kernel/REPL: an unexpected kernel exit bricks the session forever, and the watchdog has no liveness channel

[Dmatut7]

Verified against main (fb2db8ee1).

1. Kernel death = permanent session brick. ReplKernelManager state parks at shutdown; restart() has no production callers; every subsequent cell fails with "Kernel has been shut down". We had two production sessions each fail 7 times in a row. Our fix: attribute the exit (code/signal/origin), auto-revive with an hourly budget (3/h fail-closed), restore the last snapshot, and tell the model the state was reset.

2. The stall watchdog cannot tell a healthy long cell from a deadlock. Its only activity source is the event stream; a cell doing await bash(...) or waiting on a subagent produces no stdout for 900s and gets killed mid-work (we had 144 production aborts, 100% of them with ipython in flight). We added an out-of-band kernel heartbeat (loop tick + live bash handles + in-flight host requests) that vouches the cell into the existing bounded snooze. True deadlocks still die — the budget is shared and capped, with regression tests proving both directions.

Both are on our line with tests; happy to upstream as PRs. Full list: https://github.com/Dmatut7/prime-agent-rlm/blob/merge/repl-kernel/docs/fork/upstream-feedback-20260911.md