Keep each turn's outcome through engine finalization and use it for workflow and cron completion

[groeneai]

Closes #292 (the remaining scope; #428 fixed the Claude translator).

What breaks today

AgentEngine.run() returns normally for a turn that hit max turns, got an API error, was aborted or crashed after streaming. Cron then logs success, and a workflow run ends done because it reads sessions.status, which such a turn leaves (or restores to) active. An aborted turn gets no inline notice, and abnormal backend results are not logged.

Change

• The engine records a TurnOutcome (completed / failed / interrupted + reason) where each turn ends. A normally finished turn publishes it only after it is saved; a failed save publishes failed.
• run_cron / run_persistent_cron return TurnResult(text, outcome) for their own turn; cron logs error with the reason. run() still returns str.
• engine.watch_turns(session_id) reports the latest outcome of a session's turns, background continuations included. A workflow run watches its own session through the existing quiescence wait and ends failed on an abnormal outcome (a missing one fails closed). The session-status check is gone, so turn failure stays separate from session health.
• A failed or interrupted turn reported by the backend (max turns, an API error, an abort) gets an inline Turn <status>[: reason] notice and one WARNING line. The engine's own crash, stop and timeout paths keep their existing notices and log lines.

Scope is option A as agreed in #292 (#292 (comment)). Unchanged limits: a parked isolated cron run is logged from its launching turn; a workflow follow-up turn that ends without content does not change the result; the pre-query drain desync is untouched.

Behaviour changes

• Failed or interrupted cron runs (a user stop included) are logged error, so they do not anchor the interval and are caught up once at restart. I can give interrupted runs their own status if you prefer.
• A workflow run whose last turn was interrupted ends failed; a stop through Nerve is still recorded killed by the stop listener.
• A review-loop leg that hits max turns is retried, then escalated, instead of going to the verifier.

Validation

tests/test_turn_outcomes.py: 26 integration tests (real engine and database) over the engine, cron and workflow paths; each of 13 mutations of the change fails them. Full suite: 3952 passed (main: 3926).

🤖 Generated with Claude Code

[groeneai]

Internal second-model review: adjudication log (click to expand)

Pre-publication review by an independent model (engine: codex; 3 findings) plus my own review.

#	Sev	Finding	Verdict	Evidence / action
1	⚠️	Crash, hard-cancel and background stream-end paths do not get the new Turn <status> notice and WARNING	DISAGREE	#292 asked for the notice and logging on backend results; those paths already show Agent error: ... / [Stopped by user] / [Background turn interrupted ...] and log at ERROR / INFO / WARNING, and keep doing so. The PR text now says the notice is for backend-reported results.
2	⚠️	A forced cancel publishes interrupted before the best-effort save of the partial text	DISAGREE	interrupted is how the turn ended whether or not the partial row is saved; cron and workflow treat any non-completed outcome the same and only read the outcome after the turn has returned.
3	⚠️	Worker onboarding still logs "setup session completed" whatever the outcome	DISAGREE	Unchanged from main, and it writes no cron log or workflow row, so it is outside cron/workflow completion. I can add the branch here if you want it.

Severity: ❌ blocker / ⚠️ major / 💡 nit. DISAGREE verdicts carry recorded evidence and are terminal per finding.

Session id: cron:clickhouse-review-slot-8:20261001-135100