Enqueue a fresh evaluation job for a session. Used by the dashboard's re-evaluate button when a session was missed by the automatic pipeline (e.g. no agent_end ever arrived) or when the user wants fresh scores after deploying a new evaluator.

The eval_enqueued guard on agent_sessions only gates ingest-driven enqueue (so duplicate agent_end events can’t re-evaluate). Manual re-eval bypasses it: we insert directly into evaluation_jobs. The dispatcher’s terminal-result writer already upserts on session_id, so the new run naturally overwrites the previous evaluations row. Re-run the evaluation for one session.

Enqueues a job and returns 202 immediately — the scores are NOT ready when this returns. Poll /evaluation-jobs for progress, /evaluations for the result. The new run overwrites the session’s previous evaluation.

Requires evaluations:trigger, not evaluations:read — a read-only key cannot start work.

One job per session at a time: if an evaluation for this session is already queued or in flight, this is a 409. Wait for it to finish, then trigger again.