fix(daemon): avoid duplicate startup work and clean cancellation

[Marsu6996]

Problem

The daemon warmed the dispatch surface once before binding the socket and then
repeated the same encode inside the remaining boot warm-up. A timeout around
the pre-bind worker also cancelled its asyncio task even though the underlying
thread could not be cancelled, leaving that continuation implicit. Finally,
cancellation escaping the inner run loop could bypass async-write and lifecycle
lock cleanup.

The original regression test also cancelled startup after a fixed 150 ms. Once
the duplicate eager encode was removed, that sleep could expire before the real
pre-bind warm-up ran, producing a false 0 calls failure on loaded macOS
runners.

Solution

• construct and retain one pre-bind warm-up task, shielding its worker across
  the bounded wait;
• tell the remaining boot warm-up not to repeat the dispatch encode;
• keep the warmed embedder without issuing a second eager encode;
• release async-write state and the lifecycle lock from the outer cleanup path,
  guarded by current-process lock ownership;
• use the canonical daemon socket override in daemon tests;
• synchronize the pre-warm regression test on completion of the actual
  warm_dispatch_surface call instead of sleeping for an arbitrary duration,
  while always cancelling and awaiting the daemon in finally.

Guarantees and scope

The daemon protocol, socket lifecycle, and normal hibernation policy are
unchanged. The patch removes duplicate startup work and makes cancellation
cleanup explicit. The synchronized test remains red on the previous duplicate
encode path and green only when exactly one dispatch warm-up occurs.

Validation

• synchronized pre-warm regression: 20/20 repeated isolated runs;
• daemon, boot warm-up, and async-write cleanup set: 31 passed, 1 skipped;
• focused socket, cancellation, and watchdog coverage;
• exhaustive default-suite result recorded on the final stacked tip;
• adversarial review incorporated the task-lifetime and lock-ownership checks.

Designed by Marsu — Refined by Codex.

Public CI result

The amended macOS run confirms that test_prewarm_called_once_at_boot now passes. Its remaining 14 failures are all pre-existing baseline failures addressed by the next two stacked PRs: 11 by #68 and the final 3 spatial cases by #69.
The cumulative #69 tip is fully green, including the synchronized daemon regression, wheel build, and secret scan.