When SIGUSR1 fires (task timeout), resmoke was setting `HANG_ANALYZER_CALLED` before collecting subprocess PIDs. Other threads observe that flag and can tear down fixtures (killing mongod/mongo processes) before `_get_pids()` runs, causing `test_analysis.txt` to contain fewer PIDs than expected.

Fix: snapshot subprocess PIDs at the very start of both signal handlers, before `HANG_ANALYZER_CALLED` is set, and pass the snapshot forward to `_analyze_pids`.

Also change `test_analysis()` to open `test_analysis.txt` in append mode (`"a"`) instead of write mode (`"w"`), so concurrent callers from the sighandler and per-test `on_timeout` paths don't clobber each other's output.