---
name: v0670-arc-in-flight
description: "v0.67.0 NOW-SIGNAL arc — SHIPPED 2026-08-30 08:36Z, counter 102, tag==main-at-tag==tested da71b785; greenlight-to-ship in ONE session (~7h); cascade-before-publish executed; daemon restart owed to a non-endpoint actor"
metadata: 
  node_type: memory
  type: project
  originSessionId: 9298ce90-27bd-44cb-9d08-f4acba5623a0
  modified: 2026-08-30T08:42:14.414Z
---

**SHIPPED 2026-08-30T08:36:37Z** — counter 102, tag == main-at-tag == tested `da71b785`,
publish verified at source (deployah); cascade 18 min PRE-publish — the v0.66.0 card
lesson EXECUTED (#242 still surfaces on this card per spt-progress-tool#78: predicted,
filed, never a state repair). Golden: att1 sole red = FLAKE-LEDGER :664 teardown-leak,
ruled ledgered-class (signature predates the diff), att2 9/9 with cell re-execution
proven; ledger +1 OPEN, hertz cluster strengthened. Close sweep `4f0c07b6` (IR-59
SEVENTH face: per-wave pool reaping — four wave pools = 236.6GB/3.5h drove the box to
0.01GB mid-close, emergency reap +186.6GB; IR-12 stale-binary kin note). Reap party
complete (+58.2GB, 14 conhost pins killed FIRST, four trees removed first-try).
Node binary 0.67.0 applied; **DAEMON RESTART OWED to a non-endpoint actor** (guard
refuses from hosted sessions; operator flips `spt node stop/start` when convenient —
daemon-coordinated features run 0.66.0 until then). Field validation: my own transcript
received UPDATES + SEAL_BRIEF minutes post-ship. Open elsewhere: IR-73 floor reorders
(next workflow touch); hertz spool; operator queue #244/#246/#247/#254 + #16/#17 EVAL +
#11c flag question.

— arc history below —

NOW-SIGNAL #23 (GREENLIT, 10 open members) intake EXECUTED 2026-08-30 ~02:18Z at base
`aa9ac15a`; wave map `NOW-SIGNAL-23-JIT.md` (repo root) + board comment 5466200195.

- Census MEASURED: v1 core DELIVERED whole (nowsignal.rs — six gathers, SeenSet delta,
  NowSpec; #11/#14 are verify-and-close, NOT build lanes). Unbuilt categories:
  SEAL_BRIEF (#18), UPDATES (#245) — zero source hits.
- Waves: W0 todlando (verify #11/#14 with file:line evidence; #16/#17 precondition
  measurements — PROJECTS may be unblocked by REQ-GOSSIP-ADAPTER-PROJECTS; #113
  echo-commune legacy cadence DELTA report, measure-first, fork points to doyle).
  W1 #245+#18; W2 #44 channel→#45 alerts; W3 #133 shell [hints] (+#16 iff ripe).
  #16/#17 not-ripe ⇒ STOP-AND-REFER to operator before close, never silent drop.
- Riders composed in: IR-59 log-the-floor + IR-57 scripted audit tooling pair; IR-73
  floor reorders on first workflow commit; IR-14/26/27/49 audit slot = intake
  parallel-fill (OPEN, window now); IR-11/IR-33(a) opportunistic.
- **W1 GATED + LANDED 2026-08-30 ~03:30Z**: main ff 343df76e → `31020e02` (pick-audit
  1/1 MATCH; clippy-ws/treqs/xtask/bins 748/748 all 0 by exit file). #245 + #18 →
  ACCEPTANCE (github closed). One gate-rig red = MY missing fixture-prebuild leg
  (⭐⭐⭐ trap, 4th recurrence, banked as 4th face on open-the-subindex memory;
  prebuild leg now mandatory in every driver I author). Gate rig gate-ns23-w1 +
  pool HELD WARM for W2's gate, claim released for sequential takeover.
- **#113 GATED + LANDED 2026-08-30 ~04:25Z**: main ff 31020e02 → `aa5f80f6`
  (pick-audit 1/1 MATCH; all legs 0 by exit file, daemon 1141/1141, live 84/84).
  #113 → ACCEPTANCE, premise-correction record on the ticket. **W2 building**
  (feat/ns23-w2-attachment-channel off 343df76e, pool ns23-w2 from reused ns23-w1
  worktree): node-list RULED IN-SCOPE (datum reaches broker and is dropped —
  ViewerSink retains in-memory ONLY, durable triple unchanged, IR-72 kinship
  noted); shape RULED broker-push primary + thin link-edge observer (identity
  never on disk; both IPC constraints inherited from activity_write, no new
  bound); #45 away text = verbatim ticket constant (pasted from intake fetch),
  re-attach notice builder-authored short. Gate rig + pool warm for W2 gate.
- **#113 lane history (landed)** (.worktrees/ns23-113, feat/113-echo-commune-cadence off
  343df76e, own pool): THREE ticket premises measured false (todlando stop-refers,
  all ruled): fork-1 inverted — idle arm already arms echo UNCONDITIONALLY
  (docs api.md:296), missing piece = AGE GATE; RULED split sentinel by arm (two
  files: edge ungated, idle 15-min age-gated w/ legacy mechanics); fork-2 phantom —
  NO mirror-into-context exists, fix = re-kind echo_mirror→echo_commune + comment
  correction + ADR-0019 vocabulary amended by replacement + REQ-TERM-7 retag;
  fork-3 = input-before-rotation constraint (no boundary echo exists today, no
  watermark — whole-session input stands, no watermark this phase). Ticket's
  "spt-core currently surfaces echo to running context" parenthetical measured
  FALSE — comment on #113 at close.
- W1 MID-FLIGHT state (todlando communed across ~02:38Z, lane state
  .worktrees/ns23-w1/NS23-W1-JIT.md, nothing committed, pool claimed from lane
  worktree): categories on enum appended at END (ratified render order untouched),
  UPDATES per-shell-keyed, SEAL_BRIEF constant (his catch: first text cited
  nonexistent `spt seal verify`; real surface = `spt api seal verify <token>`,
  REQ title locks the correction); remaining = narrowed #11b rider (render only:
  subject named, pubkey fallback labeled — from-node half is an UNBUILT
  endpoint-level data source, minted #247, rides #11 close as exception (d)) +
  doc/unit stages (UPDATES both arms incl. version-change refire; SEAL_BRIEF
  two-sentence assertion executable) + env-scrubbed battery + sha to doyle, NO
  landing without gate. #11 close exceptions now: (a)→#133, (b/d)→#247,
  (c)→operator referral.
- REGISTER NOTE OWED at close sweep (todlando measured 2026-08-30): `xtask check`
  drift gate compares committed reference.md vs the ON-DISK spt binary —
  `cargo build -p <other>`/`cargo check` leave it stale, so it can exit 0 over
  real drift (kin IR-12 stale-artifact class). Battery remedy already structural:
  the mandatory workspace-bins prebuild leg runs before xtask.
- **W2 GATED + LANDED ~05:30Z**: main ff aa5f80f6 → `cb78f710` (respin 05a15396,
  one-hunk delta verified pre-pick, audit 1/1 MATCH). #44+#45 → ACCEPTANCE
  (comment 5466924902). ONE red closed at mechanism: xtask spacerun gate caught a
  30-space run in the new ATTACH_OBSERVER_UNSPAWNED diagnostic (lost-backslash
  class; only instrument looking at rendered diagnostics) — stop-referred,
  todlando repaired byte-diff-verified. Both-observers contention leg green
  (attach_link_push_e2e, real daemon). Shape = collapsed broker-hosted
  reconciler (approved deviation, grounds strengthened).
- **GOLDEN GREEN + RELEASE RUNNING ~08:19Z**: r1 red = resident_service_e2e :664
  teardown-LEAK, ruled LEDGERED CLASS at triage (signature predates the whole
  diff — delta test NO structurally; both first-CI cells + twohost green);
  same-sha rerun att2 = 9/9, cell re-executed PASS 5.160s (non-vacuity grepped).
  Gate pass ≠ class retirement — ledger row OPEN +1 (svcBOOT survivor,
  full-budget went_clean=false), hertz cluster strengthened. Release: main ff
  dc07e215..da71b785, tag v0.67.0 at ruled sha, CASCADE PRE-PUBLISH (v0.66.0
  lesson executed — #23 closed 08:18:56Z ACCEPTANCE, 8 members cascaded, #33
  skipped terminal; state verb wants #N ref form), counter → 102. Deployah
  holding quiet window; thin-red-at-golden-green sha = scope contradiction
  referred to me, else nothing owed. POST-PUBLISH QUEUE: close sweep (ledger +1;
  xtask stale-binary drift-gate note; --exact filter trap; spacerun 2nd live
  catch on IR-57's kin entry; IR-59 wave-pool rate face 236GB/3.5h; deployah's
  runbook full-sha pin fix — his), reap party (ns23-w1/-113/-w3 worktrees +
  gate rig, conhost pin-check first), ir57-carry: #242 will surface on THIS
  card per spt-progress-tool#78 (predicted, filed, not a new defect).
- **INTAKE PASSED ~06:00Z, golden held on CI re-earns only**: deployah's
  greenlit-form gate REFUSED my "deferred-attached" third form — #16/#17
  discharged properly (detach + EVAL, comment 5467039869, both halves of the
  drop rule; board-verified both sides). Member arithmetic corrected: 8
  ACCEPTANCE + #33 prior-arc DONE (rider ≠ member). His watcher reads run
  CONCLUSION not gh-watch exit. Reds hand back RCA-first.
- **GOLDEN HEAD HANDED ~06:00Z**: `da71b785` on release/v0.67.0-head = dc07e215 +
  shape (3 files, pre-shaped per IR-54); head gate clippy/treqs 0 post-reclaim;
  GOLDEN-HEAD-23-HANDOFF.md = the contract. Verify-closes DONE (#11 five
  exceptions incl. (e)=#254 mint; #14 clean w/ evidence-honesty note); referral
  block posted (#16/#17/#11c + BACKLOG roster #244/#246/#247/#254). 9 of 10
  members ACCEPTANCE. **DISK EMERGENCY mid-close**: C: hit 0.01GB (wave pools =
  236GB in 3.5h — IR-59 fifth-face rate confirmed); reaped 3 finished lane pool
  targets +186.6GB (targets only, JIT records intact, single-actor rule held
  when todlando's consent raced my in-flight reap); head-gate 101/101 + two
  main-CI reds classified DISK, re-run clean/re-fired; runner cache HELD
  three-way. Claim records reaped with targets → fresh pool-claim before first
  build in reaped trees (deployah banked). Golden reds hand back to me,
  RCA-first. Reap party after verdict (ns23-* worktrees + gate rig, conhost
  pin-check first).
- **W3 GATED + LANDED ~05:45Z**: main ff cb78f710 → `dc07e215` (audit 1/1 MATCH,
  6/6 legs 0, bins 753/753, e2e 1/1, ZERO respins). #133 → ACCEPTANCE. **All
  seven build members landed** (#245 #18 #113 #44 #45 #133 + #11b rider).
  Remaining: #11/#14 verify-close (todlando deltas inbound; my drafts staged in
  scratchpad close-drafts.md), operator referral block, golden head → deployah
  (shape-on-top: version bump + CHANGELOG; cascade BEFORE publish; #16/#17 stay
  attached-deferred with referral recorded — flag to deployah's intake check).
  Milestone reap party AFTER golden verdict (ns23-w1 + ns23-w3 + gate rig,
  conhost pin-check first).
- **W3 #133 history (landed)** (.worktrees/ns23-w3, pool ns23-w3): ruling (a) per-source
  hint cap with CONTEXT.md amended by replacement citing #133's greenlight as
  authority; adapter+arm-qualified shell keys, harness key spelling untouched
  (pre-lane-key no-refire unit); teaser names `spt adapter hints <adapter>`
  which the lane MINTS, parse-test built from the same constant. After W3 gate:
  #11/#14 verify-close evidence, operator referral brief (#11c/#16/#17+#246),
  golden head assembly → deployah.
- Deployah ARMED for the golden head (his words 02:15Z); drive ACCEPTANCE cascade
  BEFORE publish ([[release-verb-grace-vs-hub-card-window]]).
- hertz still offline+spool (fixup lane, 24-site population, anchored-matcher rider).
- Worktree reaps this window: ir57 (+209.5GB), asm-241, gate-w1-786d2381 (MINE — I
  mislabeled it deployah's once, todlando corrected; evidence at
  spt-preserve\v0660-close\gate-w1-786d2381\, cmp-verified). Remaining conhost pins:
  gate-w1-a8f04aff ×12, io-parser-w1 ×5 — sweep at audit slot with the PEB CWD-probe.

Related: [[v0660-arc-in-flight]], [[resumed-session-reground-before-acting]].

2026-09-01: #254 (spt-shells deprecation gate inert on cmd_bind/cmd_boundary — todlando delta-2 finding from #11/#14 verify) evidence base preserved at Documents/spt-preserve/v0670-todlando-lanes (39 files, SHA-verified, restore cost stated in NOTE.md). Candidate at next intake sweep.

2026-09-06: v0.67.1 docs-only PATCH (#274, deployah-driven, golden 34014574926 @34fdb848): ONE Linux Phase A red, brain_resume_conn_deadlock respawn_interleave 97 ticks vs floor 100 — ruled TEST-INFRA (HEAVY-qualifying binary never in the HEAVY strings; born 6 days before HEAVY-AT-BIRTH), same-sha rerun, hardening to hertz next batch. RCA-274-R1-LINUX.md at repo root. Lesson: the sweep criterion was applied FORWARD from the ruling date, never BACKWARD over pre-ruling files.

2026-09-06 DOCS-NITS LANE QUEUED (next batch, after v0.67.1 tags; all three from flynn's alchemy #79/#80 grills, all measured by me): (1) frames.md#io cross-link to json-shapes' REQ-DIGEST-SEAL-ON-IDLE paragraphs; (2) api hint doc-comment in crates/spt/src/api/mod.rs:424 'Needs --manifest' -> name the GROUP-LEVEL flag placement + the --adapter registry route (reference.md is generated from --help; drift gate); (3) CASE-2 GAP: shells/overview.md [[hints]] section states no match rule and omits the public [[hints]].regex field — rule = spt_store::matchrule::pattern_matches (literal case-insensitive substring, no word boundary, whitespace literal; regex opt-in; invalid regex silent), already documented for MONICS at networking/monics.md:466 — write it ONCE on the hints page and cross-link (one rule, two consumers), tag doc->REQ-MANIFEST-4.
  rider to (3), measured 2026-09-06 on spt 0.67.0 with exact bytes: regex rows are CASE-SENSITIVE (Regex::new, no (?i)) while literal rows are case-insensitive — state it + the (?i) idiom. flynn's 'TOML basic string \b loses a backslash' claim REFUTED: basic "\\bbags?\\b" and literal '\bbags?\b' both arrive intact and both fire (baggage negative control silent) — his writing tool collapsed one backslash level (unquoted-heredoc class, already in GATE-TEST-INDEX). No core defect; no docs sentence about spelling.
2026-09-06: hertz W2 rig-fixups lane test/w2-rig-fixups @2c9a06c3 (daemon-leak cluster): quiet-phase 30s reset on BRAIN_PHASE + 300s aggregate cap (reports class+elapsed), Linux census attributes leaked survivors by NEXTEST_TEST_NAME (nextest 0.9.137 provides it; absent/unreadable = unknown, no culprit), pre-battery POSITIVE CONTROL through real daemon->brain spawn with identity env scrubbed. Local greens (census selftest, activity 3/3, treqs 855/0, diff-check). REMAINING: kitsubito positive control + activity E2E, then PR/report. HEAVY reclass (RCA-274) stays a SEPARATE lane.
2026-09-06 07:00Z: v0.67.1 att2 = SECOND same-sha red, BOTH twins (88/97 ticks) -> pre-commitment fired: RESPIN with HEAVY reclass. I authored it (hertz offline at the time): branch fix/274-heavy-reclass-brain-resume @04e32c8c95cf09ddc2a44a51cd0233b0d13bdc64, parent 34fdb848 (ff-able onto golden/service-docs-0671): member in both <HEAVY> copies (asserted byte-equal before/after by script), nextest.toml stanza, FLAKE-LEDGER row same commit. Worktree .worktrees/reclass-274 — REMOVE after the fold lands. BOX CONFLICT: hertz resumed W2 proof on kitsubito (golden Linux box) at 06:56Z per operator — respin push sequenced AFTER his kitsubito leg reports done. Pre-registered discriminator: a Phase B red on the twins refutes the pool mechanism.
  (4) CASE-2 GAP (perri 2026-09-06): api.md io-events says "when truncated is set, follow digest_seq" and shows "digest_seq": 88 — MEASURED: IoEvent::with_seq has ZERO callers, no kind has ever emitted digest_seq; COMMUNE is not digest-backed (frame = file bytes at delete, bounded 16KB; ingest writes no digest entry). Amend: pointer reserved+unemitted (TOOL_USE-style honesty), truncated remainder not recoverable from core today. OPERATOR RULING same night: core does NOT parse !!wake!!; across-commune/wake are claude-spt concepts and do not belong in core — never propose a core-side wake bit. perri told: move the marker to the HEAD (frames head-truncate); core-side body retention = his ADDITION to file if wanted.
  2026-09-06 07:25Z: .worktrees/reclass-274 REMOVED (operator-prompted; clean, never built, 04e32c8c pushed + folded). git worktree remove hit Permission denied because my Bash cwd sat INSIDE it — content + registration were gone, only the empty entry stayed pinned; cd out, rmdir, entry gone. Local branch fix/274-heavy-reclass-brain-resume kept until v0.67.1 lands on main, then delete against a NAMED ref (merge-base --is-ancestor origin/main), never bare -d.
  2026-09-06 07:48Z: respin golden 34017906638 @04e32c8c — Linux Phase A GREEN (twins gone from it), Phase B GREEN (twins serialized; pre-registered discriminator fired the PREDICTED way = pool mechanism CONFIRMED, no re-triage), doctests + clippy green; drift/bundle, Windows, twohost still climbing (deployah). Acked. Post-land: delete local fix/274-heavy-reclass-brain-resume against origin/main ancestry.
  2026-09-06 08:05Z: DOCS-NITS lane JIT AUTHORED at repo root DOCS-NITS-0672-JIT.md (four items, all four anchors + REQ ids re-measured at 04e32c8c; item 4 widened by measurement: IoEvent.seq is ONE never-set field feeding io-frame seq, boundary-frame seq AND io-events digest_seq — three doc surfaces, zero emitters). Dispatch to todlando as a thin docs lane once v0.67.1 is on main. Wake text this session was STALE (#242-era W2 #238 grill) — re-grounded from f028fdc3 scratchpad (webserve-members.md + prior-grill-texts.md): WEBSERVE = milestone #272, grill round-1 posted 2026-09-05 03:19Z, #272 has ZERO comments as of 08:05Z — still operator-blocked.
  2026-09-06 08:03Z: respin golden 34017906638 @04e32c8c — Windows Phase B RED, ONE victim: spt::attach_link_push_e2e clause 5 (:473, `changed`=Some(node) vs expected None, 10.8s). RULED TEST DEFECT from the KERNEL not timing: attachment.rs plan_attach_pushes sets `changed` only on the State arm, which needs token_b already REMEMBERED → a Link push (None, viewers=0) had already gone out, so the viewer did NOT arrive while the link was down; the test spawns `spt rc --view` then binds with NO barrier (:450-465). Brain log: that rc's connect→subscribe 1.8s vs 33ms in clause 4. Twins GREEN on Windows → reclass holds both OS. Cell history: Win PASS @34fdb848 att1 8.06s (att2 Win row = carried copy), Linux PASS @04e32c8c 13.5s, v0.67.0 golden both OS; first red in ~5 CI executions. Ruling to deployah (acked 08:03:57Z): same-sha rerun --failed on TERMINAL (twohost still queued), pre-registered: 2nd red at :473 → barrier fix rides a respin; other red → me. Fix dispatched to hertz (test-only, post-land thin lane, barrier = viewer ATTACHED before bind by a route not riding the shell link; mutation proof = bind-before-spawn must RED; ledger row same commit).
  2026-09-06 08:05Z: perri CORE FINDING — echo-commune host (spt-live/src/echo.rs:156-166) writes the agent's OWN `<id>-commune.md` unconditionally; an authored across-commune awaiting ingest is overwritten (5/5 funnel correlation). Confirmed in source; filed via alchemy create --type bugfix (#276). Memory echo-commune-overwrites-fresh-commune.md UPDATED — writer no longer UNKNOWN.
  2026-09-06 08:06Z: echo-commune collision MINTED spt-bs-releases#276 (bugfix, backlog; first create refused at 3353 chars — alchemy create body cap is 1600, detail goes in a comment). perri replied with the ref; adapter work holds on the core ruling. #272 WEBSERVE grill still 0 comments (operator-blocked). Interim fleet rule: across-commune MID-turn + one more tool call after, never as the turn's last act.
  2026-09-06 08:45Z: CORE DEFECT found from perri's retraction — io-events seq RESTARTS at 1 once a log passes 256KB (iolog.rs last_seq_at: mid-codepoint seek → read_to_string InvalidData → 0 → seq 1; 9/9 on my log, todlando 5/perri 2/deployah 1, none under 256KB; trim arithmetic broken too). Filed via alchemy (ref pending on perch), lane JIT IOLOG-SEQ-RESET-JIT.md at repo root → todlando after #276, lands after v0.67.1. perri 0.38.4 = adapter half (refuse pre-session COMMUNE frames + per-fire trace). Golden 34017906638 att1 terminal 8/9 (only the ruled :473 red); att2 --failed running (deployah calls it).
  2026-09-06 09:05Z: golden 34017906638 att2 Windows Phase A RED, DIFFERENT cell: spt-store wtlock_two_process_int :147 — B WORKTREE_LOCK_TIMEOUT at 10005ms on its FIRST commit while A ran 6 serial commits in 13.5s (ARM1 line). Mechanism: sentinel is UNFAIR — holder re-acquires within us, waiter polls 20ms, so B's wait == A's whole run; quiet Windows A ~6.9s vs 10s bound sized for ONE sub-second op. RULED TEST DEFECT (at-budget = slow box by the code's own discriminator); same-sha att3 --failed; hardening to hertz (SPT_TEST_WT_LOCK_WAIT_MS on ARM1 children, ledger row, same thin lane as :473). Suspected load source: perri's claude-spt 0.38.4/0.38.5 builds on hfenduleam overlapping Phase A 08:52-08:56Z — build hold asked for att3 (rule of thumb: hfenduleam is the Windows golden box as kitsubito is Linux; no builds during the leg). :473 PASSED on att2. Pre-registered: att3 red same cell → respin w/ both fixes; third distinct timing cell → stop rerunning, respin + box-load audit. Random-victim-under-load family, both OS this arc.
  2026-09-06 09:12Z: emphasys omp-spt 0.5.0 field findings ruled: (a) adapter update ADDITIVE by design (cli.rs:20770-20776 apply_release_crc_swap, premise 'stale unreferenced file is harmless') — FALSE for strings/skills scanned by a harness; measured both layouts + stray claude-spt.exe (mtime 07-09) in _github/BigscreenVR-omp-spt; filed CHANGE via alchemy (ref pending): prune strings/ on update, binaries stay additive, semantics into public contract (docs gap: no update semantics stated). (b) rest_state=suspended vs ONLINE = known hazard rest-state-suspended-is-not-a-suspend-proof, not a defect. #277 iolog seq filed 08:44Z + comment 5558105715. perri monic recorded (operator-directed).
  2026-09-06 09:10Z CORRECTION (hertz measured): wtlock starvation is PROBABILISTIC, not structural — B catches a re-acquire gap on a quiet box (A24 13.2s green, A96 59.3s green, env unset); on the loaded golden box it caught none for 10s. My 'B wins only when A is done' overstated. Fix unchanged (child-only SPT_TEST_WT_LOCK_WAIT_MS=120000, ARM2 = deterministic proof), 'must red' mutation criterion withdrawn, non-repros recorded in the ledger row. Box hold extended to hertz (his local runs load hfenduleam during golden Windows legs). emphasys (a) = #278 (change, comment posted).
  2026-09-06 09:18Z: #276 GATE GREEN both OS @f8e60415 (todlando; exit files read; mutation proof measured: restoring the write reds both hazard cells with the byte-compare dump; exists() would stay green). PR held until v0.67.1 on main; my gate legs after the box hold. LOAD-AUDIT LEAD: todlando's 1240-cell Windows nextest ran on hfenduleam during att2's Phase A window (he was never on the hold — my miss); asked for his run's start/end UTC. Two process notes banked to GATE-TEST-INDEX (--no-fail-fast; run-is-the-census).
  2026-09-06 09:22Z LOAD AUDIT CLOSED: todlando's #276 gate Windows nextest (1240 cells, test phase 08:50:06-08:58:00Z, .raw/.exit provenance) FULLY CONTAINED att2's Windows Phase A (08:52:38-08:56:49Z) on hfenduleam — the named load source for the wtlock red. He was never told about the hold (my miss; perri+hertz were). Ruling stays TEST DEFECT. Standing rule → INFRA-REGISTER: golden Windows leg = quiet window on hfenduleam; driver pre-flight = gh run list --status in_progress on golden before local nextest/build; dispatch briefs name the hold. Att3 = clean measurement (all three holding).
  2026-09-06 09:25Z: IR-76 AUTHORED in docs/INFRA-REGISTER.md (golden Windows box shared with builders; job-start census blind to cargo/nextest load; fix = driver pre-flight refusal + census predicate + brief/runbook sentences). STATE: UNCOMMITTED, in the dirty golden/service-docs-0671 checkout — rides the v0.67.1 close sweep commit; not tracked in any lane yet. Att2 census fact: family_total=12 scoped=0 at 08:39:46Z while todlando's prebuild was live.
  2026-09-06 09:35Z: IR-76 COMMITTED 4633b40f + PUSHED on docs/ir-76-golden-box-quiet-window (off origin/main; worktree .worktrees/ir76 — REMOVE after the PR lands); NOT on main by design (main stays ff-able for the v0.67.1 head); thin PR after deployah's land call. todlando's amendment (three-arm pre-flight) in the entry + commit body. Main-checkout duplicate edit discarded after block byte-compare. Trap hit: persisted shell cwd broke a relative cd, && chain skipped the commit, push shipped an EMPTY branch at main's sha — always git -C <abs>, read the sha not the push output.
  2026-09-06 09:50Z WEBSERVE #272 grill round 1 ANSWERED by operator in chat (Q1/2/4/5/6 agree; Q2 needs disambiguation → `<stem>~<n>.<ext>` adopted; Q3 unsure). #271 → SHELVED via alchemy. Q3 MEASURED: fetch_file/push_file zero production callers (initiator premise TRUE), serve_xfer gate on XFER surface LIVE (xfer.rs:270/:341), zero XFER grant rows on this node, CONTEXT.md:876 ratifies minting for webservice facets → recommended mint WEB + RETIRE XFER as #246 rider (operator to confirm). Round 2 authored: WEBSERVE-272-GRILL-ROUND-2.md (Q7-Q12). Comment posted on #272. ⚠ Grep tool glob `src/**/*.rs` under crates/ returns FALSE EMPTIES — use type=rust instead.
  2026-09-06 09:52Z: golden att3 Windows RED (09:30-09:44Z, job 101463810585) = resident_service_e2e :664 teardown LEAK — 3rd classified occurrence of the ledgered row (svcmock.exe survivors pid 50484+, went_clean=false at full 60s), QUIET box (Phase A 197.8s vs 251s; todlando's last activity 09:08Z, all three holding). Three attempts, three DISTINCT Windows victims: :473 structural barrier (att1), wtlock bound under load (att2), :664 ledgered leak (att3). Load refuted for att3 by the gap. Sandbox-removal step also failed (leaked fixture holding files). Censusing leaked svcmock.exe on the box before any att4.
  2026-09-06 09:55Z RULED att3: ledgered class (row match strongest fact; deployah leaned 'stop' on the count, deferred), SAME-SHA att4 --failed after twohost terminal; my 'third distinct timing cell → stop' pre-registration OVERRIDDEN on the record (trigger was load-timing; this is a quiet-box leak row with a rerun-green precedent; a respin carries nothing for :664). HARD STOP: att4 red of any kind → no 5th attempt, release waits for hertz's leak-cluster hardening + dfcae9db on a respin, operator told. NEW datum: TWO survivors (svcboot pid 50484 + relshell pid 41892 svcmock.exe — relshell named for the first time). Box census pre-att4: zero survivors alive. todlando's one-varying-window family prior checked and dropped (quiet-box red refutes it; three distinct windows).
  2026-09-06 10:05Z WEBSERVE grill-with-docs CEREMONY (operator reminded: /grilling + /domain-modeling = ADRs + glossary as decisions settle): ADR-0056 node-prefixed URLs, 0057 registry primitive + ~n served names, 0058 pull-model attachments + spt fetch, 0059 LAN bootstrap-only; CONTEXT.md '## Web serving' section (8 terms) + docs-server alias clause. Branch docs/webserve-272-grill, worktree .worktrees/webserve-272-docs (REMOVE after PR lands), off origin/main; NOT on main until v0.67.1 lands. Q3 (mint WEB / retire XFER) NOT recorded until operator confirms; Q5's listener mechanics = Q12. Round 2 (Q7-Q12) pending in chat; each settled answer → glossary/ADR on the same branch.
  2026-09-06 10:20Z WEBSERVE round 2 ANSWERED (Q3 agree=mint WEB+retire XFER; Q7/9/10/11 agree; Q8 TTL 30d; Q12 port 5470 — IANA apsolab-col obscure, local free, Wikipedia list fetch truncated=unchecked). Ceremony: ADR-0060 (WEB/XFER), ADR-0061 (short-ID), amendments 0056 (grammar+http-only), 0058 (snapshot+30d), 0059 (5470 + separate allowlist); CONTEXT.md access-surface entry + attachment + message short-ID terms. FRONTIER EMPTY. Next for #272: wave map + REQ ids at intake (after v0.67.1 lands + #276/#277 lanes). Commit pending on docs/webserve-272-grill.
  2026-09-06 10:22Z: emphasys (omp-spt 0.7.0) asked whether 'bind for a hosted endpoint from an unknown sid is refused, not adopted' is a public guarantee. RULED case (a): api.md:152-161 states it (only guard = ownership; a live perch under a different sid is refused). BUT the line she cited, ER_HOSTED_PROBE:no-row, is NOT the refusal — engine_room_hosted() is evaluated EAGERLY as an argument at startup.rs:1111 (and :925), so the reserved-id probe prints on every bind on a node without an engine room while deciding nothing for non-reserved ids. Filed #279 (bugfix: evaluate lazily; unit = non-reserved bind emits no probe line).
- ⭐ v0.67.1 SHIPPED 2026-09-06 11:13Z: golden 34017906638 att4 GREEN 9/9 @04e32c8c, main ff'd 8a21a3b0..04e32c8c, tag == golden ref == main tip. Four attempts: att1 :473 attach_link (structural test defect), att2 wtlock :147 (test defect under builder load), att3 resident_service_e2e :664 (ledgered Windows leak row, 3rd occurrence), att4 clean on a QUIET box (IR-76 three-arm pre-flight held: perri/hertz/todlando froze cargo on hfenduleam). Preservation by deployah: 4 CI job logs at Documents\spt-preserve\v0671-golden-r1 (restore cost none); nothing else named. Post-land: fix/274 local branch deleted (ancestor-verified), docs/ir-76 @7c435564 + docs/webserve-272-grill @fbec981f rebased + pushed (PRs after release.yml 34029623827 + thin ci 34029605784 conclude — PRs fire ci.yml on the boxes, branch pushes do not).