# Paired-sequence arm — shared-machine-state candidate for the golden-v2 Windows red

Status 2026-08-19: PRE-REGISTERED, NOT RUN. Fires only on doyle's release of the hfenduleam
no-cargo hold, and only if his isolated single-cell x20 (rig `rca-182-v2` @`bd942f6`) comes back
0/20 — an isolated 0/20 is what makes isolation the wrong frame and pairing the next question.

## What this arm exists to discriminate

Golden v2 (run 32243884768) failed ONE cell on Windows only:
`engine_room_bringup_e2e::a_cleanly_offline_engine_room_is_still_brought_up_by_its_own_gate`,
FAIL 61.448s at `engine_room_bringup_e2e.rs:1213`, rc stdout `the code was accepted, but the engine
room's harness (erhost) did not come up within 30s — nothing is attached.` The SAME cell PASSED on
Linux in the SAME run (21.468s), so the red is Windows-only in this observation.

doyle's x20 is single-cell and isolated. That design is structurally blind to the candidate this
arm tests: the failing cell ran at 16/193 immediately after my W3 cell at 15/193
(`a_brought_up_engine_room_is_served_the_in_core_role_and_has_no_writer`, PASS 2.883s), and both
bring an engine room up on ONE box. If bring-up leaves machine-scoped residue — a home root, a
socket/pipe name, a port, an erhost/adapter process outliving its cell — then the failing cell's
30s attach bound is being spent on someone else's leftovers, and no isolated repeat can see it.

## Arms (registered BEFORE any data)

- **A (paired):** my cell, then the failing cell, back to back in ONE nextest invocation, x20,
  **`--test-threads=1`** (doyle's amendment 2026-08-19, accepted before first execution). Default
  parallelism would run the two cells CONCURRENTLY — two rooms up at once — which is a THIRD
  condition, not the golden's. The golden's Phase B is the serialized heavy class: cell 15
  concluded before 16 started, and `-j1` in list order reproduces that sequence.
- **B (solo control):** the failing cell alone, same rig, same env, x20, **also `--test-threads=1`**
  for symmetry, so A-vs-B differs only in the presence of the paired cell — not in scheduling.
- **Order is verified, not assumed:** `-j1` runs cells in nextest's list order, which is expected
  to put `a_brought_up…` before `a_cleanly…`. I read that order off `nextest list` and assert it
  before the first execution; if the list does not put my cell first, arm A is not the golden's
  sequence and I say so rather than run it anyway.
- Both arms run in the SAME session, interleaved A,B,A,B… so box load drifts across both equally.
  Running all 20 A then all 20 B would confound the pairing with the hour.

## The statistic, pre-registered

- Primary: **failure count out of 20 per arm**, reported as the raw vector, not a rate alone.
- Secondary and the one that actually informs: **erhost attach time per execution**, collected from
  the cell's own wall time (nextest per-cell duration) — a VECTOR plus histogram, because a red at
  the 30s bound and a green at 29s are the same mechanism and only the vector shows it.
- Conservative reading registered up front: if A and B both come back 0/20, this arm has REFUTED
  the shared-state candidate at n=20 and says NOTHING about load; a green here is not a flake
  verdict, and I will say so rather than let 0/20 read as "it was load".
- Non-vacuity: the run is only informative if BOTH cells actually executed every iteration —
  I collect the executed-cell count per iteration and report it beside the failures. A filter that
  silently matched one cell yields 20 clean greens that mean nothing (zero-match filter trap).

## Procedure (exact, so it is not re-derived under pressure)

1. Fire only on doyle's explicit release. Confirm no golden is live on hfenduleam first.
2. Own rig under `.worktrees/`, own pool, claimed with `pool-claim` at lane start from the rig's
   own cwd. Never build into another lane's `target/`.
3. Scrubbed env every leg: `env -u OWL_SESSION_ID -u SPT_AGENT_ID -u SPT_ENDPOINT_ID`.
4. Positive-control the filter BEFORE running it: `cargo nextest list -E …`, read the NAMES, assert
   both cells are in the list.
5. Report the vectors, the executed counts, and which ARM fired — never a color.

## Standing constraints

- Box is doyle's until he releases it; no cargo on hfenduleam before then.
- The failing cell is NOT from my lanes (authored at `a1b9947`, already an ancestor of main). This
  arm tests a candidate mechanism I disclosed, not a defect I own.
- My W3 int cell is DISCHARGED: first execution in golden v2, PASS on both boxes (2.883s Windows,
  10.693s Linux), no rerun involved.