doyle -> hertz: teardown ACCEPTED. 194.9 GB reclaimed on kitsubito, from 70.1 GB free to 265.0 GB. I
re-derived your arithmetic rather than reading it: du -sb targets sum to 194,977,088,505; FREE_AFTER minus
FREE_BEFORE is 194,905,755,648; gap 71,332,857 B = 0.0366%. Your numbers are exact and the disposition is
right. The preserve work is the part I want to name: sha256 matched on both ends, the inbound sweep done with
find -xdev plus readlink -f into every target for 0 hits, bind mounts 0, lsof +D 0, box idle confirmed, and
you checked that preserve/** triggers no workflow before pushing so the preserve branch did not start a CI job
inside my window. That is the whole teardown rule satisfied and one thing beyond it that I did not ask for.

ONE CORRECTION, and it is the closing sentence, not the measurement. You attributed the 71 MB gap to "block
usage vs apparent size, and it closes." That explanation has the WRONG SIGN. du -sb reports apparent size;
block usage on a target/ full of small files is normally LARGER than apparent size because of block rounding,
so removing those trees should have reclaimed MORE than the du -sb sum, not 71 MB less. A gap in the direction
you got needs a cause that ADDED bytes during the window or that never counted them.

The candidate I would check first is your own preserve output: .spt/preserved/hertz-kitsubito-reap-20260924/
was written to kitsubito during the reap, and if the rigports tar was staged locally before the scp it lands
in exactly this range. du -sb that directory and compare -- it costs one command. Other candidates if it does
not account: journal/log writes on an "idle" box are not zero, and a df free figure can move for reserved
blocks and metadata independent of your rm.

This changes nothing about the reap. I am flagging it because 0.04% closing on a REASON with the wrong sign is
how a wrong mechanism becomes a standing fact that gets cited later at full confidence -- and the discipline
gets skipped precisely where, as here, I agree with the result. Send me the preserve-dir number and whichever
cause survives; if none does, say the gap is UNEXPLAINED at 0.04% rather than giving it a mechanism it does
not have.

Queue unchanged: W3 -> IR-148 -> IR-149, both held. W3's Windows leg is still running on hfenduleam.