---
name: a-cmdline-scoped-builder-census-is-intermittently-blind
description: "A 'is my pool busy' census that filters cargo/rustc by worktree path in the CommandLine misses cargo entirely — one process carries a BLANK command line, another only the toolchain path — so it reads ALONE while three of your own builds run."
metadata:
  type: feedback
---

**2026-09-08, found by hertz on his box, reproduced on mine the same hour.** My driver's
pre-flight was `Get-CimInstance Win32_Process -Filter "Name='cargo.exe' or Name='rustc.exe'" |
Where-Object { $_.CommandLine -like '*<my-worktree>*' }`. Measured with three of my OWN build
processes live: box-wide count 3, predicate match **0**. One `cargo.exe` had a completely BLANK
`CommandLine`; another read
`"C:\Users\…\toolchains\…\cargo.exe" build -p spt-proto --bins` — the toolchain path, no
worktree anywhere in it.

**Why it is worse than a dead guard:** `rustc` invocations usually DO carry the target path
(`--out-dir …\target\debug\deps`), and `cargo` does not. So the same predicate fires correctly
mid-compile and reads empty seconds later, or during a cargo-only phase. It has a track record
of catching things, which is exactly what makes the empty readings persuasive. hertz's variant
was uniformly vacuous; mine was INTERMITTENT, and I quoted its empty output as evidence a run
was alone.

**How to apply:**
- Census **box-wide** over `cargo.exe`, `rustc.exe`, `clippy-driver.exe` and **label the field
  as box-wide, not lane-scoped**. On a shared box (golden runner + other lanes) the honest
  question is "is ANY build running", never "does one name my lane".
- A non-empty list REFUSES the run and says what it might be (the runner, another lane). That
  also makes a load-window hold MECHANICAL instead of remembered.
- Never let a process predicate carry an aloneness CLAIM without testing it against a known-live
  process first — start one of your own and check the predicate sees it.
- Corrections owed when this is found: every prior "procs-before: []" means "nothing matched a
  path predicate", not "the pool was mine". The runs' verdicts survive on their own evidence
  (exit files, one Summary, named cells, falsifiers); the aloneness guarantee does not.

Related: [[pool-scoped-sweep-blind-to-system32-leak]] (the mirror: reap BY command line, never by
image name — the two rules pull opposite ways because they answer different questions),
[[the-first-claim-on-a-foreign-pool-must-refuse]].