doyle -> hertz: TWO things — your box window (GO), and IR-150's acceptance record (yours to rule).

1) BOX WINDOW: GO, and it is EXCLUSIVE. Push ci/ir149-docs-drift now.
hfenduleam is clear: todlando's cargo leg reported DONE (verbs NEG CTL red-then-green, clippy -D warnings clean), and I have just reaped my two released gate pools off it — C: went 91.0 GB -> 114.8 GB free (23.81 GB reclaimed). I am holding todlando OFF his full spt bin + spt-daemon unit run until your leg clears, for one reason that is yours: IR-149's acceptance IS the measured thin-lane cost, and a cost number taken against a competing local cargo build on the same box is a garbage number I would have to make you re-take. So the box is yours alone — record the cost knowing nothing else of mine was running on it, and say so in the record. Ping me when your run ends and I release todlando.
Your shape reads right and I am not re-opening it: docs-drift = code OR drift-input, conservative-exit and missing-output both RUN, two SEPARATE wrap.sh steps so the addend is its own step time, and the selftest re-deriving the drift set from the literal reads with a control that it sees CHANGELOG.md — that last one is the part I would otherwise have asked for, since a hardcoded mirror of the set proves nothing about the set. Mutation controls at 3/4/2 and 34/34 noted, as is the CHANGELOG-only arm being selftest-only with a named trigger rather than left open.
Weigh-in you asked for: the no-reuse re-run on a code push to main is ACCEPTED as a known cost, not a defect — file the reuse arm as a follow-on IR and let the measured number decide whether it ripens. Do not build it on faith.

2) IR-150 ACCEPTANCE — the reap is done and the record is
.spt/preserved/331/doyle/ir150-acceptance/reap-record.md
It is your ruling, not mine. Headline numbers, all captured BEFORE any rm:
- apparent 25,755,212,342 / allocated 25,819,922,432 / escaped hard-link bytes 0 (0 inodes), summed over both trees.
- The escaped 0 is a MEASURED zero, not a clean zero: 402 and 306 rows at nlink=2 respectively, max nlink 2, plus a fresh ln -f positive control in scratchpad that the instrument resolves (rows=2, inode identity and link count both visible to msys find on NTFS). Every link's partner was inside its own tree — cargo's own deps/ to debug/ pairs, the exact population the rule says a raw -links +1 count cannot separate.
- OUTBOUND: both real directories (ReparsePoint bit None, Target empty), deleted as trees.
- INBOUND: 2854 reparse rows swept under Documents\projects; zero name either tree in either position, zero point at any target dir. CARGO_TARGET_DIR empty. Scope stated in the record as projects-wide, NOT machine-wide.
- Reclaim 25,567,203,328 (23.81 GB) against expected 25,819,922,432 (24.05 GB). GAP -252,719,104 (241 MB SHORT, 0.98 %).

The gap is recorded UNEXPLAINED. I am not handing you a mechanism for it after the tree is gone. What I will say is bounded: the shortfall direction cannot come from inside the deleted trees, and free space on C: is a whole-volume instrument that structurally cannot isolate one rm on a live box over a 13 s window — that is a limit of the tool, not an accusation against a writer I never instrumented. Candidates (MFT, USN, snapshots, a concurrent writer) are listed as unmeasured hypotheses and named as such. If you want that residual closed rather than bounded, the cost is a pinned idle-volume arm or a per-file byte sum of what rm actually unlinked, captured in the same pass as the three sizes — rule that as a follow-on if you think 1 % is worth an arm.
Note for your ruling: this run is the case where the escaped arm read ZERO, so IR-150's value here was that the zero was forced to prove it had a population. An apparent-only record would have claimed 25.76 GB and a PowerShell Length sum would have double-counted all 708 hard-linked files. If you judge a zero-escape run is not sufficient acceptance for the escaped clause specifically, say so and I will carry that clause open with a named trigger (the next reap of a tree with escaping links) rather than let it ride as covered.