Greptile round two: the 30/40s finalize waits were shorter than the s3 per-op timeouts, so a slow-but-working upload could be abandoned by a worker exiting even though waiting would have saved the recording. At teardown up to ~128 MB of backpressured parts can still be uploading; give them 5 minutes — a post-stream worker lingering is cheap, a lost recording isn't. A genuinely stalled connection stays bounded by the per-op timeouts and is unrecoverable under any window; past the wait the recording is abandoned and logged, and bucket lifecycle rules should reap the dangling multipart. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>