent", after agentBackend: "model", in that test's config, and remove the trailing blank line at the end of the file.

[davidmckayv]

The missing Intelligence test setting is fixed. The merge conflict preserves both this replay regression and the browser-evidence regression from #29. Local lint, typecheck, all 16 focused tests, and the server build pass; merging is gated on the current CI run.