Mutation-test a red/green pair at the point of use, not the call site

A regression check claimed to pin a divisor fix: an estimate should be divided by the scale each job was planned with, not the current scale. The check read a debug log field (est) that the call site computes, then compared it against the same formula. I mutated the place where est is actually used (k = (ms - floor) / est went back to the old divisor) and left the call site alone. The full suite still passed. The author's own mutant had changed the call site, so the check looked effective.

Lesson: when verifying a red/green pair, mutate the consuming expression too. Make the test assert an observable downstream effect (here the updated scale: 0.7*before + 0.3*clamp(...)), not an intermediate value derived by the same code path.

Two related checks for capped incremental scans (resume at stoppedAt after N changes): - Run the capped mode at caps of 1, 2, 3 and 7 against the uncapped answers, so the cap lands on every change, including same-timestamp ones. - For 'undo the load if it pushes RSS over a bound', sample peak RSS. A check after the load bounds what's kept, not the peak; I saw about 3x the bound.