fix(daemon): drain accepted handlers on shutdown

[jleni]

Summary

• make the accept loop own every accepted connection handler in a JoinSet
• on shutdown, stop admitting post-shutdown frames, finish current responses under one 30-second deadline, then abort and join leftovers
• preserve the shutdown request's own acknowledgement before its handler exits
• reap completed and failed handler tasks instead of detaching them until runtime teardown

Reliability coverage

• real-listener regression gates an in-flight RemoteCheck, sends an actual Shutdown, requires its ACK while the client remains open, proves the accept loop is still draining, then requires the blocked response before loop completion
• deadline regression proves a silent handler is aborted and joined
• request-boundary regression proves queued and just-completed frames are not dispatched after shutdown begins
• task-outcome regression distinguishes successful, panicked, and deadline-cancelled handlers so operational failure visibility is mutation-tested

Validation

• combined post-fix(daemon): treat remote import failures as misses #841/ci: guard Unix test resources #846 daemon suite through the resource guard: 2,185 passed, 0 failed, 2 ignored with normal process visibility
• post-rebase strict workspace/all-target/all-feature Clippy with -D warnings
• post-rebase focused drain/shutdown suite: 7/7 green
• cargo fmt --all -- --check
• git diff --check
• independent review: GO after the post-shutdown dispatch and active-polling test gaps were fixed
• independent mutation-follow-up review: GO; production classification/logging is unchanged
• two signed commits on current main (91a442c)

CI follow-up

The first changed-line mutation run found two survivors that removed panic/error observability while leaving request behavior intact. The added success/panic/cancellation truth table rejects both prior survivors plus constant and branch-condition variants.

Follow-up boundary

Tokio spawn_blocking closures are non-cancellable and can outlive an aborted outer handler. This PR does not change that existing boundary; hardening blocking handler work is a separate follow-up.