fix(memory-core): a Windows sharing violation on the head wake ticket no longer rejects the waiter

[code-yeongyu]

Summary

Fixes the Windows-only red in createKibitzerWakeSlot "ten sessions contending for two slots ... admission is FIFO" (senpi-compatibility windows, runs 36312352680 and 36300728869), together with the palace test failures that cascaded from it in the same job.

Root cause (product)

reapDeadHead in packages/memory-core/src/locks/recall-wake-domain.ts reads the head wake ticket to prove its owner dead. On Windows, opening a ticket while its owner is publishing or withdrawing it can fail with a sharing violation (EPERM/EBUSY/EACCES). That error was rethrown, which rejected a queued waiter. The FIFO test saw that as the next waiter "already settled" (wake-slot.test.ts:90). The remaining waiters kept polling after the failed test and threw inside later palace tests ("expected an acquired admission, got busy", with the stack pointing into wake-slot.test.ts:72).

Fix

• On win32, a sharing violation reading the head ticket keeps that ticket as the queue head until the next normal poll. ENOENT handling, dead-owner reclamation, FIFO order, and bounded waiting are unchanged.
• Tests: a deterministic regression injects the sharing violation through a narrow ticket-fs seam and proves both waiters acquire in order with no ticket left. The FIFO stress tests (memory-core and the omo-senpi adapter) now abort and drain every waiter in finally, so a failure can no longer leak work into later tests. The old settledAlready race helper is replaced by asserting the admitted order and the live count after every release.
• Bundles regenerated on linux/amd64 (node 24, bun 1.4.2, build-extension.mjs --check green in the container). The diff is limited to the three bundles that embed memory-core.

QA & Evidence

• Focused memory-core plus adapter tests: green. Typecheck: tsgo -p packages/memory-core and -p packages/omo-senpi exit 0. Biome clean.
• Local macOS stress (lane run): 750/750.
• The 8 palace generator tests time out locally only under extreme host load, identically on base (0/8 on base with the same load), and are unrelated to this change.
• Windows focused soaks (3 x 10 iterations) are listed in a comment.

Fixes #8953
Refs #8250, #8324

---

Summary by cubic

Fixes the Windows-only FIFO admission failures in createKibitzerWakeSlot caused by a sharing violation when reading the head wake ticket. On Windows, opening a ticket while its owner publishes or withdraws it can fail with EPERM/EBUSY/EACCES, which previously rejected a queued waiter and let remaining waiters poll into later tests. The unreadable head ticket now keeps its place until the next normal poll; dead-owner reclamation, FIFO order, and bounded waiting are unchanged. The omo-senpi extension bundles that embed memory-core were regenerated.

Tests

• Adds a deterministic regression that injects the sharing violation and proves both waiters acquire in order with no ticket left.
• The FIFO stress tests abort and drain every waiter in finally so failures can't leak into later tests.
• Sharing-violation waiters use a 10 s circuit-breaker budget, and a rejected follower rethrows its own error.

Fixes #8953.

^{Written for commit 6d601f2. Summary will update on new commits.}

[code-yeongyu]

Focused Windows soaks of the final head (windows-latest, 50 iterations each, wake-slot + recall-wake-domain + palace tests): 36327253718 success, 36327255724 success, 36327257610 success. An earlier 10-iteration soak exposed a 400 ms waiter budget in the new sharing-violation test (the follower reported busy, product unaffected); the test now uses a 10 s circuit breaker, commit 741c06a.

[code-yeongyu]

Superseded by #8961, which lands this fix together with the other win-ci fixes as one merge (all commits of this branch kept, bundles regenerated once over the merged sources). The evidence in this PR (RED/GREEN, focused Windows soaks) still applies; the issue is closed by #8961's Fixes line.