---
name: a2-golden-r2-specimens
description: "Two batch-inert specimens from golden a2-testfix-r2 run 30607903133 (2026-07-31), queued to hertz's ledger by doyle — torn-read advisory poll + kitsubito brain-bringup starvation."
metadata: 
  node_type: memory
  type: project
  originSessionId: cc1be5ee-a7e1-4b7b-861d-4c0a98570a34
  modified: 2026-07-31T10:14:49.438Z
---

Golden `a2-testfix-r2` (@edce86a off d7411f9, run 30607903133, 2026-07-31): two failures, both RCA'd **batch-inert** by doyle; batch reran same-sha. No immediate action owed; both queued for hertz.

**Outcome: attempt 2 (same-sha rerun) FULL GREEN** — all 9 jobs completed+success, run conclusion success. Flake reading HELD. **BATCH CLOSED 2026-07-31: doyle ffed main d7411f9 → edce86a, PRs #130–134 merged, gate worktrees reaped; PHASE-B-RESULTS.md docs commit rode after @dd1653e.** Same-sha green closes the trigger, NOT the class (see [[instrument-soundness-guards]]): the torn-read poll fix is still owed, and the kitsubito starvation is still HEAVY-class-if-recurs.

1. **[FIX BUILT 2026-07-31: hertz `fix/advisory-poll-soundness` @1488e39 off 0f90431, test-only — one-predicate poll riding the assert's read path + temp+rename fixture publish; reader untouched (fragment tolerance is pinned semantics). Rides the NEXT golden batch, doyle gates at assembly. Hertz's repro also minted a NEW load-class specimen: `a_wedged_service_is_force_killed_with_the_descendant_it_left_behind` :305 grandchild survived force-kill once in a cold-compile 18s run — single occurrence, baseline-clean, kin of specimen 2, watch-don't-act.]** **Windows, instrument defect (torn-read class — see [[instrument-soundness-guards]]):** `spt-daemon::servicehost_supervision_e2e::the_published_stop_marker_earns_a_cooperative_exit_inside_grace`, assert at `:223` — advisory read returned `None` vs `Some("mock advisory: serving")`. Single-shot read raced the mock's advisory write (file/line not yet present). File untouched by batch; latent unsound poll surfacing under leg load. **FIXED 2026-07-31** — lane `fix/advisory-poll-soundness` @1488e39 off main 0f90431. Cause: poll waited on `status-advisory` **existing**, assert consumed the **read**; fixture's plain `fs::write` makes the name appear before its content. Fix is test-side only: poll rides `status_registered → advisory.is_some()` (one predicate), fixture publishes by temp+rename (plain write kept as fallback — a Windows rename can lose to a reader holding the destination open). **Core's `read_status_advisory` deliberately NOT touched** — servicehost.rs:3645 requires a 50k unterminated line to still read capped, so fragment-tolerance is stated semantics and completeness must come from the publish side. Reproduced by widening the create→content window (old shape fails with the exact golden signature, fixed passes).

3. **Windows, #47 contamination class (lane-standalone only):** lane CI run 30607581061 (`fix/authenticated-test-reapers` @37e8276) — `spt-daemon psyrelay::tests::notify_reaches_own_user_only`, psyrelay.rs:382, got `notif_id "cafe:2"` vs expected `"cafe:1"`: foreign mint landed in the with_home window under threaded cargo test (see [[unit-lane-inprocess-isolation]]). Lane branch predates 57b3fcc nextest conversion; r2 merged tree has it → batch-inert. Doyle-ruled ignore/no-rebase: under ADR-0050 the golden head is the tested sha, and a lane-standalone red from the very class its batch-mate fixes is evidence FOR the batch. Ledgered by doyle as #4 of the class.

2. **Linux, load class:** `spt::activity_link_push_e2e` brain-bringup PRECONDITION panic (`:205`) after full 30s; brain stderr EMPTY (broker spawned, brain never emitted), at test 795/2169 mid-leg — spawn starvation on kitsubito under default parallelism. Failing path upstream of the batch's reaper diff in that file. Single occurrence — **not acting**; if recurs: HEAVY-class it or bump bringup deadline.

4. **Windows, tree-kill under load (NEW, unresolved):** `a_wedged_service_is_force_killed_with_the_descendant_it_left_behind` failed once at `:305` — grandchild survived the force-kill (pid 45212) — in the single run that took 18.02s vs 4–6s for every other run of the same file, i.e. right after a cold compile under load. 3 later runs with the fix green, baseline on clean main green, Linux green. Same family as #2 (load class); **not** caused by the advisory lane. Unreproduced; recorded, not swallowed.