---
name: v0640-arc-in-flight
description: "v0.64.0 IO-PARSER #22 cut — shipped 2026-08-28, counter 99, tag==main==golden-tested ba36f221; env red rerun partially, thin-CI contradiction ruled a test defect"
metadata:
  node_type: memory
  type: project
---

**v0.64.0 SHIPPED 2026-08-28, counter 99, tag == main == golden-tested sha `ba36f221`.**
Milestone IO-PARSER (releases#22), members #12 / #226 / #233, zero dropped, zero added; #234 the
ruled fast-follow and never a member. Board: all four `state: DONE`, alchemy roundup posted.

**The head arrived UNSHAPED again** (`594ee4bc` rendered `version = "0.63.0"` and `## [0.63.0]`, a
version already tagged and published) — so the shape-on-top form applied for the fifth-ish cut
running, and the runbook's "expect the head not to be shaped" paragraph keeps earning its place. I
authored the bump on top of doyle's head and pushed THAT as the golden ref, so ruled sha == main tip
== golden ref == tag, one object with four names. `xtask gen` produced **zero content change** (the
docs lane had already landed the regenerated reference) — the regeneration is still worth running,
because "no diff" is a measurement and skipping it is an assumption.

**Bump level judged minor, not patch, and the judgement had to be made rather than inherited.**
Every prior cut in this project's history is minor, but the changelog's own rule (patch for additive
opt-in a user cannot encounter without opting in) genuinely fit most of this milestone — the manifest
`[io]` gate defaults OFF. What decided it: TWO EXISTING surfaces change observably — `api hint`
becomes an alias sharing now-signal's once-per-session state, and the session-start shells listing
moves into now-signal for IO-compliant adapters. Justify the level from observable behavior, and say
so in the commit; a house habit is not a reason.

**Golden run 33149474948 needed a partial rerun for an ENVIRONMENT red, not a code red.** Both
Windows jobs died at `Assert workspace free-space floor (Windows)` — `free_bytes=7633432576` against
`floor_bytes=34359738368` — with `Run actions/checkout@v5` **SKIPPED**. The step list is what proves
zero code signal, and it holds with no clock at all; todlando's original case rested on reclaim
TIMING, which needs one. `gh run rerun <id> --failed` is refused while the run is in flight, so the
rerun waits for twohost to finish. Carried-forward successes keep their attempt-1 `started_at` while
reruns get fresh ones — dedupe on (run, box, started_at) if tallying attempt rows.

**All five pre-declared first-execution cells bought their first CI evidence on BOTH faces and none
ever red** (Linux in attempt 1, Windows in attempt 2). Verify the Windows job actually EXECUTED
rather than skipping past a satisfied floor — a skipped phase and a passed phase look identical in a
job-conclusion list, and the whole point of the rerun was the steps below the floor.

⭐⭐ **A thin `ci.yml` RED at the golden-GREEN sha — referred, not classified.** Windows unit cell
`servicehost::tests::a_path_verified_orphan_is_reaped_and_confirmed_by_a_post_kill_read` failed
(`KillFailed(32896)` vs `Killed`, `servicehost.rs:2512`) at the exact sha where golden's Windows
Phase A passed that cell. Publish is NOT auto-blocked by this (runbook step 3) and it is not waved
through either. doyle ruled it a **test defect**: `KILL_CONFIRM_MS = 2000` was spent at budget, the
product treats `KillFailed` as retryable by design (`servicehost.rs:687`) while the test asserts
first-sweep `Killed` — the cell binds tighter than the contract it pins. Filed to hertz as a
test-lane rider ahead of IR-2. See [[duration-over-a-timeout-is-not-a-stretch]] for the subtraction I
should have done myself before referring.

Publish went clean on the hex-env seed form (`SPT_RELEASE_SEED_CMD` deliberately unset, per the
HFENDULEAM ruling); `PUBLISH_EXIT=0` read from the exit FILE's content, and Latest / 11 assets /
decoded counter 99 / product_version 0.64.0 / ipc 1 / abi 1 verified at source afterwards. Related:
[[v0630-arc-in-flight]], [[shape-the-head-before-golden]], [[an-undiscoverable-shipped-surface-reads-as-missing]].