---
name: golden-twohost-is-cross-os-and-a-count-is-not-an-owner
description: "2026-09-08: golden's twohost pair is CROSS-OS — role A on hfenduleam, role B on kitsubito — so 'the golden twohost job is Windows-only' is false and a Linux box you believe uninvolved can be hosting half of it. Caught because I identified a lone cargo by cwd instead of counting it."
metadata:
  node_type: memory
  type: reference
---

**THE FACT, measured 2026-09-08 20:56Z and worth more than the story:** golden's twohost pair runs
ACROSS BOTH BOXES — role A on hfenduleam, role B on kitsubito, `cargo test -p spt-daemon --test
twohost_web` under `.github/bench/wrap.sh`. doyle had written "golden twohost is Windows-only" from
belief rather than a read, and corrected himself the same minute. So a golden run occupies BOTH
boxes, and neither is free merely because the other is where you think the job lives.

**THE HABIT THAT CAUGHT IT:** my pre-launch census read `cargo=1, rustc=0, load 0.78` — which looks
like a nearly-idle box with one stale process, and I nearly took it as one. Identifying it instead
gave `pid 1850551 etime=07:16 cwd=/home/reavus/actions-runner/_work/spt-bs-core/spt-bs-core` — a
RUNNER cwd, i.e. live CI. I was one `scp` from shipping a bundle and running a battery into golden
twohost-b.

**Why:** a COUNT tells you occupancy, never ownership, and a low count is the most misleading kind —
one process reads as leftover, and leftovers feel safe to ignore. `readlink /proc/<pid>/cwd` costs
one command and answers the question the count cannot. On this project an `actions-runner/_work/...`
cwd means CI, and the ff-push concluding a golden also starts a post-merge run on that same box.

**How to apply:** before any battery, IDENTIFY every builder by cwd (and etime), do not count them.
Refuse on any you cannot attribute — `UNREADABLE` is its own value, not an absence. And never infer
a box is free from which OS you think a job belongs to; read the job's runner_name.
Related: [[a-single-idle-sample-of-a-busy-box-is-a-lull]] (sampling over time) — this one is
attribution of a single sample. Both are needed; neither substitutes.