Sterister
baddf5f56e
Some checks failed
Test / test (push) Has been cancelled
Second flake in the same hunt, and a real bug rather than a test artifact.
`runRoundRobinUpload` and its resumed twin ended with `await Promise.all(workers)`.
`Promise.all` rejects the moment one worker throws and leaves the rest running
with nobody holding their rejections — so an abort, which fails every lane at
once, produces one propagated error and N-1 unhandled ones.
Bun counts an unhandled rejection as a test failure, which is how
(fail) Resume protocol — kill-restart-verify
→ sender crash mid-transfer → resumeUpload completes the same stream
TransferAbortError: Aborted during attempt
failed about one run in fourteen, with an error the test had already caught
properly through `handle1.done()`. The test was right; the engine was dropping
rejections on the floor beside it.
`settleWorkers` waits for every lane, then reports the first real failure —
preferring a genuine error over `TransferAbortError` when both happened, since
an abort is not the interesting one when something actually broke. Four call
sites.
This is not only a test problem. In production the same shape means a
cancelled or failed multi-lane transfer emits unhandled rejections into
whatever process is running the engine, and on a strict runtime that is a
crash rather than a log line.
Verified: 1198 pass / 0 fail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014489bKUtUEY1Zgs9xN9mt7