tests/test_watch.bats intermittently hangs an ubuntu shard until the 30-minute job cap

[fujibee]

Two ubuntu-latest shard 5/5 jobs were cancelled at the 30-minute cap while tests/test_watch.bats was running. Both PRs change nothing that test_watch.bats exercises for these tests, and a re-run of one of them passed.

run / job	PR	last passing test before the stall	stalled from
34909529861 / 104218499890	#1256	ok 5 session-end: leaves the store-owned read cursor intact (next: test 6, actas-mode ready sentinel)	01:33:51Z, cancelled 02:01:45Z
34922114263 / 104232434855	#1260	ok 29 watch: a failed pending-scan is surfaced and exits... (next: test 30, the watcher names this session's pane (#1044))	after 02:44:08Z file start, cancelled 03:12:07Z

• plain driver: an agent-native substitute for peek/poke with no pane #1256's whole-run re-run passed shard 5/5.
• The stalled test differs between the two runs, so this is not one test.
• In the first job's diagnostic process dump, a watch.sh sess-liveness.<pid> child was alive under the test with a sleep 600 sibling.
• Main's run on the 1.3.0 release commit (34911107475) passed.

Unlike #1107, the file does not progress: no test result line is printed for about 27 minutes.

Next step: capture per-test start lines and a process tree at the stall (the timed runner already prints file starts), then find which wait has no ceiling.