TranscriptRecorder

Record one streamed message when its transcript identity is stable.

While an attempt scope is open for the resolved agent_id, the record is staged rather than appended, and only reaches the transcript when complete_attempt commits it.

Open (or replace) the active staging scope for an agent.

Staged messages buffer in memory until complete_attempt commits them, so retried attempts do not duplicate transcript records.

Commit staged messages for a matching attempt, then clear it.

Duplicate or mismatched calls are ignored idempotently.

Drop staged messages and pending chunks for a matching attempt.

Duplicate or mismatched calls are ignored idempotently.

Clear all attempt scopes and chunk accumulators without appending.

Terminal teardown: these attempts never completed (abort, retry-budget exhaustion, cancellation), so their records are not transcript history. Dropping non-empty staging makes the on-screen conversation and the persisted transcript diverge, so count it rather than clearing silently.

Append checkpoint or input messages to the root transcript.

Deliberately bypasses attempt staging: these messages are turn inputs and checkpoint history, not model output, so no attempt owns them and none should be able to discard them. Callers append them between turns, never while a scope is open -- doing so would order them ahead of model output that was staged earlier.