W31 self-improvement gap #2 (reviewer-confirmed): 125 GREENs shipped, only
41 ever reviewed. Most grinds are low-risk and fine to skip, but a GREEN
that escalated past the floor tier, had a regression along the way, or grew
past a minimal-fix size is exactly the kind of result likely to hide a real
problem in code nobody looks at again. grind()'s report() now emits a
machine-parseable "AUTO-REVIEW: yes (reasons)" marker on those GREENs;
serve() parses it and fires the same @reviewer ask a human would send,
reusing the existing wake-on-address mechanism (reviewer_lurker/CLAUDE.md)
rather than adding a new dispatch path.
marfrit
c6d5e72452