# #331 SEAMLESS-UPDATES — golden-head HAND-OFF to deployah (doyle)

**HEAD: `8d402d6b8b5286ca937546b2feb22f200f5749e3`** on `gate/331-v073-r2` (origin). Parent = main `d3c90499`. Version material only on top of main.
You push the golden ref. I did not, and I will not (my misfire at 5507c487 is recorded below).

## 1. Greenlit-form delta (vs the greenlight snapshot)
- Record on #331: comment 5827919339. **ADDED #341 + #256** (W10), **DROPPED none**.
- #2 and #329 moved to ACCEPTANCE with reason comments 5827914740 and 5827914967 (#2 arm 1 measured green; #329's close rides #336).
- #305 was operator-attached before intake, so it is part of the baseline (W9).
- 20 sub-issues, all in ACCEPTANCE. Re-derive with `gh api …/sub_issues` before you run.

## 2. Release shape (a CHECKED leg)
- `git show 8d402d6b:Cargo.toml | grep -m1 '^version'` → `version = "0.73.0"`
- `git show 8d402d6b:CHANGELOG.md | grep -m1 '^## \['` → `## [0.73.0] - 2026-09-25`
- **Re-shaped fresh** on the respin tip. 8d402d6b does NOT descend from the discarded 5507c487 (`merge-base --is-ancestor` false).
- **W9 release-note line is IN the changelog** (top of the section): restart the daemon once (`spt update --restart`) to move docs serving to the brain. W9 is BRAIN_ONLY by ruling (JIT-PLAN W9).
- Counter: v0.73.0 = 109 (follows published 108). Compatibility constants unchanged.

## 3. Never-executed cells + changed-seam consumers
- Lists: `.spt/preserved/331/doyle/intake-5507c487/handoff-lists.md`. The batch adds 153 cells: 26 int cells in 14 binaries, 127 units. The 2 MUST-RUN-FULL crates are `spt/tests` and `spt-daemon/tests`.
- **Discharged BEFORE hand-off, at this exact head 8d402d6b, both OSes, full `(package(spt)|package(spt-daemon)) & kind(test)` in golden's Phase A/B shape:**
  - Windows (hfenduleam, doyle, OWL_/SPT_ env scrubbed): A 206/206 (2 leaky: reaper_guard, false_promote); B **243/243** (3 slow), 1207.8 s. Logs: `doyle/intake-8d402d6b/r2-win-{a,b}.log`.
  - Linux (kitsubito, hertz, 0 agent vars): A 199/199; B **235/235** (2 slow), 1607 s. Record: `hertz/gate331-linux/` (r2 report).
  - So every cell the lists marked "never" on either OS has now executed green at this head (hertz's report names the first-ever Linux cells one by one).
- **Still last run at 73f19d51 — golden-only legs:** n1-gate (old broker + new brain: exactly W9's accepted transition. A new brain on an old broker must log EADDRINUSE loud while the old broker keeps serving, and the W9 field arm already pins this), twohost-a/b, notify, oneliner (source not read: UNKNOWN whether it touches serving), doctests, budget --release. This golden is their first run on the batch.
- **Phase B shift to expect:** 5 new HEAVY binaries (adapter_peer, source_ladder, auto_apply, busy_window_axis, serving_in_brain), plus 8 binaries moved to Phase B by IR-156 (7c315d52). Linux delta measured by hertz; Windows unmeasured before this golden.
- **auto_apply_e2e** Linux ≈ 263 s (3 samples), under #282's 540 s per-binary backstop; Windows ≈ 82 s.
- **Watch items, not reds:** (a) r1 Windows `job_escape_e2e` (reachable=false) and `refresh_freeze` (sustained ack false) failed ONLY in an env-contaminated run (my agent shell's OWL_/SPT_ vars). Both went green in the scrubbed rerun and at r2. **No mechanism read**: if either reds in golden, it is not "known flake". (b) Linux `transport_death_eof` took 65.5 s at r2, just past the slow period; not read.

## 4. What changed since the first candidate (5507c487) and why
- r1 full-crate run found 3 reds on Linux / 1 product red on Windows:
  - `webserve_attachment_e2e` = **W9 product race**, both OSes (autostart readiness = seed ping; port only after the brain announces). Fixed by **#283** (todlando): DocsStatus `pending` + `{gen,0}` failure announce + CLI-only 15 s bounded wait → `DOCS_LISTENER_PENDING`. REQ-WEB-BOUND-PORT-GENERATION-SCOPED amended in place. Red-first shown at 456c8554.
  - `source_ladder_e2e` (tar hidden with gh) + `auto_apply_e2e` (240 s nextest backstop) = rig defects. Fixed by **#282** (hertz).
- #282 landed at 0c467bfc. Its PR CI Windows `DISK floor end` red was a RESOURCE red caused by my gate pool, and `Unit tests` was success (ruling comment 5829221926). #283 landed at d3c90499. Both by ff, tested == merged, and post-merge CI is green on both.
- Record: `.spt/preserved/331/golden-intake-5507c487.md` (full trail).

## 5. Compile gate (isolated .worktrees/golden-331, fresh pool)
- `cargo build --workspace --all-targets` at 8d402d6b: exit 0, 6m51s. `traceable-reqs check` exit 0. `xtask gen`: 0 drift.

## 6. Pool reap + disk floor (IR-145)
- **Gate pool** `.worktrees/golden-331/target` (compile gate + both full Windows runs), reaped 09:49:40–09:52:19Z. Outbound: a real dir. apparent 92,527,127,232 B; allocated 92,620,314,624 B; escaped 0. Free C: 34,142,511,104 → 126,750,789,632 (+92,608,278,528 = allocated −12 MB). Census: 0 cargo/rustc/Runner.Worker in the window.
- Floor after that reap: **NEGATIVE** (126.75 GB vs 137.44 GB), so I did not hand off.
- **Runner workspace target** `C:/actions-runner/_work/spt-bs-core/spt-bs-core/target`, reaped 09:59:48–10:00:06Z with 0 queued runs and 0 Runner.Worker/cargo. apparent 69,992,327,762 B; allocated 70,053,636,096 B; escaped 0. Free C: 126,750,687,232 → 196,769,304,576 (+70,018,617,344 = allocated −35 MB). **Consequence: golden's Windows legs build COLD** (expect several extra minutes before Phase A).
- UNEXPLAINED: between 08:20Z (178.4 GB free) and 09:49Z (34.1 GB), about 51.7 GB went to writers other than my pool. The Everything index attributes 12.8 GB to System Volume Information (dated today). The runner target was rewritten but did not grow (70.0 GB both times). About 39 GB has no attribution, and I captured no census for that window, so it stays UNEXPLAINED. It is not assigned a mechanism.
- **Floor line at the head** (`git show 8d402d6b:.github/workflows/golden.yml`, `DISK start floor (Windows)`): `$floorBytes = 128GB` (= 137,438,953,472 B)
- **Free C: 196,803,489,792 B at 2026-09-25T10:00:16Z → margin +59,364,536,320 B (+59.36 GB).**
- This is an INSTANT reading: re-measure at push time. The unexplained ~39 GB draw above says the margin can move.

## 7. Owed at release close (not golden blockers)
- **Field acceptance on #341:** a real 5-minute detach that produces the away notice.
- Sweep `docs/INFRA-REGISTER.md` at release close (my standing duty). New candidate from this intake: auto_apply_e2e's Windows LEAK (inferred: handle inheritance through the apply chain; seen once at r1, absent at the #282 gate and r2).
- My misfire: pushing `golden/331-v073` fired golden run 36104664271 before intake. It was cancelled in about 2 min and the ref deleted. Ignore it.