Deleting a running local channel previously accepted the kill acknowledgment and immediately removed SQLite files while its worker was still closing them. This caused a disk I/O error and left the worker alive after the channel disappeared. Local kill now waits for that connection's cleanup receipt before deletion; a ten-second timeout preserves the channel's files. The worker keeps its shutdown refusal guard reachable until storage/tool cleanup finishes, so another delete during teardown cannot mistake it for dormant. Hosted kill behavior is unchanged.

Closes #119; discovered during #8 validation and tracked under #5.

Validation: full types/lint checks and independent review pass. A real isolated source host/gateway exercised five live SQLite channel deletes, five pairs of concurrent deletes, dormant deletion without starting a worker, and a deliberately paused worker. The timeout left catalog/SQLite/WAL bytes unchanged; resuming and retrying completed deletion. No storage/unhandled errors remained, and all disposable workers and the host exited. Admission of a new worker after complete shutdown is a separately identified lifecycle concern, recorded in #119; this change does not claim to serialize every channel operation.

### Built in Ace

Nate explicitly allowed direct tools for this #8 pass. Codex shell/file tools and review agents reproduced, implemented and validated the cleanup fix; GitHub CLI created this PR. The reproduction and final checks used real Ace gateways, workers and SQLite stores in disposable profiles. No existing channel data or desktop permissions were changed. The observed Ace gap is #119; external harness support remains #9.