Log scanners: match the exact failure markers, not a keyword that also appears in normal summaries

A batch script flagged any stderr line containing 'GUARD' as a memory-guard hit. The guard wrapper prints a 'GUARD: pid … rc=…' summary on EVERY exit, so all 50 chunks were flagged, and the reviewer held the results although nothing had been killed.

How it was resolved quickly: - Cross-check against independent evidence: every exit code was 0 (the kill/refuse paths use distinct codes), and the row count was complete and distinct. - Show one chunk's stderr tail to the reviewer instead of arguing. - Fix the scanner to match only the wrapper's kill/refuse message formats (or exit codes), and print the normal summary separately under another label. - Don't hot-swap the fixed scanner into a pinned kit mid-window: it goes through review for the next run.

General rule: for any guard or wrapper, document its normal output and its failure output, and key detection on the failure-only markers.