---
name: uniform-zero-second-pass-including-the-must-skip-cell-is-env-that-never-reached-the-child
description: a two-host rig raw where EVERY cell PASSes in ~0.00s, including the cell that must SKIP on this box, means the rig env (SPT_TWO_HOST) never reached the test process — a vacuous PASS with Summary==1 that reads exactly like a witness
metadata:
  type: feedback
---

todlando 2026-09-07 11:10Z, W2 F17 xbox verify run: all FIVE role-A cells "finished in 0.00s"
PASS in ~0.013 s each — deny, fetch, helper, range AND role_b, the cell that is SUPPOSED to skip
on box A. Uniform skip = `Rig::from_env` returned None for every cell = SPT_TWO_HOST != "1" in the
child (twohost_web.rs:126, the first gate). Summary==1 on both raws (single, empty window, not
double-run). The DECISIVE corroboration was the owner side: `WEB_SERVE_FOR` appeared ZERO times in
b.raw — the same absence that was F17's evidence — so the run could not credit the fix; B looped
"A not ready" and timed out at nextest's 240 s cap.

**Why:** a nextest PASS line carries no elapsed threshold and no "did the rig arm" bit; a cell
that gated itself off before doing anything prints the same PASS as one that drove two daemons.
Summary==1 is necessary (not double-run), never sufficient (not empty).

**How to apply:** for every rig raw read (a) the cell that MUST skip on this box — if it reads
identically to the cells that must run, the env never reached the child; (b) the elapsed of the
cells that must run — a real round trip cannot finish in 0.00 s; (c) the OWNER-side breadcrumb the
cell exists to elicit (`WEB_SERVE_FOR`, `WEB_STREAM`), counted in the owner's raw. Put all three
in the driver's pair readout so the driver refuses to call the leg a witness when they read empty.
Measure the child env (psutil on the nextest pid), do not reason about the launch wrapper.
See [[a-source-read-cannot-witness-an-integration-cell]], [[a-battery-launched-from-a-perch-inherits-the-endpoint-identity]],
[[mutation-proof-does-not-validate-a-green]].

CAUSE FOUND 11:15Z (todlando): NOT an env bug — the rig script was REWRITTEN under his running bash (my arm D edit, 04:00:08 local, two minutes into his run); bash resumed mid-line past the env prefix. See [[editing-a-script-under-a-running-bash-mangles-the-live-run-into-vacuous-greens]].