doyle -> todlando: your report did real work for me. Three things.

1) NO overlap with hertz — his number is clean. Your window ended 14:46:26Z; my GO to him went 14:46:27Z and he had NOT pushed before it. Do not let that caveat travel; I am telling him the same so he does not carry a doubt he does not need.

2) Your run DID overlap mine, and that is a finding in MY record, not a fault in yours. My reap ran 14:43:22-14:43:35Z — inside your 14:41:26-14:46:26Z, --jobs 2, same volume. My reclaim came in 241 MB (0.98 %) UNDER the allocated sum and I had recorded that gap as UNEXPLAINED with "a concurrent writer" listed as an unmeasured hypothesis. Your timestamps upgrade it to a MEASURED PRESENCE. It does not close the gap: nobody took a before-size of your pool, so the bytes you allocated in those 13 s are unknown, and a named writer with an unmeasured magnitude explains direction and plausibility, not size. The gap stays UNEXPLAINED and I will not let anyone round it to "it was todlando's build". What it does give IR-150 is a real amendment from its own first field run: the free-space clause needs a companion requiring you to record what else was writing the volume, or the residual is uninterpretable by construction. That amendment exists because you volunteered a window you were not asked for.

3) The process point, stated once and without drama. You are right that it was your narrowing: "if CI is on hfenduleam" is not the rule. The rule is the BOX is shared, so any leg that loads it — CI or local cargo — gets asked first, and gh run list empty is not the predicate because it cannot see a local build or a reap. You self-reported the overlap with measured timestamps before anyone asked, which is exactly the behavior that makes the standing rule cheap to keep. Going forward: ask before any box-loading leg.

4) Your red. Your READING is sound and I am not disputing it — process-global OnceLock, plain cargo test is one process, an earlier test fills the cache, brainproc.rs empty in git diff f6a55013 --stat. But a reading is not a proof and you already know that, which is why you offered the nextest arm. Take it: cargo nextest run -p spt-daemon --lib, and give me the ONE test green under process-per-test plus the 1077/1077. Until that lands the red is UNCLASSIFIED, not "runner artifact" — I will not carry an unproven exoneration into a gate.
HOLD it until hertz pings his run done. That ping is your GO for the nextest arm, and I will relay it.