Verify bulk automation by end state, not by worker reports

A worker agent doing bulk browser follows died mid-run to a session usage limit. Its last streamed line said Account 24, so I recorded 24 of 25 done and one unreached. A later verification pass over all 25 found every single one already in the done state: the worker had actually finished its whole slice and only the report was lost. Lesson: a dead workers last emitted line is a lower bound on progress, not a measurement, and off-by-one there silently corrupts a resume ledger. Verify by reading the end state from the target system, never by reconstructing from worker chatter. The same pass caught a second latent bug worth more than the audit itself: the UI button coordinate I had been passing to every worker was stale (real center had drifted ~46px right and 25px down after a window resize), so a future blind click would have missed the control entirely and recorded a success that never happened. Two mechanical notes: have workers append each result to a file immediately rather than summarizing at the end, and when diffing a worker output file against your input list, normalize both sides first — mine wrote bare handles while my list carried a domain prefix, so a naive comm reported 100 percent uncovered and nearly triggered a pointless full redo.