mirror of
https://github.com/stablyai/orca.git
synced 2026-09-22 00:02:31 +00:00
Jinwoo-H
38baeec7c7
Round 4 on #21809. H1. The mount claimed the page before its dynamic import and handed back a promise, so a component cleanup that ran while the chunk was still in flight had nothing to dispose: the claim outlived the mount it was made for, and Reload — the recovery ruling 20 names — was refused as a second document. The claim, the markup and the handle are now made synchronously, `ready` settles on its own, and a mount disposed while its import was in flight releases without starting anything. Pinned in the render check by holding the document chunk 20 s past the 15 s readiness watchdog, clicking Reload and waiting for the second mount to become live; red at that wait before the change. M1. The frame case's precondition asserted that a frame had been asked for while the document owned the page, not that one was owed when it was disposed. The fit retry commits on its first attempt whenever the grid still measures, so a dispose between two refits owed nothing and agreed with an empty leak list for exactly the reason under test — one run in five. The refit and the unmount now share one discrete click, which React flushes before the event returns, and a mutation observer reads the registry at the instant the host is emptied. Five red runs without `cancelDocumentFrames`, all on the leak and none on the precondition, and five green with it. M2. Two mounts handed the same element, which is what the token is for: the other six cases use a different element each, so a host comparison passes all of them. L1. A throw inside the start sequence released the token but ran no stop, leaving the host-notify error listener installed until the next reset nulled its undo. The sequence now unwinds the starts that completed, in reverse, before it rethrows. L2. A render case comparing the window and document listeners the page holds with no terminal on it, before and after a mount, so a stop that forgets one is a failure rather than a second copy per terminal ever shown. L4. Separated the stacked docstrings in the parse-time-effects census. The render check's bundle, server, browser and page helpers move to their own fixture module: the cases are what is under review and the scratch route tree is not, and the file was 16 code lines under its cap. Claude-Session: https://claude.ai/code/session_01JNnE9qzUZMMnqpZWCqM3nb