(remote): bind watch-channel handlers to their child, add ssh keepalive and failure logs

[devsuitup]

Audit 2026-09-11, findings 1, 4 and 11 (.work-files/switchboard/audit-fable-2026-09-11.md, local note).

Why

stop() followed by start() on the same host raced the previous child's asynchronous close: its handler was bound to the per-host state, so when it fired it dropped the new child from the state and scheduled a third one. Result: an unkillable ssh -tt … inotifywait per host on every manual refresh, per-host reconnect, or host toggle in Settings (the toggle path ships in v0.0.75). Reproduced with a fake child whose close arrives one tick late: 2 children alive after stop+start, 3 spawned after the restart timer, 1 survivor after stopAll().

What

• Every child handler (data, stderr, close) checks s.child === child before touching the state; a superseded child's late close or chunk is ignored.
• The watch ssh gets ConnectTimeout=10, ServerAliveInterval=30, ServerAliveCountMax=3, so a half-open session dies and gets restarted instead of staying "running" forever.
• A quick failure logs the exit code and the last 200 bytes of stderr once per backoff tier (same throttle as the inventory backoff) — a dead watch channel is no longer silent in main.log.
• stop() clears pending coalesce timers so onEvent cannot fire after stop.
• Test helper now emits close asynchronously, like a real ChildProcess.

Proof

Mutations: remove the identity guard in onExit → the F1 race test and the F4 duplicate-close test go red; drop the stderr capture → the warn-content test red. task check: 1256 pass, 0 fail, 8 pre-existing skips.