[Fix] End a Fast turn's inference once its closeout is delivered

[mrubens]

Stacked on #2029 (diagnostics); retarget to develop once that merges.

Problem

After send_chat_reply posts a closeout (or a closeout reaction or ignore_event ends the turn), OpenCode still starts one more model request to consume the tool result. The prompt tells the model to do nothing after a closeout, but the request happens regardless: it re-sends the full context, produces text nobody sees, and holds the conversation's turn lock for as long as the provider takes. On the nightly this was 10–25 s per turn (inferenceDurationMs minus firstResponseDurationMs), during which any follow-up in the same conversation queues behind a turn that is already done.

Change

• When a new assistant message starts while the current instruction is closed, the turn aborts the OpenCode prompt with a dedicated reason. The prompt runner's existing abort path issues session.abort server-side, so the model stops generating rather than just the HTTP request being dropped.
• Before aborting, the handler waits for any in-flight native steer drain. A queued follow-up that reopens the turn (new instruction version) wins and the abort is skipped; the existing "steer after closeout" test covers that race.
• The inference attempt treats this abort as success (the visible reply already went out) and returns an empty prompt text, which the closed turn ignores. No retry, no failure classification, no session invalidation, so the warm OpenCode session survives for the next turn.
• Diagnostics gain abortedAfterCloseout on the log line and aborted_after_closeout on the settled-turn telemetry, so the saving is measurable against postReplyInferenceDurationMs from [Feat] Report OpenCode setup phases and model usage in Fast turn diagnostics #2029.

Follow-ups admitted while a turn runs are durable parent events delivered as fresh turns once the lock frees, so ending the turn earlier cannot strand one.

Verification

• New test: closeout on the first assistant message, second assistant message starts, the prompt signal is aborted only then, the turn resolves to the closeout, postReply is called once, no failure is classified, and telemetry reports success with abortedAfterCloseout: true.
• src/server/fast-agent: 356 tests pass. Typecheck, oxlint, knip, and pre-push clean.

[roomote-community]

No new code issues found. See task

_{Reviewed a459478}