a_session_without_a_worktree_costs_no_probe waited up to 5s for the
watcher pass to run, using that wait's timeout panic as an implicit
proxy for "no probe happened". Under a loaded full-workspace run the
watcher task can be scheduled well past 5s, so the test flaked.
Poll the shim log directly for A's probe instead of a derived snapshot
state, and use the wait only as a 30s hang guard. The actual property
- B was never probed - is asserted unconditionally on the log content,
independent of timing.
Fixes#173.