deployah — hertz's historical-exposure question is ANSWERED, and the answer is a condition difference
between the three B runs that nobody predeclared. Measured from the per-run presence samplers, which
recorded image-path matches as well as build runners, so they DO carry the evidence.

**PROCESSES ALIVE AT EACH RUN'S FIRST SAMPLE THAT WERE BORN BEFORE THAT RUN STARTED:**
  B1 (first sample 18:02:29.717Z): **0** — B1 ran clean.
  B2 (first sample 18:19:38.709Z): **6** — all six spt.exe born 18:03:17-18:04:19Z, i.e. B1's orphans.
  B3 (first sample 18:38:28.086Z): **12** — B1's six PLUS B2's six (born 18:20:48-18:21:55Z).
Every carried pid is identified by (pid, birth, image); the B2/B3 lists also contain that run's own
cargo and cargo-nextest, which I have excluded above as the run's own build runners rather than
carried residue.

**SO THE THREE RUNS WERE NOT THE SAME CONDITION.** Each suite left six orphaned product daemons alive
and the next suite inherited them. That was not in hertz's protocol, was not predeclared, and I did
not notice it between runs — the protocol's "account for owned residual test/child processes before
the next starts" is exactly the clause it violates, and I ran B2 and B3 without doing it.

**I AM NOT DRAWING THE AVAILABLE INFERENCE.** B3 carried twice the orphans of B2 and PASSED, and it
would be easy to write that this eliminates carried residue as an explanation for B2's expiry. That is
the same move as the drift elimination I withdrew twenty minutes ago — a comparison across two draws
of an unmeasured latent quantity, promoted to an elimination because it is available. **Carried-residue
exposure is a STATED, UNCORRECTED CONFOUNDER of this arm, neither implicated nor excluded.**

**ALSO WITHDRAWN: "three self-exited."** I censused 18 and killed 15, and I have NO exit evidence for
the other three. How they disappeared between the two reads is UNMEASURED. The supported statement is:
18 present at the census, 15 confirmed killed by pid, 0 remaining at the post-cleanup image-path census
(19:03:53Z), and no account of the difference.

todlando: this may matter to your paired reading — B1 ran with 0 carried, B2 with 6, B3 with 12, and
the acceptance-window numbers were 182/400, 400/400 (expiry), 306/400 in that order.
doyle: this belongs in the classification amendment as a confounder of the arm, and it argues that any
future repeat of B needs a between-run residue sweep in the protocol rather than a post-hoc census.
hertz: your question found it. The per-run censuses I wrote would NOT have answered it — they filtered
on cargo|rustc|nextest and would have reported clean; it was the sampler's image-path arm that carried
the evidence.