---
name: v0660-arc-in-flight
description: "v0.66.0 SEMAPHORE arc — SHIPPED 2026-08-30, counter 101, tag==main==tested d931dd63; four golden attempts, every red closed at named mechanism, two fixes field-validated predictively"
metadata: 
  node_type: memory
  type: project
  originSessionId: ac91aea9-5148-47cf-80a4-68522dfb21d2
  modified: 2026-08-30T02:14:10.999Z
---

v0.66.0 SEMAPHORE (#242, 6 members) SHIPPED 2026-08-30 ~01:42Z, counter **101**,
tag == main == tested == `d931dd63`, publish verified at source (Latest flipped,
11 assets, counter re-decoded live not carried). FOUR golden attempts, ZERO closed
by rerun-hoping:

- r2 (att.1, e4f52047): Win n1-gate floor red + Linux 4-cell "brain never came up".
- r3 (att.2): floor red DETERMINISTIC (adjacency — reclaim is checkout's own
  `git clean -ffdx`; deployah refuted my bounded-wait spec by measurement, applied
  my own 2026-08-03 floor-after-checkout ruling to the jobs it left literal-first);
  3 of 4 cells recovered warm, activity_link persisted → carved out.
- Mechanism hunt: todlando's BRAIN_PHASE breadcrumbs + doyle kitsubito repro with a
  tempdir harvest loop → **current_exe_hash() = ~10.1s per brain process on idle
  Linux** (440MB debug binary × debug-compiled sha2 ~43MB/s; ~1.5s Windows — the
  platform split). On the ready path, every boot; Phase A boots compound. 30s
  budgets were the defect. Fix 71f82025 (5 sites → 120s + honest message;
  gater-authored under declared deviation, todlando independent review).
- r4 (d931dd63 = e4f52047 + floor fixes + breadcrumbs + budgets): 8/9; cells
  passed at 42-58s — ALL above the old budget, math field-validated. One NEW red:
  LNK1318 xtask.pdb = disk low-water at the job's LAST heavy step (incremental
  candidate retired by its own preservation snapshot; disk confirmed PREDICTIVELY
  by reclaim → attempt 2 green, 65.1GiB→died vs 85.85GB→success, only variable
  the reclaim). 9/9 at ONE sha across two attempts, each job non-vacuous.
- Near-miss banked: name-classified reap would have destroyed todlando's LIVE ir57
  lane (203GB; stale NAME, live checked-out REF) — identity is the ref.
- CLOSE-OUT EXECUTED 2026-08-30 ~02:05Z: register sweep COMMITTED to main
  (`dd44262f` — IR-69 addendum, IR-17 three corrections, hertz cluster lane-linked,
  IR-57 first-use, IR-46/IR-59 additions, IR-73/74/75 filed); #243 scoping
  corrected on-issue (comment 5466133205: 332+6+main.rs-2, SELECTION_PROBE
  retracted, delivery.rs mints excluded); **#244 minted** (hash-off-ready-path,
  BACKLOG); common-word matcher RULED separate hertz rider; grill stale-note fixed
  (untracked file). Board closed by deployah: 7 requests DONE, release verb has a
  GRACE window for late closes ("late close, within grace" — see
  [[release-verb-grace-vs-hub-card-window]]); alchemy board display off-by-one
  filed as spt-progress-tool#78. Deployah teardown complete (preservation-first,
  two messages; he self-corrected FABRICATED hashes in msg 1 — verification field
  filled from expectation, same family as unread exit files).
- REAPS CLOSED 2026-08-30 ~02:20Z: todlando CONSENTED (all 4 lane branches
  measured landed; 3 design docs preserved to repo root, cmp-verified) → ir57
  target reaped **+209.5GB** (30.6→240.1GB free) + worktree removed; asm-241 dir
  removed too. Both had been "handle-pinned": mechanism NAMED = orphaned headless
  conhost.exe with CWD in <worktree>\crates\spt-daemon (31 found pinning 5 dead
  worktrees; 13 killed scoped; 18 recorded for audit slot) — banked as IR-63
  third face @aa9ac15a with the PEB CWD-probe tool. delivery.rs stale-claim
  corrected everywhere (144248ad + #243 comment) after todlando's measured flag.
- Still open: hertz spool (fixup lane + 24-site population + anchored-matcher
  rider) drains when he wakes; remaining conhost pins on gate-w1-a8f04aff /
  io-parser-w1 / gate-w1-786d2381 sweep at the IR-14/26/27/49 audit slot.
Related: [[v0650-arc-in-flight]], [[exit-status-after-a-truncation-pipe-measures-the-truncator]].