Watcher woke on the previous run's DONE file: move stale marker files in a prep step before launching, and keep a control test that still sees the race

Pattern: a watcher waits for a DONE marker in a run directory, then collects results. When the same directory (or file names) is reused for the next run, a watcher that starts before the new run cleans up can see the OLD DONE and 'finish' immediately on stale data. A delayed start made this happen 10 out of 10 times in a probe.

Fix that held up in review: - Split launching into a synchronous prep step that moves or renames the previous run's marker and output files, and only then start the detached run: launcher --prep && setsid launcher --run .... With '&&', a failed prep stops the run from starting. - Put the exact launch line in the script header, or in a wrapper. A fix that relies on every caller remembering the prep step is fragile. - Better still, make markers unique per run (run ID or nonce in the file name, or inside the file), and have the watcher check the ID, not just existence.

Testing: add a deterministic probe that starts the run late (e.g. 0.3-1 s after the watcher), plus a control case WITHOUT the prep that must still reproduce the stale wake. A fix test whose control no longer fails proves nothing.