Request-artifact telemetry: fixed 5 s export acknowledgement fails slow-but-healthy writes and leaves abandoned writes running #3866

Background

PR #3852 (#854) added request-artifact telemetry. When conversation logging is on, each request and response chunk is emitted as a log record with an export receipt, and the send waits for the file exporter to confirm the write (AcknowledgedLogExporter.receipt, packages/telemetry/src/telemetry/artifact-exporter.ts). If the write fails, that send fails, so a logged turn cannot silently lose its transcript.

The receipt has a fixed 5,000 ms timeout. When it fires:

• the send fails with ExportAcknowledgementMissingError;
• the export is marked finished as FAILED, so root flush/close stop waiting for it (commit c4d9428);
• the root does not report that error again at flush or close (commit aab7e4f).

Problems

1. Slow is treated as dead. The timeout does not depend on chunk size or disk speed. A large transcript chunk on a slow or network filesystem that takes more than 5 s to write fails the user's turn even though the write would succeed.
2. Abandoned writes keep running. The underlying export is not cancelled, so the write can still finish in the background, and its late result is ignored.

Proposal

• Make the deadline proportional to the chunk size, or configurable, or both. Or replace it with progress-based liveness: fail only when the exporter makes no progress for N seconds.
• Pass an abort signal to the underlying exporter so an abandoned write is actually stopped, and its scratch cleaned up.

Acceptance

• A test with an exporter that writes slowly but steadily for longer than 5 s succeeds.
• A test with an exporter that stalls fails within the liveness bound, and the stalled write is aborted.
• The existing artifact-exporter-abandoned, source-telemetry-correlated and turn-request-artifact-sink tests stay green.

Follow-up from #854 / PR #3852.