[Other] managed runtime host husks forever after its server child exits — no child watcher, no reap, no self-heal; profile stays locked until manual kill (0.1.49.2, verified present in 0.1.51) #4574

Area

Terminal — managed runtime host (strata runtime start, strata-control).

Desktop app, Terminal or source?

Terminal (managed runtime host on POSIX).

What happened

A managed runtime host supervises its server child not at all. If the child exits for any reason — clean or not — the host process keeps running as a husk: it stays in its accept loop, keeps the state lock and the control socket, and the profile is unusable until a human kills the host PID.

Code path (strata-control 0.1.49.2, strata_control/runtime.py):

• Host.serve() is an accept loop with a 1 s timeout. There is no join(), no waitpid(), no SIGCHLD handling, no poll of self._child anywhere in the host loop.
• _status() correctly reports the child as failed (it polls), but nothing consumes that: no restart, no exit, no log line on child death.
• --shutdown-timeout only bounds _stop_child() on deliberate stop/restart paths; it never applies to an unexpected child death.

Verified still present in strata-control 0.1.51 (code read 2026-10-28): serve() accept loop never polls the child; reap happens only in the finally/signal paths.

Impact: any local client (agent harness, benchmark, GUI) sees a wedged runtime with a live-looking lock and no self-heal; recovery requires knowing to kill a PID that status does not point at as the problem.

Smallest steps or command that reproduce it

1. strata runtime start <profile>; note the host PID and the server child PID (strata runtime status).
2. kill <server-child-pid> (any signal; also observed with the child exiting cleanly on its own).
3. Observe: the host process is still alive, strata runtime status reports failed, every strata runtime start/launch for the profile refuses with "this profile already has a live managed runtime" (exit 6), and the HTTP port stays dead. The husk never exits on its own; only kill <host-pid> + strata runtime start recovers.

Expected: the host should watch its child — on child exit, log it and either restart (bounded, like the engine auto-restart in serve, issue #27) or tear down cleanly (close owner pipe, release state lock, exit non-zero) so the next runtime start recovers without manual kill.

Reproduction status

Observed in the field (one real occurrence, 2026-10-27 ~11:02 local; not a controlled re-run). The code path was additionally confirmed by reading 0.1.49.2 and 0.1.51. A controlled kill <child> repro would confirm the full matrix but was not run.

Strata version

Bundle 0.1.49.2 (linux-x86_64, python 3.12.15); strata-control 0.1.49.2, code re-read at 0.1.51.

Graphics card, memory and system

Arch Linux x86_64, single GPU with 24 GiB VRAM, 64 GiB DRAM.

Model, size and settings

Profile launch-iq3_s-128k (IQ3_S quantization, 128k context), single profile.

Logs

Watchdog blocks from the 2026-10-27 incident recorded: strata runtime status = child failed while the host PID stayed alive, and repeated strata runtime start attempts returning exit 6 ("this profile already has a live managed runtime"). The in-flight generation was severed at the transport level when the child exited. Recovery was kill <host-pid> + runtime start. Sanitized excerpt available on request.