---
name: drain-side-instrument-suppresses-the-race-it-hunts
description: "An instrument that rides the racing thread can deterministically SUPPRESS the race — 21/21 green with tap vs 3/3 red without; instrument the LOSING side's consumer instead. Also: an environmental slowdown can be an accidental mitigation, so FIXING the environment exposes the race."
metadata: 
  node_type: memory
  type: project
  originSessionId: 0e501f1f-bba6-49e0-81cd-db288f425892
  modified: 2026-08-25T21:56:41.130Z
---

FIELD-SEAL #225 RCA, 2026-08-25. The resume-suite victim (seed needle = producer's FIRST chunk,
discarded by `spawn_session_pid`'s Spawned-wait loop when it races the Spawned reply, KH 6.9's
loop, same-session face) was **near-deterministic in BOTH directions** depending on the
instrument: `SPT_OBS_BYTE_TAP=1` (per-chunk `create_dir_all`+open+hex+write ON THE DRAIN THREAD)
= 21/21 GREEN; unset = 3/3 RED. The tap's per-chunk fs work delayed the first chunk ~the width
of the ~2ms race window.

**Why:** the race was chunk-arrival vs reply-write. Any probe attached to the arriving side adds
latency exactly where the race is decided. The winning instrument was on the OTHER side — an
eprintln in the brain's discard arm (test-process, off the drain thread) caught
`SPAWNWAIT_DISCARD: sid=1 seq=0 b="SXMARK01\r\n"` on a red run without perturbing the outcome.

**How to apply:**
- Before instrumenting a suspected race, name which thread LOSES it — put the probe on the
  consumer/loser side, never the producer/racer side.
- A 100%-flip discriminator (green with probe, red without, many trials) is itself EVIDENCE:
  it localizes the race to the probed thread's timeline. Report it as a finding, not a nuisance.
- Corollary measured the same day: an environmental slowdown can be an ACCIDENTAL MITIGATION.
  The kitsubito kernel-audit printk storm slowed the box like the tap did — the race hid for
  releases; FIXING the box (auditd consuming the storm) is what exposed it. After remediating
  an environment, expect latent knife-edge races to surface and re-run the victim family.
  Same family as [[correct-fix-deletes-accidental-mitigation]].
- Two-regime unification: a DISPATCH stall (audit backpressure) stretched the same ~2ms discard
  window to seconds → ANY marker prefix lost → random victims, long deaths, spawn-intensity
  scaling. One mechanism, two regimes — check whether a "random victim family" collapses to one
  window whose WIDTH varies with load before positing several mechanisms.