# #307 JIT plan — v0.71.0 "REFRESH-FREEZE & RC-HITCH remedies" (fast-follow of #304)

doyle, 2026-09-16 01:40Z. Operator directive (2026-09-15): pull relevant/troublesome requests into #307 by doyle's judgment, plan, dispatch, drive to publish. Base: main `289e3cdf`. Target: v0.71.0 (counter 107).

## Members
- Carried from #304 relocation: #49, #267 (refresh freeze), #302 (rc hitch).
- Added 2026-09-16 by doyle's judgment (resolves the pending "milestone home" ruling both mints carried): #308 (serve-registry lost-update, prep draft `fix/308-registry-process-lock` @fefc4d64 exists), #309 (ER briefing lost when liveness reconcile offlines a bringing-up session). Both are intermittent CONTRACT losses found by the #304 consumer leg with hertz RCAs on file.
- Added 2026-09-16 by operator ask: #310 (`spt update` concludes with outdated links — point Changelog/UPDATE_DOCS_LANDED at the locally hosted docs `http://localhost:5474/<node>/docs` and `.../docs/changelog.html`, plus any sibling surfaces). Small product change, todlando lane, queued between #308 and #309.
- Excluded on purpose: #296 (field PASS on omp-spt 0.9.2 = adapter-side; no core defect shown). #25/#263 adjacent but unscoped; stay BACKLOG.

## Waves
**W1 (parallel, now)**
- **W1a — todlando, #308.** Rebase fefc4d64 (base 7890ead3) onto 289e3cdf; build; new fail-closed unit + two-process int (`serving_registry_two_process_int.rs`) via nextest; targeted `serving`/`webserve_attachment_e2e` suites; activate the REQ stages in the same commit; traceable-reqs 0.4.1 check; thin PR. Pool: claim from the lane's own worktree.
- **W1b — hertz, #49/#267 observation rig (evidence, not fix).** Per `docs/49-267-REFRESH-OBSERVATION-PLAN.md` (measurement/teardown rules) + `FAST-FOLLOW-PLAN-draft.md` §1–4: isolated `SPT_HOME`, `SPT_PUMP_TRACE=1`, binary pinned by sha (main 289e3cdf build), disposable brain-owned controller/PTY with counter+generation-stamped output, real `spt daemon refresh`; capture `CONTROLLER_SLOT_CLOSED` (Q1) and `BRAIN_PROMOTE_GATE_NOT_DRAINED` (Q3). Q3 needs a real `AppliedPending` trial — report BLOCKED if not admittable, never fabricate. Bound 300 s + 120 s teardown. Output: receipt + RCA candidate that explains why prior fixes failed.
- **W1c — hertz, #302 discriminator.** Isolated-home daemon on hfenduleam (NEVER the resident fleet daemon), local rc session, `SPT_RC_HITCH_DIAG=1 spt node status --json` during healthy/hitch/recovery; the five-line cadence experiment ONLY on the isolated home. Plus: doyle asks lia to authorize ball-b (ENLYZEAM) to run the read-only probe on the operator's box during a real hitch. Output: #1 (runtime starvation: seconds-stale canary + hitch-correlated dials) vs #2 (synchronous broker/stream wait: fresh canary).

**W1d — todlando, #310** (after #308 is handed to gate, before #309): grep every site printing the releases URL / UPDATE_DOCS_LANDED path; replace with the node-scoped local docs URLs; one unit cell pinning the rendered lines. Operator authorized ball-b readings on ENLYZEAM and the ssh route decid@enlyzeam (read-only) for W1c.

**W2 (after W1 receipts)**
- **W2a — todlando, #309.** Coherent live/starting/absent observation at the spawn/reconcile authority boundary: livehost session poll honours the broker's `bringing_up` ledger; false-death mutations (incl. control cleanup) protected. NOT a longer briefing wait. Red-first unit on the reconcile arm + `er_briefing_presented_e2e` clause 4 green. hertz: regression cell hardening if the e2e needs a deterministic barrier.
- **W2b — todlando, #49/#267 remedy** from the W1b RCA (RCA-first; agreement never promotes to proven). hertz: regression cell on the W1b rig. Rider if the path is touched: IR-82 (daemon.pid predecessor/death record, todlando-owned).
- **W2c — todlando, #302, two steps (ruled 03:05Z on hertz's source RCA `.spt/preserved/302/hertz-meet-rotation-rca.txt`; #302 comment 5691417538).** STEP 1 discriminator: monotonic stamps around interface enumeration / portmapper gateway lookup / CancelMibChangeNotify2 / meet endpoint bind+retirement / NTP refresh, plus both-worker stack capture when the 25 ms canary goes stale; fielded on ENLYZEAM under the unchanged three-subnet rotation. STEP 2 conditional: offload the PROVEN blocking OS boundary with bounded admission, or isolate the COMPLETE pre-trust meet lifecycle on a bounded separate executor. Rejected: family-probe cache (already offloaded), dial-only seam B, more workers.

**W3 — assembly.** doyle assembles the golden head (merge chain off current main, compile-gate + traceable check), hands to deployah; deployah verifies GREENLIT FORM (every member fulfilled or relocated-with-reason on the issue) → golden → publish v0.71.0. Red → doyle RCA-first triage.

## Gate discipline (every lane)
Isolated worktree under `.worktrees/`, own pool claim, targeted suites + compile-gate, `traceable-reqs check` exit 0, ff-only landing of the tested sha, co-author trailer space spelling. Register sweep at intake done (hertz, 82333ca1; IR-86 retired, IR-11 built).

## Findings log
- 2026-09-16 02:00Z — ENLYZEAM at spt 0.69.0: hitch gauges UNSUPPORTED until the operator updates that box to 0.70.0 (asked via lia). #302 field reading blocked on that, not on the knock.
- 2026-09-16 02:01Z — hertz W1c cadence arm on isolated hfenduleam 0.70.0 home: `node refresh` froze PTY ACK/progress; replacement controllers hit the 15 s write-poison bound; canary fresh, dials 0. Classified #49/#267 refresh-freeze reproduced (W1b subject), NOT #302 candidate #1. Artifacts preserved under `.spt/preserved/307/hertz-w1bc/`.
- 2026-09-16 02:00Z — todlando #308 red-first proved (native 100, sentinel-unlocked witness), source restored; green inventory running.
- 2026-09-16 02:03–02:11Z — Operator told doyle to update ENLYZEAM over ssh. Attempt 1: UPDATE_FETCH_REJECTED:GhAuthRequired (stale stored gh login SaberMage; release channel private). Attempt 2 with transient GH_TOKEN: same refusal — measured `gh auth status` exits 1 with a valid token while the stale entry exists (#312 filed). Attempt 3 (logout stale entry + update) wedged pre-install; ball-b (ENLYZEAM resident) objected via lia; run STOPPED 02:09:40Z, nothing landed (0.69.0, pid 16832, no bounce, three independent readings), my leftover shell reaped. #313 filed (access asymmetry, needs-operator). Enlyzeam next step = Reavo repairs gh login there, then update, then ball-b probe. #302 field reading still blocked on that.
- 2026-09-16 02:06Z — hertz W1c RECEIPT `.spt/preserved/307/hertz-w1bc/w1c-run1/RECEIPT.json`: NO steady-state hitch repro on isolated hfenduleam (130 probes, canary 0–26 ms, dials 0); refresh-induced viewport failure preserved as #49/#267 evidence; config restored byte-for-byte; teardown clean. W1b running.
- 2026-09-16 02:07Z — todlando #308 Windows green native 0, 44 selected cells (fail-closed, three process regressions, both webserve composites); traceability + changelog page in progress.
- 2026-09-16 02:14Z — hertz W1b COMPLETE, gated PASS: refresh-freeze reproduced (two arms); Q1 mapped (conn17 detach; replacements 37/47/58 send_deadline then write-poison); Q3 genuine trial, NOT_DRAINED not observed (rc controllers by=Some excluded from the gate). Leading candidate: acked pending-input flush attach.rs ~754 on the sole consumer thread; competing: synchronous output forwarding. W2b dispatched to todlando: STEP 1 discriminator instrumentation, STEP 2 remedy, both on hertz's rig. hertz W2 test lane: reusable refresh-freeze regression rig + nextest cell, red on 289e3cdf. Lane order on todlando: #308 -> #310 -> W2b (#49/#267) -> #309. Board record: comment on #49.
- 2026-09-16 02:19–02:29Z — ENLYZEAM updated: doyle staged the 0.70.0 signed set by hand from hfenduleam (hash-verified both ends) + `spt update apply` (coordinator 0.70.0, broker layer 0.69.0); operator then bounced the daemon manually => broker 0.70.0, pid 48932. Probe live: canary 187 ms, dials 0. #302 field reading unblocked: ball-b PTY-correlated triple asked; hertz 1 Hz passive series asked. Bypass path recorded on #312. Accidental no-op `spt update apply --json` at 02:17Z disclosed.
- 2026-09-16 02:20Z — todlando #308 handoff PR #226 (c9e80511; fixup f6b1dd9a = one hazard sentence). Gate driver running (worktree gate-226-c9e80511). todlando on #310.
- 2026-09-16 02:32-02:48Z - #315 filed (WAN MSG admitted while knock pending on the 0.70.0 broker; refused by 0.69.0 with tier NodeMode). Source: gate.rs identical across releases; Allow never logs; an ImplicitOpen bottom is silent. ball-b posture: BIGNET unmoded => chain bottoms out open. Hypothesis: origin subnet set gained BIGNET across the restart. hertz RCA queued. Not in #307 unless a confirmed regression; a posture defect is operator territory (#313).
- 2026-09-16 02:44Z - hertz W2 rig gated PASS as instrument (50876f70; cfg(windows); HEAVY filter only; negative control RED with partitioned checks). LANDING HELD: rides todlando's W2b PR as commit 1 of 2; python under Runner.Worker to be pinned (SPT_REFRESH_RIG_PYTHON) or evidenced.
- 2026-09-16 02:35Z - #226 gate mechanics PASS on my side (cargo check --tests 0; spt-store 641/641 incl. 3 two-process cells; checker 923/0; fixup delta = one doc sentence). Landing waits on the CI Windows leg for f6b1dd9a.
- 2026-09-16 02:48Z - hertz ENLYZEAM 1 Hz series (180/180, 02:44:36-02:47:35Z, broker 48932/0.70.0): canary age p50 18 ms, p95 1872, p99 3991, max 4912; 17 samples >= 1 s; active_dial_tasks 0 in ALL 180. Stale canary WITHOUT counted dials on a box with 4/7 peers dead at quic-connect for ~4.9 days. Reads as candidate #1 (runtime-progress starvation) with the starving work NOT in the counted dial population - suspect pump-round submit_dial to dead peers (10 s bound, PUMP_PEER_FAIL x89). Correlation asks out: ball-b PUMP_PEER_FAIL timestamps in the window; hertz pump period + whether active_dial_tasks counts submit_dial. If it holds: W2c remedy = seam B covering BOTH dial and submit_dial (REMEDY-DESIGN already says so), plus the gauge population fix.
- 2026-09-16 02:50-02:54Z - #302 DISCRIMINATED (field, ENLYZEAM 0.70.0): ball-b's log grep falsified the dead-peer-dial sub-hypothesis (zero PUMP_PEER_FAIL in the window), then aligned every stale-canary peak 5-10 s after a PAIR_MEET_UP round start (~30 s registry period; 3 subnets => 3 NET_FAMILY_GATE "binding IPv4-only" + 3 PAIR_MEET_UP per round), with the log silent during peaks and the CONN_LIFECYCLE burst off-peak. hertz: no daemon.json on ENLYZEAM (defaults 30/15/60 s); active_dial_tasks excludes synchronous dial(), inbound proof and unpolled submits. Candidate #1 holds; blocker = in-round non-dial work; leading pointer = per-subnet synchronous IPv6 probe/rebind (kin: broken-ipv6-poisons-iroh). hertz source read dispatched ahead of #315. W2c remedy shape to follow that read (seam B may not be the right cut if the stall is the family probe).
- 2026-09-16 03:02Z — #308 LANDED: PR #226 ff-pushed, main = f6b1dd9a; CI 35047933904 5/5 green; branch deleted; #308 comment 5691417348.
- 2026-09-16 03:02Z — hertz #302 source RCA delivered (LMT7OBIT): 30 s = TOTP meet rotation, NOT registry pump; family probes already spawn_blocking; candidates = synchronous netdev/portmapper/netmon boundaries in iroh deps at the v0.70.0 lock, none measured. W2c re-ruled two-step (above). hertz back on #315.
- 2026-09-16 03:07Z — hertz #315 RCA (`.spt/preserved/315/hertz-reply-exemption-rca.txt`): admission = node-keyed 15-min ReplyExemption (ball-b→lia SENT 02:24:45Z; doyle inbound 02:29:02Z), returns before ACL; knock not an input; BIGNET falsified by tier precedence; no 0.69→0.70 access delta. Ruled: out of #307; policy fork flagged needs-operator; ACCESS_ADMITTED audit line = later rider. hertz next: #314 defunct reattach RCA.
- 2026-09-16 03:22Z — hertz #314 RCA (`.spt/preserved/314/hertz-defunct-leaf-rca.txt`): zombie classifier rules a live adapter-labelled root with no descendants a zombie after 30 s grace; W1b Python producer is a legitimate leaf client (attach 1.9 s ok, reattach 81.4 s refused, same PID/birth alive). Core lifecycle-contract defect, NOT #307; AlreadyLive reap hazard source-predicted. Ruled out of v0.71.0, next-milestone candidate; hertz next = bounded fresh-home leaf reproduction (no refresh) with raw KIND_SESSIONS + descendant census.
- 2026-09-16 03:29Z — #310 gate PASS on 4b6b03cc (doyle rebased e1036ef9 onto f6b1dd9a; product diff byte-identical): 10/10 cells, xtask gen/check 0, checker 0.4.1 0; #310 comment 5691599796. CI 35050708554: 4/5 green, Windows leg queued. LAND when green.
- 2026-09-16 03:35Z — DISK: free fell 197→94 GB during my #310 gate (the 310 pool grew 12.4→67.6 GB under the nextest build) and todlando's W2b rig held at its 96 GiB floor. Classified (real dir, no inbound links, no resident process, both pool claims released) and reaped ONLY `.worktrees/310-local-update-docs/target`: 125.5 GB reclaimed in 15 s, free 219.9 GB. todlando cleared to launch W2b.
- 2026-09-16 03:38Z — W2b STEP 1 first discriminator RED (native, owned teardown): replacement worker 12604/13 buffered 31 inputs; pending send #4 entered 1789529783358, returned only at conn 61 write-poison 1789529798338 (~14.98 s); Request/output-forward calls had returned; no reads during the pending flush. Blocked location MEASURED. todlando adding narrow broker reply probes to separate the ACK cause from Subscribed/output contention before remedy.
- 2026-09-16 03:48Z — W2b STEP 1 DONE (run 2, atomic probes): broker applies pending op1 then blocks sending Applied while worker blocks writing pending op4; two-way write cycle, released by poison 14979 ms later; repeats on conn 79. STEP 2 ruled: replay pending input via existing send_effect_no_ack (dedup preserved); gate = unchanged rig green + attach/rc seam suites + repeat-proof. Recorded on #49.
- 2026-09-16 03:53–03:56Z — W2b no-ACK differential GREEN (recovery 3485 ms, 23/23 tags, 31 pending flushed in 43 ms, zero poison); sole 2.5 s outlier = first post-refresh tag waiting 2437 ms for the replacement worker, attributed (slow-tag-attribution.json). Final producer: probes stripped, seam suites, repeat rig, docs.
- 2026-09-16 03:58Z — hertz #302 STEP 1 driver delivered (STEP1-DRIVER.md + STEP1-DELIVERY.json sha a81c4d39); #302 comment 5691820885. hertz next: #309 regression rig on current main, branch test/309-er-briefing-regression.
- 2026-09-16 04:00Z — #310 LANDED: PR #227 ff-pushed, main = 4b6b03cc; CI 35050708554 5/5 green; branch deleted; #310 → acceptance. Post-merge ci run now occupies the runner (merge-push re-occupies the box).
- STATUS 04:52Z: landed #308 (f6b1dd9a), #310 (4b6b03cc), W2b #49/#267 (39241675, PR #228, gate receipt .spt/preserved/307/gate-228/). main = 39241675. In flight: #309 rig (hertz: fixture stdin-capture fix, then native RED + knob-unset control receipts). Queued: #309 product (todlando, on hertz rig), W2c #302 STEP 1 instrumentation → ENLYZEAM field run with hertz driver → STEP 2. Then W3: golden head → deployah.
- 2026-09-16 03:35Z — hertz #314 confirmation (matched pair, comment 5691648539); hertz now on #302 STEP 1 field-driver contract (.spt/preserved/302/STEP1-DRIVER.md).
- 2026-09-16 02:53Z - todlando #310 native proof: 10/10 selected, real signed apply printed node-local changelog/index URLs (HTTP 200), offline + no-listener controls kept old links; docs/traceability/PR next.

## One decision that may come back to the operator
#302's remedy cannot be CHOSEN without a field reading. If W1c yields neither an hfenduleam repro nor an operator-box reading, #302 RELOCATES (never drops) with the reason on the issue before golden. Everything else proceeds without operator input.

## Pre-flight (role discipline)
Cost: W1a ~1 lane-day (prep exists); W1b/W1c ~half-day each, light rigs. Box: hfenduleam runner idle (last run 289e3cdf 15:54Z), 182 GB free (above the 128 GiB golden floor, but tight — reap own retired targets before golden). Parallel fill: all three W1 legs run concurrently; todlando's cargo vs hertz's single isolated daemon = CPU contention only.
- STATUS 04:55Z: resumed post-recharge. #302 STEP 1 dispatched to todlando IN PARALLEL with #309 rig (hertz still building: er309_gate.rs + fixture stdin capture uncommitted, branch at 4b6b03cc, needs rebase to 39241675). Pinged hertz for ETA. Alchemy #302 comment spooled (shx-1789534537528281700). Next: hertz receipts -> #309 product to todlando; todlando STEP1 PR -> gate -> stage ENLYZEAM -> hertz collector verdict.
- STATUS 05:39Z: BLOCKED ON OPERATOR. hertz + todlando (omp-spt, provider openai-codex model gpt-6-astra) return HTTP 429 usage_limit_reached on EVERY turn since 04:30Z ("Try again in ~7162 min" ~5 days). Evidence: ~/.omp/agent/sessions/-Documents-projects-spt-core/<sid>.jsonl last assistant entries stopReason=error; bun.exe 0 CPU/5s; endpoint list still says ONLINE+CONTROLLED. #309 rig state: original scheduled RED-on-main captured (hertz 04:29Z); corrected-fixture final runs (baseline-red-final, knob-unset-final) staged argv+env at 04:30Z but NEVER executed (no nextest logs); 8 uncommitted files in .worktrees/hertz-309-er-briefing, branch at 4b6b03cc. #302 STEP 1 dispatch to todlando delivered but never read by a working model. Options for operator: (a) switch hertz/todlando provider/model, (b) wait ~5 days, (c) authorize me to finish #309 rig + run STEP 1 solo.
- STATUS 06:03Z: builders RECOVERED 06:02Z (same provider openai-codex/gpt-6-astra; quota lifted ~90 min after the 429 said ~5 days). hertz resuming #309 rig: rebase onto 39241675, both corrected-fixture runs, commit, receipts on #309. todlando started #302 STEP 1 (reading driver, lane off 39241675 next). Waiting on hertz receipts -> #309 product handoff to todlando.
- STATUS 06:18Z: #309 rig ACCEPTED (c0ef9571, receipts verified). todlando dispatched to #309 product (fix/309-bringup-liveness from c0ef9571); #302 STEP 1 parked at clean commit (forks netdev a1241fa3, netwatch 603d6189; pool migrating from 49-267 sequential takeover, C: free 133 GB). Next: gate #309 PR -> ff -> alchemy acceptance; then #302 STEP 1 proofs -> stage ENLYZEAM.
- STATUS 07:31Z: #309 LANDED — PR 229 eed9ff36 ff-pushed, main = eed9ff36. Receipt .spt/preserved/309/gate-229/. Remaining in #307: #302 STEP 1 (todlando: diag branch, symbolization fix in progress, field staging blocked on resume-failure path + symbols), then hertz collector verdict, STEP 2 conditional; hertz LEAK rider. Then W3 golden head -> deployah -> v0.71.0 (counter 107).
- STATUS 07:40Z: resumed post-recharge. #302 split contract raw-copy-resume-v1 APPROVED (ruling FB4BCXZD sent todlando+hertz; hertz accepted, amending analyzer + successor DELIVERY hash; todlando owes 4 proofs before ENLYZEAM staging). hertz #309 LEAK rider 84e359aa (test-only: livehost fixture blocks shell builtin on owned stdin, RAII-owned children + IR-108 entry) — gate running .spt/preserved/309/gate-leak-84e359aa (hertz pool sequential takeover, cell x3 + livehost 27). Next: PR + ff-push rider, alchemy #309 comment; then wait on todlando proofs -> stage diag binary -> hertz collector verdict; then W3 golden head -> deployah.
- STATUS 08:05Z: PR 230 (hertz LEAK rider 84e359aa) gated green, receipt .spt/preserved/309/gate-leak-84e359aa; CI 4/5 pass, Windows unit pending (waiter armed); alchemy #309 comment 5693857643. #302: hertz analyzer pinned (6f36563c, DELIVERY bd7f8f3e, shas verified); todlando forced-2 proof 3 blind-read ACCEPTED (18 marked pairs, raw 0-1ms, FALSIFIED via boundary-outside arm); scrub of 370 fmt-churned files done; waiting on post-scrub rebuild re-run + proof 4 + sha triplet -> stage ENLYZEAM.
- STATUS 08:21Z: hertz LEAK rider LANDED — PR 230 ff-pushed, main = 84e359aa (CI 35070000101 5/5). Post-merge ci run now occupies this box's runner alongside todlando's clean-commit proofs (3142753a). #302: todlando seams 255/255, resume units 3/3, smoke+forced parse pass on 3142753a; awaiting handoff (branch sha, stat, spt.exe sha256 distinct from net_diag_smoke.exe, analyzer sha, receipt) -> stage ENLYZEAM -> hertz collector verdict.
- STATUS 08:31Z: #302 handoff 3142753a verified + branch pushed. RULED release profile (debug confounds starvation); todlando: release rebuild + re-proofs + ENLYZEAM-matched OS PDBs (DLL pull running bg b7xecg6n0 -> .spt/preserved/302/enlyzeam-19045.6466-dlls). hertz: isolated-home rehearsal (neg control knob-unset, pos control knob-set) + STAGING-KIT.md for operator. ENLYZEAM: 19045.6466, task-launched daemon, no debug pin, hosted ball-b/engine-room, 15 GB free. Swap = NEEDS-OPERATOR window (alchemy #302 comment). Then collector -> verdict -> STEP 2 conditional -> W3 golden head.
- STATUS 08:40Z: resumed post-recharge (main 84e359aa). ENLYZEAM DLL pull COMPLETE + verified (15/15 sha, VersionInfo 19041.x ≠ local 26100; stray PS error line in SOURCE.txt replaced). hertz accepted ITLGY7RI (isolated HFENDULEAM daemons only, unset/set controls, task-env inheritance verified not assumed, kit w/ rollback). Waiting: todlando release rebuild+re-proofs+PDB manifest; hertz rehearsal+STAGING-KIT.md. Then blind-read -> NEEDS-OPERATOR comment on #302.
- STATUS 08:47Z: #302 OS-symbol leg ACCEPTED (todlando): enlyzeam-symbols/ flat 15 PDBs, manifest.json sha 766df70a…, each entry PE CodeView GUID/age vs PDB identity (OpenValidate4 rule), official msdl download http 200; doyle blind-verified 15/15 flat shas + identity flags + independent PE parse of ntdll/kernel32 CodeView = manifest. Note for kit: manifest search_path_environment carries the HFENDULEAM path; ENLYZEAM env must point at its local copy; copy flat *.pdb + manifest only (skip attempts/). todlando: release build next, proofs held for hertz window; release spt.pdb to be added to the same dir.
- STATUS 08:41Z: hertz ENLYZEAM read-only probe: task principal decid/InteractiveToken, action 'daemon run --detached', daemon 48932 session0 parent gone (matches doyle 08:25Z). DISCREPANCY noted by hertz: normal status shows 4 local endpoints vs my '+5 suspended' — hertz preserving full census separately; incomplete status never authorizes touching endpoints. Kit will verify the new daemon's session+env, not assume. todlando release build running (3142753a, opt3, line-tables, jobs2).
- STATUS 08:55Z: W3 prep — #307 greenlit-form parity (gh api sub_issues): 6 children; #49 #267 #308 #309 #310 closed+ACCEPTANCE (all on main 84e359aa); #302 open GREENLIT = sole remaining. No drops/adds vs greenlight. hertz: rehearsal rig + kit DRAFT ready (helper smoke pass, traceability 924/924), controls held on todlando release build. Critical path = release build.
- STATUS 09:00Z: #302 RELEASE capsule ACCEPTED build-only (todlando): artifacts-release-build-3142753a/ manifest c5ac46ca…, spt.exe 67dbe711…, fixture 31ae6ed0…, spt.pdb 61b41db2…; doyle verified: manifest+receipt shas, native_exit 0, after-tracked.diff = empty sha (source unchanged @3142753a), env CARGO_PROFILE_RELEASE_DEBUG=line-tables-only OPT=3 STRIP=none JOBS=2; my PE parse spt.exe CodeView 64E2A01E… age1 == manifest pdb_identity; bundle enlyzeam-symbols/spt.pdb byte-identical (16 flat PDBs, manifest 6431f18c…, application_symbols identity_match). STILL OWED: proofs (knob-unset seams, resume units, smoke+forced parse under analyzer 6f36563c) on release exe AFTER hertz controls window; hertz controls + kit finalization with these pins. No staging authorization yet.
- STATUS 09:25Z: #302 RELEASE PROOFS ACCEPTED (todlando handoff .spt/preserved/302/todlando-release/handoff.json, status local_release_proofs_complete_field_staging_requires_doyle). doyle blind-read: exe 67dbe711… identical in artifacts-release-3142753a and -build- dirs; seams 255/255 (1015 skipped, empty tracked diff); resume units 3/3; smoke (knob-set pos control) parses under analyzer 6f36563c: 15/15 usable, 0 stale, dials zero, INSUFFICIENT_EVIDENCE (14 s window, expected); forced: 18/18 complete bounded pairs, raw capture spans 0-0 ms (gate 100), processing 11-43 ms (debug was 89-415 — profile ruling confirmed), serialized, no unexpected issues, synthetic FALSIFIED. Limitations recorded honestly (forced runs fixture not spt.exe; ResumeThread proof is helper unit; no clippy rerun). Staging decision = mine, pending hertz controls + kit.
- STATUS 09:40Z: #302 STAGING KIT ACCEPTED (hertz): STAGING-KIT.md 01246ff6…, DELIVERY b00541cc…, CONTROLS 3f057e31…, ZIP 302-operator-3142753a.zip 03f1d9c7… (21 entries: release spt.exe 67dbe711/spt.pdb 61b41db2 MATCH, 15 OS PDBs + manifest 6431f18c MATCH, staging-manifest 3993d261 with pins/flags/suspended_ids, collector). Controls 4/4 (debug/release × unset/set): unset = zero NET_DIAG, set = fresh START + IF_ENUM/GATEWAY/NTP boundaries, all INSUFFICIENT_EVIDENCE on 55 s windows as expected, native COMPLETE/CONFIRMED_GONE. doyle audited whole kit (655 lines): bounded helpers, exact-handle stop, hash-pinned payload, USER env + process-level Process Explorer gate, rollback restores prior values; CLI fields local_endpoints/broker_image/--workers and env knobs SPT_RC_HITCH_DIAG/SPT_NET_DIAG_SYMBOL_DIR confirmed at 3142753a; task canonical pin 4fcc291c in receipt. NEXT: NEEDS-OPERATOR comment on #302 (flag verb refused on GREENLIT) -> operator runs swap -> collector verdict.
- STATUS 09:47Z: #302 NEEDS-OPERATOR comment spooled via alchemy-0 (shell send, 2941 chars; ack pending). BLOCKED ON OPERATOR: ENLYZEAM swap per .spt/preserved/302/STAGING-KIT.md with ZIP 302-operator-3142753a.zip. After swap: ONE collector (kit §7 or step1-field.py --out .spt/preserved/302/field-release-3142753a-operator) -> frozen analyzer verdict -> STEP 2 conditional on CONFIRMED; else #302 relocates before v0.71.0 golden head. Builders idle by design (todlando lane diagnostic-only, hertz native lane clear).
- 09:48Z: alchemy ack — #302 NEEDS-OPERATOR staging comment = 5695200444.
- 2026-09-17 08:45Z: OPERATOR stopped ENLYZEAM daemon (spt daemon stop; inhibit present) and delegated the swap to doyle over ssh. DONE by doyle: ZIP scp'd (pin ok), extracted to preserved\302\payload-3142753a (19/19 hashes), run dir preserved\302\operator-20260917T084033940Z (original.json: exe 81b4d0a2… spt 0.70.0, no pdb, USER env both empty; spt.exe.original backup hashed; old 26 MB log moved aside), symbols-3142753a staged (16 + spt.pdb), installed exe 67dbe711/pdb 61b41db2, USER env set, 'spt daemon start' from a shell with SPT_RC_HITCH_DIAG=1 + SPT_NET_DIAG_SYMBOL_DIR in-process (WMI rung forwards every in-process SPT_* via cmd /c set — daemon.rs wrapped_daemon_command @3142753a — deterministic env proof). RESULT: broker 42904 + brain 47684 session 0, inhibit cleared, fresh NET_DIAG_START broker_pid 42904 run_id 976b346f… 2 workers canary 25 dropped 0 full caps; ball-b + engine-room DAEMON_RESTART_RESUME'd (not killed by the stop). DEVIATION from kit: task path not used — task LastRunTime 7/30 vs daemon start 9/15 proved the resident was never task-launched; ssh+daemon start = operator's prior method. OPEN: daemon status --json returns empty from ssh shell (knob inherited from USER env?).
- 2026-09-17 08:47Z: STAGING VERIFIED: status --json (knob unset) running true pid 42904 net_up true 55 ms; diag sampler (knob set) canary age 8 ms — earlier empty status was the Start-Process Handle/ExitCode trap, not the daemon. hertz dispatched (SENT) to run the single collector -> .spt/preserved/302/field-release-3142753a-operator. #302 swap-done comment spooled to alchemy. Receipts mirrored: .spt/preserved/302/field-staging-20260917T084033940Z/.
- 2026-09-17 09:05Z: FIELD READING (field-release-3142753a-operator, hertz collector exit 0, analyzer 6f36563c exit 2): 180/180 usable, 4 stale intervals (peak 4011 ms), each round-aligned (NET_MEET_ROUND steps 59654492/3/5). v1 verdict INSUFFICIENT_EVIDENCE = INSTRUMENT MIS-SPEC: all 4 intervals have accepted both-worker in-peak captures, tid-matched sync NET_IF_ENUM spans, frames ntdll!NtDeviceIoControlFile<-NSI!NsiGetAllParameters<-iphlpapi!ResolveIpNetEntry2<-iphlpapi!SendARP<-netdev::os::windows::interface::interfaces (via iroh::socket::bind<-bind_scoped<-meet::bind_rendezvous, and via netdev::route::get_default_gateway<-netwatch HomeRouter::new<-portmapper::ip_and_gateway); SYMBOLS[NET_IF_ENUM] regex does not match these names. Symbolization WORKED. Blocked-time: tid 5852 150452 ms, tid 18952 123623 ms of 179 s; IF_ENUM n=248 (92 @1.1-1.2 s, 8 @2.1 s, 2 @3.1 s); GATEWAY_LOOKUP n=200; NET_MEET_BIND n=48 async 231-4373 ms. hertz dispatched: analyzer v2 (widen NET_IF_ENUM/GATEWAY symbols only), deterministic re-run on preserved capture + forced positive control.
- 2026-09-17 09:08Z: ANALYZER v2 ACCEPTED (6314743d…, DELIVERY 68d442df…; diff = 2 regex lines only; forced control byte-identical; deterministic replay): field summary mechanism CONFIRMED (interval 1: 5 both-boundary captures; 2-4 single-capture INSUFFICIENT). STEP 2 DISPATCHED to todlando: branch fix/302-meet-offload off 84e359aa — (1) meet-listener lifecycle on dedicated bounded spt-meet runtime, (2) rendezvous listeners PortmapperConfig::Disabled; gate = existing suites + diag-stamped field re-run on ENLYZEAM (0 stale intervals ≥180 s, 3 subnets meeting). alchemy dispatch #302 -> WIP. Finding comment on #302 spooled. ENLYZEAM stays on diag 3142753a meanwhile.
- 09:12Z: alchemy ack — #302 finding/STEP 2 ruling comment = 5711763508; #302 dispatched to todlando (WIP).
- 2026-09-17 10:20Z: hertz reaped hertz-309-er-briefing/target + hertz-307-evidence/target (27.9 GB logical, +22.6 GiB free, final 135.9 GiB; receipt .spt/preserved/302/landed-target-reclamation/DELIVERY.json). todlando remedy COMMITTED a593ece0 (fix/302-meet-offload, 13 files +473/-29, diag-free): spt-meet 2-worker runtime owned by NetHost, MeetRotation stop/JoinSets, EndpointPurpose::Rendezvous => portmapper Disabled, lib test pairhost::meet_runtime_tests::three_subnet_rotation_keeps_lifecycle_off_broker_workers (2 binds held, canary peak < 250 ms, all events on spt-meet, retired listeners refuse), nextest+golden HEAVY filters both updated, REQ-MEET-RUNTIME-ISOLATION + REQ-MEET-PORTMAPPER-OFF, ADR amended. Producer receipts: meet 1/1 (red control exit 100 canary 471 ms; restored 1/1), guards 253/253 x2, starve 2/2 x2, attach 17/17 x2, trace 926/926, clippy 0. doyle read: parking_lot dev-dep only; broker drop context sync; NetHost already owns a Runtime so drop-in-async was already forbidden.
- 2026-09-17 10:22Z: PR 231 opened (spt-bs-core, fix/302-meet-offload @a593ece0 -> main, fix-only, field acceptance pending). todlando field assembly 0f80fef8 (fix+diag) building in 302-meet-offload-field on the relocated warm pool; I gate a593ece0 by sequential takeover after his release. hertz standing by: 185 s collector -> field-fix-<sha>-operator + v2 + broker-tid >=250 ms predicate; gate = >=180 s span, 0 stale intervals, 3 subnets meeting.
- 2026-09-17 10:50Z: PR 231 CI 35209931657 = 5/5 SUCCESS (changes, traceability, unit Linux, unit Windows, lint); Windows leg 10:20-10:49Z under todlando's field build contention. todlando field assembly 0f80fef8: release compile exit 0 (4m12s), cohorts held on my ruling until this leg concluded; static bundle prep done (app PDB GUID 8291CB6B-28F4-4CC5-BD74-899BCC26EC46 age 1). Resuming him now.
- 2026-09-17 11:22Z: FIELD CAPSULE VERIFIED (artifacts-release-0f80fef8: manifest 5b668591…, spt.exe 73f22826…, spt.pdb 554e1979…, CodeView 8291CB6B… age1 by my PE parse, symbols/manifest 4d7c716e…, 16 PDBs incl. new spt.pdb). todlando handoff .spt/preserved/302/todlando-step2-field/handoff.txt: meet-diag 1/1 (meet tids runtime other, disjoint from broker tids), guards 253/253, starve 2/2, attach 16/16 + provisioned wedge 1/1, stack 3/3, clippy 0, trace 926/926; two honest exceptions retained (CI contention wrapper1; missing-fixture rig red). GATE LAUNCHED: .worktrees/gate-302-a593ece0 (detached a593ece0) on sequential takeover of the released field pool; receipts .spt/preserved/302/gate-231-a593ece0/. Next: transfer capsule exe+pdb to ENLYZEAM, prep symbols-0f80fef8; lifecycle step (daemon stop with live ball-b/engine-room) referred to operator.
- 2026-09-17 11:25Z: ENLYZEAM RESTAGE PREP DONE (no lifecycle): payload-0f80fef8 (exe 73f22826…, pdb 554e1979…, bundle manifest 4d7c716e…, hertz step2-staging json 184c74d0…) verified on-box; symbols-0f80fef8 = 15 OS PDBs (hash-verified vs bundle manifest) + new spt.pdb + manifest. Resident still diag 42904/47684 (log 47 MB), USER SYMDIR still symbols-3142753a, 13.1 GB free. Gate attempt 1 refused by SPT_POOL_FOREIGN (no claim on the released field pool; hatch required for the claim step) — archived under gate-231-a593ece0/attempt1-pool-refusal; attempt 2 running since 11:21Z with SPT_POOL_UNCHECKED=1 on pool-claim only. BLOCKED: daemon stop on ENLYZEAM with two live hosted agents — operator stopped it last time; asking operator to stop again or authorize 'spt daemon stop --force'.
- 2026-09-17 23:06Z: GATE a593ece0 PASS (attempt 2, .spt/preserved/302/gate-231-a593ece0/: pool-claim hatch, build 0, meet 1/1, starve 2/2, spt-net 73/73, daemon-int 29/29, clippy 0, trace 0, pool released). ENLYZEAM RESTAGED to fix+diag 0f80fef8 (operator stopped daemon; doyle swap 23:04Z; run dir operator-20260917T230446072Z-fix; old diag exe/pdb + 243 MB log preserved): broker 5752 run_id 101c51d4… tids 7052/9968, symbols-0f80fef8, both endpoints resumed. hertz dispatched (SENT): 185 s collector -> field-fix-0f80fef8-operator + v2 + broker-tid predicate. OPERATOR RULING 23:03Z: doyle may start/stop the ENLYZEAM daemon going forward (agents inactive, auto-resume). Next: hertz verdict -> if PASS: ff-push PR 231 (tested a593ece0 == merged), #302 -> acceptance via alchemy sweep, then W3 golden head -> deployah -> v0.71.0.
STATUS 2026-09-17T23:16Z: #302 STEP 2 field verdict FAIL (sync-span gate; user metrics all PASS). RCA: ordinary-node portmapper actor -> HomeRouter::new sync on spt-broker-net every ~25 s (STEP 1 finding (b) incomplete). STEP 3 ruled + dispatched to todlando: PortmapperConfig::Disabled on every endpoint, same lane. PR 231 NOT landed. ENLYZEAM stays on 0f80fef8. Ruling: .spt/preserved/302/step3-ruling.md; #302 comment posted.
STATUS 2026-09-17T23:33Z: DISK reclaim for STEP 3 admission (todlando HOLD at 96 GiB floor, free 59.7 GB). Census: .worktrees/*/target = field pool 59.3 GB (todlando's warm pool, kept) + 3 MERGED-lane targets (297-enforcement-shape 8.5, 297-rule-heavy-admission 8.5, release-S2-527cd8e8 1.4) reaped 55.6->71.9 GB; C:\actions-runner\_work\spt-bs-core\spt-bs-core\target 62.4 GB real dir, runner idle (last run 35209931657 10:20Z, only Runner.Listener up), reaped in 7 s 71.9->133 GB (142,792,941,568 B). Cost: next CI run cold-builds. todlando cleared for STEP 3 native admission.
STATUS 2026-09-18T00:30Z: 712c987f (STEP 3) reviewed+approved; CI 35289449769 4/5 green, Windows unit in progress. My gate run1 VOID (two drivers concurrent in one pool after a TaskStop'd background call survived; attach wedge red = SPT_ENDPOINT_ID=doyle leaked from perch env, scrub list omitted it). run2 single detached driver launched with SPT_ENDPOINT_ID scrubbed, receipts .spt/preserved/302/gate-231-712c987f/run2/. deployah #307 form pre-check PASS 6/6. hertz STEP 3 gate ready (step3-gate.py 40f804cb). Next: run2 PASS + CI green -> todlando builds fix+diag capsule off 712c987f -> stage on ENLYZEAM -> hertz STEP 3 capture.
STATUS 2026-09-18T00:37Z: 712c987f GATES GREEN — CI 35289449769 5/5 (Windows unit 31m47s cold) + doyle isolated gate run2 PASS (all legs, attach 17/17). Capsule requested from todlando (fix+diag off 712c987f). Then: doyle stages on ENLYZEAM -> hertz STEP 3 capture (same predicates + NET_GATEWAY_LOOKUP on spt-broker-net == 0, control 13) -> PASS => ff-push PR 231 @712c987f, #302 -> ACCEPTANCE, W3 golden head to deployah.
STATUS 2026-09-18T00:42Z: DISK reclaim ERROR (mine): at 00:39Z I reaped .worktrees/302-meet-offload/target believing it the finished remedy pool; todlando had relocated the STEP 2 FIELD pool into that path at 23:32Z (pool-relocation.json), so the warm release diag artifacts went with it. Census was stale (23:27Z) and the reap read no POOL-OWNER.json. No evidence lost (artifacts-release-0f80fef8 preserved). Cost: STEP 3 capsule builds cold. Free 125 GB. Capsule assembly 0611a058 committed locally by todlando; cold build in progress next.
STATUS 2026-09-18T01:30Z: STEP 3 capsule 0611a058 verified (exe 5211b9da…, pdb 65ce0cce…, manifest 15e721b2…, bundle b1cb1c27…; policy = unconditional Disabled) and STAGED on ENLYZEAM 01:28:39Z (stop --force of 5752 under the standing rule; backup operator-20260918T012834490Z-step3; rollback = its spt.exe.fix0f80fef8). New broker pid 20572, run_id eb9b43a2-e280-490f-83d2-f40334ff868b, tids 39708/37848. hertz authorized: 185 s collector + STEP 2 gate + STEP 3 additive gate. Verdict ~01:40Z.
STATUS 2026-09-18T01:46Z: #302 FIELD PASS (capture #2 on 0611a058: 185/185, 0 stale, canary 38, 0 broker sync spans, gateway 0/0; capture #1 INSUFFICIENT coverage, retained). LANDED: main 84e359aa -> 712c987f ff (tested==merged), PR 231 MERGED 01:44:24Z, branch deleted. alchemy sweep issued (#302 -> ACCEPTANCE). #302 acceptance comment posted. W3: golden head 712c987f handed to deployah for v0.71.0 (counter 107). Deferred cleanup: .worktrees/gate-302-a593ece0 (permission denied on remove, retry), .worktrees/gate-302-712c987f (stray target/ inside — classify then remove). ENLYZEAM stays on 0611a058 until v0.71.0 ships. Next: await golden result; on green, release close -> sweep docs/INFRA-REGISTER.md.
STATUS 2026-09-18T01:52Z: deployah intake: (a) he authors the v0.71.0 release shape on top of 712c987f (golden runs on the successor); (b) intake artifact (never-executed cells + changed-seam consumers, batch ddd7fc7f..712c987f) delegated to a background agent -> .spt/preserved/307/GOLDEN-INTAKE-712c987f.md; review then send to deployah. Orphaned dirs .worktrees/gate-302-a593ece0 and gate-302-712c987f (4 GB target inside) are pinned by my own tool shells' cwd (crates/spt-daemon); already out of git's registry; delete later when unpinned. Free 184 GB.
STATUS 2026-09-18T02:12Z: GOLDEN-INTAKE-712c987f.md authored by doyle (bg agent never produced it) and sent to deployah with cover. 7 first-golden int fns / 5 binaries; watch refresh_freeze (Win Python rig), disk floors 128GiB/130GB, traceable-reqs precheck. Waiting on deployah release shape + golden.
STATUS 2026-09-18T02:20Z: deployah corrected two cells (docs_server_e2e/twohost_web fns are integration, no CI evidence) — artifact fixed by replacement, count now 9 int fns / 7 binaries. Orphan dirs gate-302-a593ece0 + gate-302-712c987f REMOVED (3 orphan headless conhost pins, dead parents, my gate times — killed). Free 123.5 GB and falling = post-merge ci rebuild on the runner. deployah: release tree registered (.worktrees/release-307-v071 @712c987f), measures disk after main CI exits before Windows 128 GiB admission. Drafted register entries for release close: .spt/preserved/307/INFRA-REGISTER-draft-v0710-close.md. Memory: background-agent-does-not-survive-recharge. WAITING on deployah golden.
STATUS 2026-09-18T02:25Z: DISK reclaim for golden admission: 121 -> 150 GiB (+29), floor margin 22 GiB; reaped wit-what/target (idle since April), spt-core root target, capsule pool 0611a058; reserve = spt-alchemy/target 14 GB. Delegated intake agent DID surface (25 min after recharge) and overwrote GOLDEN-INTAKE-712c987f.md (mine -> .prev-doyle.md) with 3 corrections; superseding artifact sent to deployah. Memory background-agent-does-not-survive-recharge REPLACED (wrong: it survives, arrives late).
STATUS 2026-09-18T02:31Z: GOLDEN 35299506382 launched by deployah on f348ed0c (release shape on 712c987f; GO comment #307 5724230693; admission Win 157.8 GB / Linux 190.5 GB). Waiting; red returns to me RCA-first, no auto-retry.
STATUS 2026-09-18T04:25Z: GOLDEN 35299506382 RED (Win Phase B, refresh_freeze only, 241/242). RCA = rig ceiling: one tag 2016 ms vs 2000 bound after recovery; zero-tolerance sustained predicate; 4 prior PASS same signature; ambient box jitter 1.4-2.6 s in negative-control baselines; no CI co-tenant. hertz dispatched (brief in golden-35299506382/): repin predicate with replay proofs, thin PR off 712c987f. deployah holds; reshapes on successor. Same-sha rerun refused.
STATUS 2026-09-18T04:35Z: hertz branch test/307-refresh-ack-jitter @046bad44 (3 files: rig.py ack_recovery, test_predicates.py, refresh_freeze.rs unit wrapper). Doyle blind read APPROVED; my replay matrix 10/10 matches; unit 6/6. Ceiling noted: periodic stall period>10s passes (=#302 domain). Waiting hertz live cell -> my isolated second producer run -> PR -> ff -> deployah reshapes v0.71.0.
STATUS 2026-09-18T05:00Z: hertz handoff 0ab1e48b (author proofs: live cell PASS 21/22, replay 10/10, clippy/trace 0; receipts .spt/preserved/307/ack-jitter-repin/). My second-producer gate RUNNING (driver pid 56488, .spt/preserved/307/gate-307-0ab1e48b/) after 3 launch misfires (Start-Process unquoted -c payload; memory written). deployah shape queued behind ff-landed sha.
STATUS 2026-09-18T05:08Z: doyle second-producer gate on 0ab1e48b PASS (.spt/preserved/307/gate-307-0ab1e48b/: refresh_freeze 2/2, live cell 80.9 s, trace 0, pool released; attempt1 = unhatched pool-claim refusal, archived). Rig receipt numbers not retained (SPT_REFRESH_RIG_OUTPUT unset; retained only on failure) — hertz's live receipt carries them. hertz opening thin PR; next: CI 5/5 -> ff main (tested==merged) -> deployah fresh shape + golden.
STATUS 2026-09-18T05:36Z: PR #232 (hertz refresh ACK repin) CI 35309569389 5/5 -> ff main 712c987f..0ab1e48b, branch deleted. Golden head 0ab1e48b handed to deployah with intake delta + post-merge-runner warning. hertz stood down. Next: deployah fresh v0.71.0 shape on 0ab1e48b -> golden (~1.5-2 h) -> green: ship c107, release-close sweep (register draft ready), ENLYZEAM restage; red: RCA-first.
STATUS 2026-09-18T05:40Z: post-recharge resume. main 0ab1e48b unchanged; post-merge ci 35311316391 IN PROGRESS (05:34Z); no golden #2 yet. Register draft now 4 entries (IR-138 endpoint-id scrub, IR-139 pool-relocation reap, IR-140 refresh ACK ceiling BUILT + residual, IR-141 Start-Process unquoted payload discipline) — ids re-derived at apply (next free = IR-138 as of now). WAITING on deployah launch msg.
STATUS 2026-09-18T06:08Z: post-merge ci 35311316391 GREEN. deployah census 104 GB free (post-merge rebuild ate ~46 GB) -> asked me to reclaim for golden #2 admission. Reaped 3 target subtrees (my gate pool 38.9 GB under hertz-307-ack-jitter, deployah's released old release-307-v071 8.1 GB, spt-alchemy 14.1 GB idle): 95.76 -> 153.26 GiB; receipt reclaim-20260918T0605Z.txt. r2 pool + runner untouched. deployah shaping v0.71.0 in release-307-v071-r2 @0ab1e48b; WAITING on his golden #2 launch.
STATUS 2026-09-18T06:15Z: GOLDEN #2 35313912102 launched 06:12Z by deployah on aa8a2978 = 0ab1e48b + 1 shape commit (CHANGELOG/Cargo.lock/Cargo.toml/docs changelog; f348ed0c NOT an ancestor — verified). Admission #307 comment 5725967377 (Win 158.5 GB / Linux 190.4 GB). Expect verdict ~07:45-08:15Z. GREEN -> deployah ships c107, my release close (register draft 4 entries ready; alchemy sweep+release; ENLYZEAM restage). RED -> RCA-first.
STATUS 2026-09-18T07:55Z: GOLDEN #2 35313912102 RED (Win Phase B, sole red rc_attach_truth::resume_never_bound_reads_unbound_and_attaches; refresh_freeze repin PASSED both cells; Phase A 3471/3471, B 242/243). RCA golden-35313912102/RCA-resume-unbound.md: PRODUCT defect — brain liveness reconcile offlines an UNBOUND perch during a CLI-driven resume spawn (no custody minted on the CLI path => resume_in_flight false; CAS on session_id passes because the UNBOUND stamp preserves it); window = UNBOUND stamp..row_inserted, normally ~40 ms, this run 4 s (pty_created 2585 ms). No product delta vs golden #1 pass; first FAIL in 30+ records. Same-sha rerun refused. todlando dispatched (BRIEF-todlando.md): custody-on-CLI-path or census counts in-flight spawn; REQ-HAZARD + KNOWN-HAZARDS + deterministic repro. deployah holds; reshapes on successor. hertz on call.
STATUS 2026-09-18T08:00Z: golden #2 SECOND red: twohost-b (kitsubito) rendezvous timeout — twohost-a started 10m39s later, queued behind n1-gate Windows on the single-slot hfenduleam runner; golden.yml has no n1-gate->twohost edge (golden #1 paired by dequeue luck). RCA-twohost-b.md; hertz dispatched (needs: [test, n1-gate] on both twohost jobs + IR entry; acceptance = successor golden paired startedAt). twohost-a still running, will fail same rung ~08:00Z. Both fixes (todlando product, hertz workflow) land on one successor head; deployah reshapes once.
STATUS 2026-09-18T08:50Z: PR #233 (hertz) thin CI 35322019343: 4/5 green, Windows unit CANCELLED at 40:01 = ci.yml unit timeout-minutes 40, at test 3236/3254 — VOID leg caused by MY overlap authorization (todlando jobs=2 build + clippy on the same box). Golden #2 recorded terminal CLOSED_RED (#307 comment 5727154153). todlando: baseline red 3/3 + green 3/3 deterministic barrier cases proven; clippy/trace running. Sequence: todlando finishes clippy/trace -> runs timing-sensitive suites with runner idle (I hold CI) -> I rerun #233 Windows unit same sha uncontended -> ff d0d123dd -> post-merge ci -> todlando PR CI -> my gate -> ff -> deployah reshape. Lesson: compile overlap with a runner unit leg can push it past its job timeout; overlap only when the CI leg has >15 min slack.
STATUS 2026-09-18T09:18Z: #233 Windows unit RERUN fired (same run 35322019343, same sha d0d123dd) at 09:17Z uncontended — todlando's lane is past all cargo (final proofs on f27fc4d5: barrier green 3/3, rc_attach_truth 6/6, compat 125/125 incl. attach_idempotent_replay, clippy -D warnings, trace 927/927). Next: rerun green ~09:50Z -> ff d0d123dd -> todlando rebases f27fc4d5 onto it, thin PR, CI (queues behind post-merge ci) -> my isolated second-producer gate -> ff -> deployah fresh shape + golden #3.
STATUS 2026-09-18T09:36Z: PR #234 (todlando, f27fc4d5 on 0ab1e48b) blind review APPROVED: CliSpawnCustody (pid+birth identity, per-perch OS lock, fail-closed acquire, post-lock winner recheck, release-before-attach, Drop cleanup) acquired via before_unbound in resolve_home_and_write_skeleton for fresh/true-resume/fresh-fallback; reconcile untouched (comments only); REQ-HAZARD-CLI-SPAWN-UNBOUND-CUSTODY doc/impl/int tagged; KNOWN-HAZARDS 7.67; int test = real CLI + held broker spawn reply + in-process reconcile, 3 branches + dead-harness negative control. Pending: rebase onto hertz's ff, my isolated gate (rc_attach_truth 6, cli_resume_custody 3, driven_by_selfheal, attach_wedge_e2e, livehost pull_liveness_*, clippy, trace). #233 rerun in progress since 09:17Z.
STATUS 2026-09-18T10:00Z: #233 rerun GREEN 5/5 (Windows unit 09:17-09:51 uncontended). ff-pushed d0d123dd -> main (tested==merged), branch deleted. Post-merge ci starts on the runner now (~30 min). Next: todlando rebases f27fc4d5 onto d0d123dd -> new sha -> my isolated gate (driver template staged in scratchpad) -> PR #234 CI -> ff -> deployah reshape + golden #3.

STATUS 2026-09-18T10:12Z: todlando rebased #234 -> 71165e35 (parent d0d123dd; delta vs f27fc4d5 = 3 inherited non-Rust files; PR head updated, PR CI 35333276404 pending behind post-merge ci 35332370390). doyle gate LAUNCHED 10:11:29Z detached (pid 48836, quoted -c payload verified) in .worktrees/gate-307-71165e35 on sequential takeover of todlando released pool; receipts .spt/preserved/307/gate-307-71165e35/ (head.txt = sha, env scrub clean). Legs: pool-claim(hatched) build clippy custody rcattach wedge selfheal livehost storeunit trace pool-release. Overlap = post-merge ci only. Next: PASS + CI 5/5 -> ff 71165e35 -> deployah reshape + golden #3.

STATUS 2026-09-18T10:40Z: doyle gate 71165e35 PASS 11/11 legs (custody 3/3, rcattach 6/6, wedge 1/1, selfheal 2/2, livehost 38/38, store 12/12, clippy 0, trace 0; pool released; census-after cargo = runner post-merge ci, parent pwsh<-rustup, not mine). Gate worktree removed. PR CI 35333276404 queued behind post-merge ci 35332370390 (in progress). Next: PR CI 5/5 -> ff-push 71165e35 -> deployah reshape + golden #3.

NOTE 10:42Z: IR ids at main = ..137, 142 (hertz, landed d0d123dd). Draft ids 138-141 remain FREE => no renumber; IR-143 (overlap-timeout VOID lesson) appended to draft. Apply at release close.

STATUS 2026-09-18T11:08Z: post-merge ci 35332370390 Windows unit CANCELLED 10:51:12Z at 40m41s = timeout VOID; cause = MY gate (10:11-10:36Z) overlapping it (IR-143 mechanism, 2nd instance; overlap I had ruled acceptable because the run is informational: d0d123dd tested==merged via #233 CI). No rerun: main moves to 71165e35 next and its own post-merge ci follows. PR #234 CI 35333276404: 4/5 green, Windows unit in progress since 10:51:14Z uncontended (expect ~11:25Z). Box RAM 6.4/63 GB free = runner linkers + operator apps; two bg waiters killed by low-memory, re-armed one. Gate worktree removed.

STATUS 2026-09-18T11:29Z: PR #234 CI 35333276404 5/5 GREEN (Windows unit 10:51-11:19Z). ff-pushed 71165e35 -> main (tested==merged), PR MERGED 11:26:52Z, branch deleted. Golden head 71165e35 HANDED to deployah (reshape once on r2, exclude f348ed0c/aa8a2978, admit after post-merge ci exits, IR-142 startedAt record). todlando stood down; hertz briefed for IR-142 acceptance. Free disk 82G. Next: golden #3 verdict -> GREEN: deployah ships c107, my release close (register draft 5 entries IR-138..141+143 apply, alchemy sweep+release, ENLYZEAM restage 0611a058 -> 0.71.0); RED: RCA-first.

STATUS 2026-09-18T11:41Z: DISK for golden #3: census 11:29Z free 128G (runner target wiped by post-merge ci 35339675565 checkout, will regrow ~50 GB). Reaped 307-cli-resume-custody/target 72.6 GB (finished lane; stamp at reap = gate-307-71165e35 mine; 0 inbound, 0 live) 134.25 -> 210.10 GB; receipt .spt/preserved/307/reclaim-20260918T1136Z.txt. r2 + runner excluded. deployah/todlando/hertz all acked handoff; deployah shape edits ready, admits after post-merge ci exits. WAITING: post-merge ci exit (~12:10Z) -> deployah golden #3 launch (~1.5-2 h).

STATUS 2026-09-18T11:42Z: post-recharge resume. post-merge ci 35339675565: changes/trace/lint/Linux unit green, Windows unit in_progress since 11:27:05Z (expect ~12:01Z, then n1/twohost legs). Free disk 183 GB (>128 GiB floor). IR draft re-verified: 138-141+143 free at main (register has 130-137,142). No deployah golden #3 message yet. WAITING ~30 min, no cargo.

STATUS 2026-09-18T12:22Z: post-merge ci 35339675565 = VOID (Windows unit cancelled 12:07:24Z at 40m19s, ALL steps green 3254/3254, floor PASS, cleanup done). RCA .spt/preserved/307/deployah-intake-r3/RCA-postmerge-35339675565-void.md: box-wide CPU starvation 11:48-12:07Z, median per-test 1.88x vs PR CI 105562570404 (pure-compute crates 6-8x), 1384/3248 tests >2x slower; no agent cargo in window; contender = operator desktop load (1000xRESIST 2.3-4.1 cores + Ableton/OBS; box 51-64% CPU at rest 12:15-12:18Z, 16 cores). Golden refs: Windows test leg 67 min (#1) / 69 min (#2) vs cap 80; twohost-a 30-31 vs 40; n1 10.6 vs 40. RULING: golden #3 HELD (resource hold, not red). Admission predicate: sampler (.spt/preserved/307/box-sampler-golden3.log, pid 66172, 30 s cadence to 16:15Z) shows CPU <=20% sustained 5 min -> GO to deployah. Referred to operator: pause desktop load ~2.5 h / wait for quiet / approve cap raise (test 80->120, twohost-a 40->60) as IR-143 rider via hertz. 3rd timeout VOID today on hfenduleam (10:51Z d0d123dd, 12:07Z 71165e35, + gate overlap earlier) -> IR-143 draft gains: ci.yml unit cap 40 has 12 min slack over 28 min uncontended; golden test cap 80 has 11 min over 69. Peers asked for 11:45-12:07Z box activity; replies pending (not needed for ruling).

STATUS 2026-09-18T13:09Z: box quiet since 13:02Z (game exited; CPU 31-38%, privileged 21%/user 10%, Defender 0.4 + System 0.3 cores, attributed procs 1.4 cores; disk idle ~90%; RAM 15 GB free). Reference condition for golden #1/#2 leg walls (Ableton present, no game) => GO sent to deployah for golden #3 at 71165e35 r3. Sampler pid 66172 continues to 16:15Z. Note for IR-143: PR CI 105562570404 ran 28 min UNDER the game (07:17Z+), faster than the 32-35 min uncontended reference, so the game alone is not proven to be the 11:48-12:07Z contender; 20% predicate replaced by reference-condition test. No peer replies to the 11:45-12:07Z activity question (todlando/hertz/deployah silent 60 min). Next: run id from deployah -> job-timing watch at ~30 min cadence, foreground gh --json jobs only.

STATUS 2026-09-18T13:16Z: DEFECT IN MY OWN COMMS: zero MSG_OUT since resume (io-events 11809+) — three shortform @< tags (12:10Z peer question, 12:20Z HOLD, 13:09Z GO) never dispatched; no DISPATCH_RESULTS appeared in any briefing (the tell). deployah had no ruling 12:09-13:14Z. Resent via CLI 13:14Z: deployah GO (SENT), todlando/hertz FYI+question (SENT). Waiting: deployah run id.

STATUS 2026-09-18T13:18Z: deployah shaping build running in r2 pool; pre-build free 146.7 GB (136.7 GiB, 9 GiB over his 128 GiB floor); he remeasures after generation. hertz/todlando NONE for 11:45-12:07Z (RCA appended). Memory shortform-tag-silent-nondispatch-after-recharge written+indexed.

STATUS 2026-09-18T13:22Z: pool census (.spt/preserved/307/pool-census-13Z.txt): 304-release-note-s4/target 0 GB, hertz-S4-foreign-apply/target 0 GB (husks), root target 0.5 GB. NO reclaim lever left except deployah r2 pool (his) and runner target (cold-build cost). If his 128 GiB admission floor fails after the shaping build, the ruling is: admit anyway if free >= 100 GiB (in-run floor is 32 GB; runner target regrows ~50 GB; golden #2 ran from a similar level), else he trims r2. Waiting: deployah run id.

STATUS 2026-09-18T13:21Z: GOLDEN #3 LAUNCHED 13:19Z: run 35349501073, sha b25b2b77 (parent 71165e35 verified, "release: shape v0.71.0"), branch golden/307-v071-r3, intake GO #307 comment 5730568713; post-build free Win 145.0 GB / Linux 258.0 GB. Windows test leg started 13:19:29Z (ref 67-69 min -> ~14:28Z, cap 80 -> 14:39Z). Watch cadence: foreground `gh run view --json jobs` at ~13:50, 14:20, 14:35; sampler running. NO cargo on hfenduleam from anyone until golden exits.

STATUS 2026-09-18T13:52Z: GOLDEN #3 Linux test leg RED 13:27Z: 3427/3428, sole red relay_backlog_and_live_deliveries_stamp_once_even_after_hook_poll (5.8 s; passed 9-16 s in golden #1/#2). Presented asserts passed, receiver log has 1 MSG_IN (backlog) not 2. 71165e35 cli.rs diff = cmd_endpoint_run spawn custody only (not the listen path on its face). Repro 5x on kitsubito from runner workdir binary at b25b2b77 launched 13:52Z (bg). Run continues for evidence. deployah+hertz told.

STATUS 2026-09-18T14:06Z: golden #3 Linux RED RULED: product ordering (startup.rs deliver closure emit-before-publish, eb38b71a) vs test polling at the print edge (110c7c4c). Todlando lane dispatched (brief .spt/preserved/307/brief-todlando-relay-order.txt): reorder publish->record->emit, no test change, hazard REQ. No cargo on hfenduleam until golden #3 exits. deployah told: #3 runs to completion for evidence; successor = r4 + golden #4. Faithful kitsubito repro (bg bkn5uoaym) pending; rate only.

STATUS 2026-09-18T14:02Z: faithful repro 35/35 pass (rate low; ruling unchanged). todlando: reorder + 5-file rider written; sweep: cmd_send and agent poll journal-first; delivery.rs:700 drain = ADR-0020 shell transport not MSG_IN; PTY inject after broker verdict; ready has no journal — unchanged, reported. Golden #3 Windows test leg in_progress since 13:19:29Z (ref ~14:28Z, cap 14:39Z); n1 Windows queued; twohost after. Box cpu ~46% (Ableton, bun).

STATUS 2026-09-18T14:08Z: PR #235 (2f7df27a on 71165e35) diff reviewed OK; PR CI 35353733267 queued behind golden #3. GATE PLAN: Linux e2e (io_events_undriven_kinds_e2e, 5x) on kitsubito from the PR head via bundle into ~/spt-core-deploy (own target; kitsubito idle until twohost-b ~14:40Z; overlap with twohost-b allowed: 17 min ref vs 40 cap) + PR CI green (compile+unit both OS, clippy) + trace -> ff 2f7df27a (tested==merged) -> deployah reshape r4 -> golden #4. No separate Windows int leg (golden #4 test leg covers; rationale: 3-line reorder on a shared closure, PR CI compiles it on Windows). NO cargo on hfenduleam until golden #3 exits (~15:15Z).

STATUS 2026-09-18T14:10Z: kitsubito gate for PR #235 RUNNING: ~/spt-core-deploy at 2f7df27a (wide bundle f1a841afe.. 11 MB sha 87308388), cold target, script /tmp/doyle-gate235.sh -> /tmp/doyle-gate235.log (5x nextest io_events_undriven_kinds_e2e, env scrubbed). Overlaps PR CI 35353733267 Linux unit on kitsubito (5 min ref vs 40 cap: allowed). Two instrument faults on record: pkill -f / pgrep -f self-match killed my own ssh shell twice; first gate launch ran on stale head f1a841afe (deploy origin = shallow runner workspace) — killed by cwd, relaunched. Golden #3 Windows test leg due ~14:28Z. Pending: gate log, golden Windows legs, PR CI.

STATUS 2026-09-18T14:27Z: INCIDENT: kitsubito OOM 14:11:15Z (15 GB box; my cold gate build + runner Phase A) -> user slice + GNOME session killed, runner unit FAILED, GitHub runner OFFLINE. PR CI Linux unit 105627984731 exit 137 = VOID (mine). Gate #235 result: 5/5 runs 7/7 green (/tmp/doyle-gate235.log; deploy target 2f7df27a). Restoring runner; peers told; operator must be told (their kitsubito desktop session died).

STATUS 2026-09-18T14:29Z: kitsubito runner restored 14:26:07Z (sudo -n worked), GitHub online 14:26:59Z; operator GNOME session on kitsubito was OOM-killed (gdm on seat0) — operator to be told. Memory kitsubito-15gb-build-overlap-ooms-the-runner written+indexed. Peers acked; hertz preserved outage beside IR-142 evidence. PR CI 35353733267: Linux unit VOID (137, mine), Windows unit queued behind golden; rerun --failed after the run exits. Golden #3 Windows test leg 68 min in (cap 14:39Z).

STATUS 2026-09-18T14:43Z: golden #3 Windows test leg CANCELLED 14:40:00Z at 80m31s (VOID at cap; box baseline 31-38% pre-launch, sampler cpu 36-74% during). n1 Windows in_progress 14:40:02Z. PR CI Windows unit still queued. Peers told the sequence. Measuring phase stretch vs golden #2 leg.

STATUS 2026-09-18T14:50Z: golden #3 Windows test job 105614091181 = all steps SUCCESS, cancelled by cap at Complete job (80m28s) = VOID with green content. Step sums: build 8.6 / A 26.4 / B 31.1 / doctests 1.6 / clippy 5.3 / E2E 1.2 / docs 3.4 = 77.6 min vs green refs 55.4 (34957614641, 09-15) and 66.3 (34474627303, 09-10). Golden #2 69-min "ref" was a Phase-B-red leg (skipped tail steps) — my admission reference was wrong by class. RULED: cap raise lane -> hertz (brief .spt/preserved/307/brief-hertz-cap-raise.txt): golden test 80->120, twohost-a/b 40->60, ci unit 40->60, + IR-143 register entry. deployah told. Next: golden #3 exit; PR CI Windows unit; rerun Linux unit; ff #235; ff cap PR; r4; golden #4.

STATUS 2026-09-18T14:54Z: PR #236 (c7241ada) reviewed OK; its CI 35358690610 queued. Golden #3: n1 Windows green 14:48:18; twohost-b running (role B waiting) since 14:48:21; twohost-a queued behind PR #235 Windows unit (started 14:48:20, ~15:18 end) -> twohost-b likely VOID at 15:28 = INTERLEAVE-SKEW (new IR-142 mechanism: PR jobs interleave golden legs on the single-slot runner). Ruling: PR CI keeps priority. Runner chain: #235 unit -> twohost-a -> #236 unit -> post-merge runs -> golden #4 (~17:30Z+).

STATUS 2026-09-18T15:16Z: golden #3 twohost-b FAILED 15:05:35Z (peer never started; interleave-skew), twohost-a still queued behind PR #235 Windows unit (14:48:20Z, in progress 27 min). hertz told. Next: unit green -> `gh run rerun 35353733267 --failed` (Linux unit VOID) -> both green -> ff 2f7df27a.

STATUS 2026-09-18T15:18Z: PR #235 Windows unit GREEN 15:15:27Z. Linux unit rerun (attempt 2) launched 15:17Z. twohost-a (golden #3) running on hfenduleam since 15:15:29Z (interleave gap 27m08s recorded by deployah). Next: Linux green -> ff 2f7df27a.

GOLDEN #4 ADMISSION PLAN (15:20Z): (1) ff #235 2f7df27a on Linux green; (2) ff #236 c7241ada after its CI green (Windows unit runs after twohost-a ~15:45Z) — if #236 needs a rebase onto 2f7df27a for ff-only, hertz rebases (CI-only diff, register hunk independent of #235 files); (3) deployah reshapes r4 on the new main (both riders in); (4) EMPTY QUEUE gate: `gh run list --limit 5` shows nothing queued/in_progress on either runner (post-merge ci runs from both ffs must have EXITED), box CPU baseline <=40%, free disk >= 128 GiB, kitsubito free -g available >= 12 GB; (5) NO PR pushes, no reruns, no local cargo on either box from golden #4 launch until its terminal state (~2.5 h with raised caps). Reserved-runner rule from deployah adopted.

STATUS 2026-09-18T15:30Z: ff 2f7df27a -> main DONE (PR #235 MERGED 15:28:01Z; CI 5/5 incl. Linux rerun 15:17-15:24; my Linux gate 5x green; branch deleted). Peers told; hertz rebasing #236 onto 2f7df27a. Golden #3 twohost-a still running (15:15:29Z+). Next: #236 rebased head -> its CI green -> ff -> both post-merge ci exit -> deployah r4 -> empty-queue check -> golden #4.

STATUS 2026-09-18T15:33Z (post-recharge): #236 rebased head 629e33e1 VERIFIED (base 2f7df27a; range-diff = on both commits; 3 files +27-17). Its CI 35362720511 queued (traceability/changes queued). Post-merge ci 35362584091 for 2f7df27a: Linux unit in_progress on kitsubito since 15:28:34Z, Windows unit + lint queued on hfenduleam. Golden #3 twohost-a (hfenduleam) still in role-A ladder step since 15:23:12Z; b failed 15:05:35Z -> a fails at its 900 s pairing deadline ~15:38Z; NOT cancelled (self-terminates, keeps the A-side timeline for IR-142). Chain: twohost-a -> post-merge Win unit (~35 min) -> #236 lint+unit -> ff 629e33e1 -> post-merge #236 -> deployah r4 + empty-queue check -> golden #4 (~17:30Z). Box: sampler cpu 37-50% (one 87 spike 15:30), availMB ~12000. IR candidate for close: post-merge ci re-tests the exact sha PR CI already ran green (2f7df27a: 35353733267 5/5 then 35362584091) = ~35 min/box per merge, twice per milestone tail; ci.yml could skip push-main when the sha carries a green pull_request run.
STATUS 2026-09-18T15:51Z: golden #3 35349501073 TERMINAL failure (twohost-a failed 15:41:25Z on its own; not cancelled). Post-merge 35362584091: Linux unit + lint + traceability + changes green; Windows unit still queued (since 15:28:16Z). #236 CI 35362720511: traceability/changes/Linux unit green; Windows unit in_progress since 15:41:27Z (GH scheduler placed it ahead of the older-queued post-merge Windows unit). Next: ~16:16Z Windows unit green -> ff 629e33e1 -> post-merge Windows units run serially (2f7df27a then 629e33e1) -> ~17:30Z r4 + empty-queue -> golden #4. No local cargo on either box.
STATUS 2026-09-18T16:18Z: #236 ff DONE: main=629e33e1 (PR MERGED 16:17:01Z; CI 35362720511 5/5, Win unit 15:41-16:05 = 24 min; branch deleted). Post-merge 35362584091 Windows unit in_progress since 16:05:37Z. Auto post-merge for 629e33e1 next. Sampler box-sampler-golden3 ended 16:16Z -> relaunch for launch window. Next: both post-merge runs exit -> deployah r4 -> empty-queue predicate -> golden #4 (~17:15Z).
STATUS 2026-09-18T16:45Z: post-merge 35362584091 (2f7df27a) SUCCESS. Post-merge 35367547382 (629e33e1): all green except Windows unit in_progress since 16:31:07Z (~16:55Z). Sampler relaunched pid 45312 (16:18-20:18Z); cpu 53-69 while the leg builds, availMB 7-10k. Next: leg exits -> fresh admission measurements (hfenduleam disk>=128 GiB, cpu<=40%; kitsubito disk>=130 GB, free>=12 GB; queue empty) -> confirm reservation to deployah -> golden #4.
STATUS 2026-09-18T17:02Z: both post-merge runs SUCCESS; queue empty. deployah r4 shaped on 629e33e1 (build/gen/check exit 0). ADMISSION GO sent: hfenduleam cpu 28/34/36, 134.1 GiB free, cargo 0; kitsubito 13 GB avail / 174 GB / runner active. Golden #4 launch imminent; reserved-runner window ~2.5 h from launch. Next: run id from deployah -> watch legs (test 120-min cap now) -> GREEN: c107 ship + my release close (INFRA-REGISTER draft IR-138..141 + candidate same-sha post-merge redundancy; alchemy sweep; ENLYZEAM restage; memory v0710-published). RED: RCA-first.
STATUS 2026-09-18T17:08Z: GOLDEN #4 = 35372054865 at 7e35ba43 (parent 629e33e1, release shape only: CHANGELOG/Cargo.toml/Cargo.lock/docs-site changelog). Launched 17:03Z; test legs both started 17:03:42Z (cap 120 now); n1 x2 queued; twohost a/b gated on [test, n1-gate] (IR-142 barrier). Terminal est ~19:00-19:15Z. Pre-staged release-close docs PR: worktree .worktrees/docs-307-close, branch docs/307-infra-close, commit 3d33df1c on 629e33e1 (IR-138..141 + IR-144 same-sha post-merge redundancy; traceable-reqs check exit 0; blob LF). NOT pushed (reserved-runner). At close: rebase onto post-ship main, append IR-142 acceptance (golden #4 twohost startedAt pair + conclusions) and IR-143 "new-cap golden completion measured" lines, then push + PR.
STATUS 2026-09-18T18:04Z: GOLDEN #4 Windows test RED at Phase A 17:37:46Z: spt-daemon::sync two_tier (55.6 s) + concurrent_writes (20.3 s), sync.rs:109 stream-appear bound. Linux test, both n1 GREEN; twohost-a/b started 17:50:15/17:50:14Z (1 s skew, IR-142 timing met), running. RULED test-shape red (RCA golden-35372054865/RCA-windows-sync-stream-bound.md): iteration-counted 10 s poll vs requester's FS-bound pre-send; every green leg today passed the pair SLOW (91-132 s), red is the fast case. #235 not causal. Deployah: same-sha rerun --failed after twohost terminal. hertz: wall-clock bound rider (git-only until the rerun is terminal). Register at close: IR-95 lineage or IR-145.
STATUS 2026-09-18T18:22Z: golden #4 attempt 1 terminal shape: twohost-a GREEN 17:50:15-18:06:39Z, twohost-b GREEN 17:50:14-18:21:09Z (1 s skew) => IR-142 acceptance MET. Only red: Windows test (test-shape, RCA written). deployah told to rerun --failed now (attempt 2: Windows test ~80 min + twohost dependents ~30 min => terminal ~20:15Z). hertz rider c484ebb1 (test/307-sync-stream-wall-bound, parent 629e33e1) reviewed+approved, held git-only for proof after attempt 2 terminal.
STATUS 2026-09-18T18:27Z: attempt 2 of golden #4 in_progress since 18:22:35Z (Windows test ~19:45Z, twohost dependents ~20:15Z). Docs close commit amended -> db1a5fdc (docs/307-infra-close): + IR-142 DONE (hertz's close line, acceptance bullet reconciled), + IR-95 third-instance line. Pending at close: IR-143 "new-cap golden completion measured" line from attempt 2's green Windows leg wall; rebase onto post-ship main; push + PR after reservation lifts.
STATUS 2026-09-18T19:00Z: attempt 2 Windows Phase A GREEN 18:31:40-18:56:34Z (24.9 min; sync pair passed on a quiet box, cpu ~25%). Phase B since 18:56:34Z (~19:28Z), leg end ~19:40Z, twohost dependents ~20:10Z.
STATUS 2026-09-18T19:43Z: attempt 2 Windows test GREEN 76m28s wall (IR-143 measured; line folded into docs commit f3df8b55). twohost-a/b both started 19:39:04Z (0 s skew), terminal ~20:10Z. GREEN -> deployah ships c107 -> my close.
STATUS 2026-09-18T20:13Z: GOLDEN #4 attempt 2 TERMINAL SUCCESS 9/9 @ 7e35ba43 (twohost-b 20:10:50Z; paired starts 19:39:04Z, 0 s skew). deployah shipping c107 (ff/tag/release). hertz told: git free, cargo after post-merge+release build terminal. Next (mine): main=7e35ba43 + release published -> rebase docs/307-infra-close, push, PR, ff; alchemy sweep + release v0.71.0; ENLYZEAM restage; memory v0710-published + ledger; remove worktree; sampler ends 20:18Z.
STATUS 2026-09-18T20:55Z: v0.71.0 PUBLISHED 20:52:04Z (deployah; counter 107; 7e35ba43; 11 assets). PR #237 ff'd -> main=7ee5aa2d; branch deleted; worktree removed. hertz CARGO CLEARED (rider proof + PR next). alchemy sweep: nothing to reconcile; release v0.71.0 verb spooled. Next: ENLYZEAM restage; memory v0710-published + ledger; alchemy roundup read.
STATUS 2026-09-18T21:03Z: CLOSE DONE except board ack: alchemy release v0.71.0 -> 6 DONE; #307 close comment + state done sent (spooled). ENLYZEAM restaged 0.71.0 (spt node stop --force after DAEMON_STOP_REFUSED for 2 live sessions; node start; pid 1772; ball-b + engine-room ONLINE; backup preserved\307\operator-20260918T205630Z). Memory v0710-published.md written. hertz holding rider proof behind post-merge 35393844345 (docs-only push ran both unit legs — IR-144 note for next sweep). Remaining: MEMORY.md ledger line; final commune.