fix(apple-runner): fence prep spawns after teardown or last-waiter cancellation

[thymikee]

Purpose

A runner start must stop creating preparation processes after its execution host is torn down or its last interested waiter cancels. Stopping the currently registered children once is insufficient.

PR #3193 measured this during close with a cold build: close killed the first build-for-testing before obtaining the runner session lock, but the detached prewarm health retry spawned a second build while close waited for that lock. Close reached its timeout; the second build remained until daemon stop. Original evidence: #3193 (comment). This observation is at b6cdd3c49, not a fresh reproduction on main.

Required behavior

• The in-flight runner start owns whether further preparation is admitted. Device teardown closes that admission before stopping current prep children or waiting for the session lock.
• Canceling the final interested waiter closes admission for that start. Canceling one waiter must preserve work still needed by another.
• Check admission at every preparation spawn, including retries after a killed build. A check only when registering an already spawned child is too late.
• A later independent open may create a fresh start; a retired start cannot resume and publish a runner into it.
• Prefer one start-owned cancellation/admission mechanism. Reassess whether the request-owner prep filter introduced by fix(daemon): scope client timeout recovery to the timed-out request #3193 remains necessary once the reachable callers and waiter semantics are covered.

Done when

• A control holds the runner session lock during a cold build, begins non-retained close, and attempts a retry after the first child is stopped. No second prep child is spawned; close settles without waiting for a replacement build.
• Controls cover cancellation before the first prep spawn, last-waiter cancellation, and another still-interested waiter.
• Repeat the live close-during-build route with evidence of no respawn, completed close, and cleanup. Retain the original evidence attribution.
• Focused tests and the affected gate pass; process signaling remains scoped to owned children.

Depends on #3193. This is start admission, not another host-wide timeout sweep or a shorter join timeout.