[DaXu999]

Why

authority.test.ts, "refuses the later of two opens that race under one execution id", failed about one run in 50 on main. The test was right and the kernel was wrong: openTurnAuthority let both opens load authority and gave the turn to whichever load finished first. Each load hashes with crypto.subtle.digest on the libuv thread pool, so the two can finish in either order, and now and then the later call held the turn while the first was refused with execution_in_progress. Found while verifying #116.

What

• A context with an execution id claims its turn before it loads. A second open under the same id is refused at once and without a store read, as it already was once the turn was held.
• A cancelled or failed load frees the claim, so the turn can be opened again.
• Opens without an execution id are unchanged: they still load and join the lease.
• The race test now makes the later load finish first, so it fails on the old kernel every time. It checks that the later open is refused before any second store read. A new test frees a turn whose open was cancelled while it loaded.
• docs/api/sharedos-core.md is regenerated for the openTurnAuthority doc comment. CHANGELOG entry under Fixed.

kernel.ts: 22 lines added, 12 removed. Tests: 39 added, 6 removed.

Related

Independent of #116; either can merge first.

Checks

• pnpm check is green in a plain clone: format, typecheck, vitest 1156/1156, test:release 7/7, docs:api:check and conformance:check.
• The authority test file passed 50 runs out of 50. With the old kernel.ts, the rewritten race test fails every time.