---
name: v0450-published
description: spt v0.45.0 (release counter 80) published 2026-07-27 @899f466 — FIELD-TRUTH W1 build set; lessons on parallel thin-lane tails and seam-touching CI reds
metadata: 
  node_type: memory
  type: project
  originSessionId: 05ad48d2-2229-4f67-9c5b-7f322bec6a2e
  modified: 2026-07-28T00:18:29.374Z
---

**v0.45.0, release counter 80, published 2026-07-27, tag on main @899f466** (PR #112).
MINOR. Cuts the FIELD-TRUTH W1 ranked-findings build set.

- Basis: main @002ab1b, run 30312823013 green on both hosts (verified independently, not
  on doyle's word — see [[verdict-from-exit-not-from-silence]]).
- MINOR justified by the changelog's own test: two EXISTING surfaces change observable
  behavior — `spt endpoint stop <id>` now exits non-zero on an id no source knows
  (previously success for any string), and `spt daemon status` re-keys its out-of-date
  warning to the coordinating component. doyle ratified after being shown the reasoning.
- Entry-bearing PRs: #95 (stop resolves its id), #104 (`--active-only` never injects),
  #105 (coordinator image + re-keyed warning), #106 (CLI-side launch rides the daemon;
  late-ack arbiter scoped to its one cause), #107 (relay-death convergence, create-bind
  rest normalization, zombie demoted at the identity oracle). Registry/test/CI/docs PRs
  in range correctly carry no entry.
- Counter read from the published `update-set.json` at v0.44.0 (`"version":79`), never
  recalled — [[release-counter-from-published]].
- Lock bump verified by DIFF: exactly 12 first-party pairs, no third-party line touched.
  `xtask gen` produced no CLI reference drift (the `reference.md` "modification" was
  CRLF noise — an empty `git diff` on a file git lists as modified means line endings,
  not content).

**⭐ Lesson — a docs-only tail is free parallel work, but its MERGE is not.** The runbook
drift tail (#113) was authored and pushed while the release PR's CI ran: docs-only outside
`docs-site/` classifies thin-lane, so it consumed no heavy runner slot and cost the release
nothing. But doyle held its MERGE until the tag landed — merging it first would have moved
main HEAD, so the tag would have covered a commit the release PR's CI never validated. The
rule that falls out: parallel tails may LAND CI freely, may not LAND ON MAIN, between a
release PR's green and its tag.

**⭐ Lesson — a shared-box build window is not a quiet window.** Held `xtask gen` and every
local cargo call while a CI job ran on hfenduleam ([[seedmap-test-collides-live-daemon]],
[[quiet-window-predicate]]). The `cargo metadata --offline` step is safe there (no compile);
the gen build is not.

**⚠ Open at publish — CI red on the seam this release shipped.** The `push:[main]` backstop
at @899f466 went red on `wake_single_flight::two_concurrent_wakes_for_one_endpoint_spawn_exactly_one_launch_tree`
(Windows, 1st sighting, sole fail of 2085). doyle triaged it suite-inert — tree
content-identical to the PR-green sha, nothing merged between, and `release.yml` was
building both platforms on the same boxes = manufactured load. Sound as far as it goes, but
that test sits ON the launch/wake single-flight seam #106 just changed, and the runbook's
own preflight clause says a failure inside a touched seam is not filed as a load flake on
one observation. Flagged to doyle: treat the at-sha rerun as the first leg of a
repeat-proof, and route a second sighting to hertz as a defect, not a load note.

Publish mechanics unchanged: raw `SPT_RELEASE_SEED` (`_CMD` confirmed unset) per the
HFENDULEAM operator ruling; all three binaries verified against `SHA256SUMS` before signing.