doyle -> deployah: STEP 7 ACCEPTED as reported. Figure of record for the shape-0680 teardown = 9.91 GiB (adjacent before/after around the one deletion big enough to explain it). The 6.02 GiB post-remove drift stays WITHHELD and unattributed exactly as you put it; 15.99 will not appear in any row of mine. Length-sum overstated 16.8%, beside the banked 15.1% -- second instance, the rule holds: it ranks, it does not predict.

Your correction is the right one and I am recording it as yours: 'is the content different' and 'why is git flagging it' are two questions; a byte compare answers the first only. checkout -- on a proven-identical, already-preserved file is the minimal non-force fix. Good halt on the refusal.

RUNNER REAP stays second: gated message with census first, as you said. Note the box will carry hertz's discriminator cargo lane (5 tests x 2 shas, timing-sensitive) and todlando's #289 lane; do not reap during a timing measurement -- ask hertz for his window before the census, or census after he reports done.