fix(spawner): reap terminal task containers instead of waiting for a nonexistent exit

[dkrattiger]

Problem

When a task reaches a terminal state (COMPLETE/DROPPED), its Docker container is never removed, so containers leak. We found 18 orphans on one host, causing significant CPU/memory pressure and typing lag in live task panes.

Root cause chain:

• Spawner.cleanup only removed the per-task workspace, and only once self._runner_for(task).is_running(...) already read False — it waited for the container to "exit naturally". It never removed the container.
• But the container never exits on its own: the liveness loop while running(): in container/entrypoint.py uses _until_signalled(), which flips only on SIGTERM/SIGINT and never checks task state.
• Nothing called the runner's stop() (LocalRunner.stop → docker rm --force + tmux kill-session) on a terminal transition.

So cleanup waited forever for an exit that never came; the container ran indefinitely.

Fix

Spawner.cleanup now, when a task is terminal and its container is still running, actively stops+removes it via the runner's stop() before releasing the claim and cleaning the workspace. It runs for every task on every host pass, so:

• terminal + running → stop() (docker rm --force + tmux kill-session, idempotent), then release claim, then remove workspace.
• terminal + already gone → stop() skipped (safe no-op), workspace still cleaned — unchanged.
• non-terminal → untouched (early return): container never touched, workspace never deleted.

The cleanup docstring is updated — the "wait for it to exit naturally" assumption was wrong.

Plan: see the plan.md task artifact.

Follow-up (out of scope here)

Container-side self-exit on terminal state ("option 1") is a complementary improvement: have running() in container/entrypoint.py also return False once the task is terminal, and/or run the container with docker run --rm so it self-removes. It fails differently from the host-side reap (defense-in-depth), but carries real trade-offs — it needs the container to poll task state (new REST traffic in the one LLM-bearing package), docker run --rm would remove crashed containers too (hiding docker logs the crash-loop/heal machinery relies on), and immediate self-exit removes the post-completion grace window for attaching to a finished task's pane. Worth its own design discussion rather than bundling here. The host-side reap in this PR robustly closes the leak on its own — including for wedged containers a self-exit path couldn't reach.