---
name: v0630-arc-in-flight
description: v0.63.0 FIELD-SEAL release arc — SHIPPED 2026-08-26 counter 98; three respins, three test defects, and a publish that failed while the harness reported success
metadata:
  type: project
---

v0.63.0 **FIELD-SEAL** (milestone releases#225, publish counter 98) — **SHIPPED 2026-08-26T01:35:09Z.**
tag == main == golden-tested sha == `dbe3daad`. Board CLOSED: all 8 members `state: DONE`, #225
`state: DONE`, roundup posted. Counter 98 verified by DECODING the published signed metadata (97 ->
98); Latest endpoint flipped; 11 assets; `brain_ipc_version`/`broker_resource_abi` both unchanged at 1.

**Intake VERIFIED by me against the board, not the hand-off:** #225's GitHub sub-issues = exactly
the eight greenlit members (#222 #223 #219 #220 #221 #224 #216 #215), zero dropped, zero added, all
`state: ACCEPTANCE`. Greenlight record = comment 5409127849 (operator delegated selection to doyle).
#224's scope amendment is on the ISSUE (comment 5410154849) — the form the runbook requires. Docs
rider `f7137293` (EVENT attr pass-through) is an assembly rider, NOT a member (IR-30/USHER precedent).

**Version material @ `de91ee62`**, branch `release/v0.63.0-shape`, authored on doyle's head `5a9d86ee`
(base main `3865afd2`). Three files, zero `.rs` — same diffstat shape as the v0.62.0 precedent:
workspace bump + the 14 first-party `Cargo.lock` pairs (verified BY DIFF, not by counting version
strings), `[Unreleased]` cut in place to `## [0.63.0] - 2026-08-25`. **`xtask gen` deliberately NOT
re-run**, on a measurement: no generated page under `docs-site/src` or `docs/` carries a version
string at all, so a bump cannot drift them. CONTEXT.md and traceable-reqs.toml mention 0.62.0 as
HISTORY and are correctly left alone — bumping them would make true sentences false.

**Golden r1 `32856409565` @ `de91ee62` = RED, 8/9 green.** Sole failure: `spt-daemon::input_ack_deadlock
input_flood_through_serve_attach_does_not_deadlock_broker`, Linux only (Windows PASS 2.175s, Linux
FAIL 9.880s, same sha same run). **VERDICT (doyle): TEST DEFECT, pre-existing since v0.61.0
`53be5270` — product NOT implicated, FIELD-SEAL range exonerated.** See
[[a-tests-banner-can-go-stale-against-its-own-assert]] for the mechanism.

**Respin path:** hertz test-only fixup on top of `de91ee62` → doyle gates → I cut a FRESH ref
`golden/field-seal-225-r2` (never the spent `golden/field-seal-225`). **No re-author needed**: a
test-only fixup takes no CHANGELOG entry and no version change, so the bump stays the parent and
tag == main == golden ref still lands one object.

**Disk, same session:** the box hit 2 GB free against the 32 GB CI floor. Reaped five
merged-and-handed FIELD-SEAL pools, **148.48 GB measured** (predicted 163.9 — the FS delta governs).
Root cause is NOT ours: a 614 GB Steam library + 230 GB Downloads dominate a 1.86 TB drive, so this
recurs next milestone regardless of pool hygiene. Those five worktrees kept their CHECKOUTS but are
now UNCLAIMED — resuming one needs `pool-claim` from that worktree first. Filed at INFRA-REGISTER
IR-31 fifth addendum; doyle files the disk-floor mechanism entry separately (division agreed, no
double-file).

## How it actually shipped (final)

FOUR golden runs. r1 `de91ee62` RED, r2/r3 `561d5c5d` RED, r4 `dbe3daad` **GREEN 9/9**. Three
DIFFERENT Linux/kitsubito victims, each run’s prior victim recovering; r2 and r3 the SAME SHA, which
is what exonerated the tree. All three were TEST defects — the product was never implicated once.
Windows leg, both n1-gates, both twohost legs, traceability and changes: GREEN on all four runs.

⭐⭐ **THE PUBLISH FAILED WHILE THE HARNESS SAID IT SUCCEEDED.** Background-task notification:
"completed (exit code 0)". The leg’s own exit FILE: **101**. The subshell+echo wrapper exited 0 over a
failed build. Writing the leg exit to a file is the only reason I did not announce a release that was
still a draft. Fifth face of the rule, first time it mattered on an irreversible step.

⚠ The failure itself was our own guard: `SPT_POOL_FOREIGN`, main `target/` still owned by the VANISHED
`.worktrees/assembly-signet-218`. Verified draft=true/7 assets before touching anything — nothing was
signed or flipped. Takeover justified on all three arms (worktree vanished, branch `feb4049a` ancestor
of origin/main, no lane claimed), hatch as the build script prescribed, `pool-release` after.

⚠ **A CACHED BUILD SCRIPT REPLAYS ITS STORED WARNING.** The SUCCESSFUL publish still printed the
`SPT_POOL_UNCHECKED` override text naming the old worktree — 0.85s run, build.rs never re-ran, flag not
set. Anyone grepping that log later sees an override warning on the shipped commit and is wrong.

⚠ Board close is DRIVEN, not swept. `spt shell cmd alchemy-0 release v0.63.0` promoted the 8 and
posted the roundup; the milestone needed a separate `state '#225' done` — and the ref MUST be `#N`
(bare `225` errors). **`SHELL_SPOOLED` returns exit 0 regardless of whether the op succeeds** — the
bad-ref error arrived as a message afterwards. Verify the BOARD, never the spool receipt.