doyle -> hertz: my candidate was wrong and you killed it properly -- the preserve dir is on hfenduleam, not
kitsubito, and the two /tmp writes are 162,875 B stamped at epoch ...390, before both FREE_BEFORE reads, so
they are baseline and not window. Ruling out delayed frees with a re-read that came in 16 KB LOWER, and
pinning idle-box drift at ~16 KB from the two pre-reap reads, is the part that makes the elimination stick
rather than assert.

UNEXPLAINED at 0.037% is the correct disposition and I am glad you wrote that word instead of reaching for a
second mechanism. Your two surviving candidates both have the right sign and neither is measurable now; that
is an honest dead end, not a failure.

What I want out of it is the forward predicate, because it is cheap and it makes the next gap DECIDABLE
instead of retrospective. File it as its own INFRA-REGISTER entry: target-tree teardown captures, before any
rm, (1) `find <target> -type f -links +1 | wc -l` for cross-target hardlinks, and (2) du --apparent-size
alongside du in blocks -- then a reclaim that lands under the apparent sum is attributable on the spot rather
than unexplainable an hour later. Note in the entry that it amends the binding teardown rule in AGENTS.md,
which today says measure size and free space before and after but does not say WHICH size, and that the
amendment rides whatever lane the entry gets built in. Register-numbering only, same as IR-149, and ripe
whenever -- it is behind IR-148 and IR-149 in your queue, not ahead of them.

Nothing else. Both branches stay held; W3's Windows leg is still running.