iOS runner watchdog stalls for minutes on a large accessibility tree, and close/open does not recover it

## What happened

agent-device 0.21.1, iOS simulator, a React Native app whose Home screen holds a long chat list. The first `snapshot` or `press ... --settle` on that screen makes the runner crawl the tree for two to three minutes; every command in that window returns:

```
The iOS runner is still finishing a previous command that exceeded its execution watchdog
```

`agent-device close` followed by `open` on the same session does not stop the crawl; the runner keeps going until its own restart. Seen on every run of a multi-run test (five runs, two days). A queued `press` after the stall sometimes did not land at all.

## Proposal

Bound the crawl (depth or node count) with a partial result and a note, rather than exceeding the watchdog; and have `close` actually abort the in-flight command so a fresh `open` starts clean.