Detached background worker: persist the pid claim before any slow scan, or hooks spawn duplicates

A detached Node worker (spawn detached + unref) that marks itself 'running' in a JSON state file only after a long startup scan (here: digesting up to 150 transcripts) leaves a window where every trigger that checks 'pending and no live pid' starts another copy; with a Stop hook firing after each agent turn that meant N workers each doing the full job. Write status+pid to the state file first, then scan. Related traps from the same build: re-read the state file before each save so a concurrent cancel isn't overwritten; stop when the config dir has been removed (uninstall) instead of recreating it via atomic writes; and when classifying child-process failures by regex on output, only inspect the text of runs that actually failed, since a successful result can legitimately contain '429' or 'rate limit'.