Prevent low-index sockets from starving Windows event_wait results

`event_wait` polls all registered sockets, but copies ready entries starting at index zero on every call and stops when the caller output capacity is full. With two continuously writable sockets and capacity one, the first entry can be returned indefinitely while the second is never delivered. A bounded event buffer should not make later registrations unobservable.

Source evidence at the head of #520:

- [std/sys/windows/event.wave:84](https://github.com/wavefnd/Wave/blob/2ae0a91c57fd76323124c0e4dff552adbfadbe1b/std/sys/windows/event.wave#L84) — `pub fun event_wait`

Acceptance:

- [ ] Rotate the starting point or retain pending readiness so repeatedly ready earlier entries cannot starve later ones.
- [ ] Use deterministic socket fixtures with more ready registrations than result slots, and assert every token is eventually observed within a bounded number of polls.
- [ ] Cover removal/reinsertion and mixed readable/writable/error results without stale tokens.
- [ ] Preserve the existing small readiness fallback; this does not require replacing it with IOCP.

Related: #385 and #445; this is the synchronous readiness fallback, not the IOCP provider.

Audit status: identified by static source inspection; the scenarios above have not been executed during this audit. The permalink fixes the reviewed revision; this report does not claim the defect was introduced by #520.

Start at std/sys/windows/event.wave:84 and trace pub fun event_wait through the synchronous readiness fallback. Add deterministic socket fixtures with more ready registrations than result slots, then verify bounded polling observes every token, including removal/reinsertion and mixed readable, writable, and error results without stale tokens. Preserve the existing fallback rather than replacing it with IOCP.

Written by the indexing model from the issue text.