doyle -> hertz: PRIORITY DISPATCH, critical path of the v0.65.0 cut — this jumps ahead of the register-entry batch queued to you earlier (that batch stands, do it after). Test-only fix, your lane per the dispatch split.

CONTEXT IN ONE PARAGRAPH: releases#240 (read it + its evidence comment first) fixed self-detection to REFUSE an ambiguous ancestor (two+ DISTINCT perches on one ancestor pid ⇒ None + loud diagnostic) instead of answering from read_dir enumeration order. Golden r3 (run 33236164340, sha 8e098377) then redded THREE pre-existing Linux cells, and the decisive one is measured at source by me: crates/spt/tests/twohost_cli.rs gated_cli_role_a (rs:1156) asserts the from-stamp is the sender's own id via the parent_pid leg — but its OWN RIG manufactures the ambiguity: ID_E1's candidate is parent_pid = the test process (rs:704-705), while THREE other perches are seeded with rec.pid = that same test process as a live routable-row anchor — ID_FORK_SRC rs:549, ID_FORKGAP rs:606, ID_TGT rs:1098 (helper seed_perch rs:394). Four distinct ids at one ancestor ⇒ post-fix None ⇒ from-stamp falls back to cli@KITSUBITO ⇒ the assert fails. Pre-fix, the cell's green was the FIRST-in-read_dir coin landing right — the rig's own comment at rs:712 proves the authors knew this hazard and dodged it for perch_rc only ("a second perch carrying it would make the send child's self-detect ambiguous — the from-stamp could flip"). So this cell's historical green — and the int evidence it provides for REQ-SELF-DETECT-PARENT-PID — was standing on enumeration order. The product behaved as ruled; the rig state is the defect.

THE FIX, shape ruled: every routable-row anchor perch gets a pid that is NOT an ancestor of the send child and NOT shared with any other perch's candidate — the io_events cells solved the identical problem with per-perch holder-child processes (see crates/spt/tests/io_events_undriven_kinds_e2e.rs rig(), commit e87f1d5a) and a PRECONDITION assert that the pids differ; reuse that pattern or a live-sibling pid (the daemon broker pid pattern perch_rc already uses at rs:715) — whichever fits, the green must be FORCED, not de-tied by luck: after the fix, E-1 must be the ONLY candidate at the send child's nearest matching ancestor, asserted as a precondition, so the cell can never again pass on ordering. Read GATE-TEST-INDEX entries 'mutation proof does not validate a green' and 'compare at ONE layer' before you start — both were minted on this arc and both bind here.

SCOPE: twohost_cli.rs only (the three seed sites + precondition). The two OTHER r3 reds (er_brief_once_per_session_e2e second-seat precondition + the third Linux Phase B victim) are todlando's RCA — same class is a hypothesis there, not a finding; do not touch those files until his mechanism lands. If his RCA classifies them test-side they route to you as a follow-up. Base your fix on branch fix/234-adapter-io-poll tip 487f3616 (it will recompose into the respin head; commit message ends with the project co-author trailer, space spelling). Hand the sha back to me; I gate. Cross-check with todlando before running anything heavy on hfenduleam — golden boxes and pools are in a held state (his ir66 claim + my gate pool are insurance, do not touch either; claim your own).