--- name: rig-inherits-the-defect-it-studies description: "A diagnostic rig reaches for the same broad-selector shortcut as the bug it investigates; on a shared dev/CI box that selector reaches CI's processes." metadata: node_type: memory type: feedback originSessionId: 1d37fd4a-b44f-4b62-b50e-c5470fd1c3c9 modified: 2026-08-03T05:39:27.615Z --- Hold a diagnostic rig to the SAME identity discipline you would demand of the product code it investigates. Building the positive control for IR-18 (a bare `taskkill /PID /F /T` fed by stale breadcrumbs) I wrote the rig's own killer as a machine-wide `Win32_Process Name='spt.exe' | where CommandLine -match 'daemon run'` → `Stop-Process -Force`. Caught before first execution; corrected to `ParentProcessId=` for the daemon and `ParentProcessId=` for its service. **Why:** hfenduleam is the Windows CI runner AND the dev box, so a machine-wide selector is not test-local — it matches CI's test daemons. Running it during a live leg force-kills another job's daemon and then reports the result as your own measurement: a fabricated red in someone else's run, produced by the instrument built to study fabricated reds. `CREATE_BREAKAWAY_FROM_JOB` does not weaken parent-id scoping — breakaway leaves the JOB, not the parent-id record, which is also how an orphaned child stays attributable to the daemon that spawned it. **How to apply:** in any rig that kills, scope by proven descent or captured identity, never by image name or cmdline match, and never machine-wide on a shared runner. Treat "my instrument used a bare selector" as evidence the pattern is easy to reach under pressure — argue the fix should make the authenticated path the only one in reach, not "be careful". Being blocked on the box is what bought the re-read: load discipline buys correctness catches, not just courtesy. Related: [[rig-ownership-reap-only-your-own]], [[reap-root-needs-authenticated-ancestry]], [[stale-breadcrumb-tree-kill-class]], [[live-agents-lock-target-debug-spt-exe]].