---
name: duplicate-work-population-is-processes-not-code-paths
description: "When something was served twice, enumerate the PROCESS population before the code paths — my 'two loops in one process' hypothesis was right about the effect and wrong about the locus, and only the pid field showed it."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 1e15e6ca-6830-40c8-b051-25a9b1850b5c
  modified: 2026-08-19T00:35:04.558Z
---

A duplicated serve invites a code-path story (a loop re-entered, a latch that
does not cover a second pass). Ask FIRST which processes were running the
dispatcher at all: a per-process claim/ownership state cannot exclude a SIBLING
PROCESS, and no amount of reading one process's code shows the second one.

Measured (releases#181 layer 3, run 32200175631, reproduced 2/2): one redeem
stream was served by TWO processes on one home 21 ms apart — a livehost `spt`
(147 serves that run, six families) and the test-hosted daemon (670) — both from
`floor_seq=0` on a row already `finished=true`. The winner consumed the code and
notified; the loser refused at claim time and its reply reached the caller 88 ms
FIRST. My pre-registered hypothesis said "doubled feed loop in one process": the
effect was right, the locus wrong, and the `proc_pid`/`exe` stamp on every diag
line is the only reason the correction was possible in one run.

Two riders worth keeping:
- **Name the arm you predict, and score it honestly.** I nominated the
  consume-time refusal as the candidate; the real one was the CLAIM-time refusal
  (the winner had already consumed). Same class, different site — and the fix's
  guard belongs at the site that actually fired.
- **The visible family is not the affected population.** Redeem surfaced this
  only because its store is exactly-once enough to refuse the second serve. Sync,
  Notif, Registry, Update and Attach took the same double-serve in the same run
  and said nothing. Report the silent families beside the loud one — it is what
  decides the size of the fix.

**How to apply:** stamp pid + exe on every diagnostic line (the pump188 sink
already does). Before theorizing about re-entrancy, count the distinct pids
emitting the instrumented event and list what each one served. Kin:
[[read-the-other-legs-GREENS-not-just-its-red]],
[[anti-oracle-refusal-is-opaque-to-the-test-too]], [[two-daemon-one-box-rig]].