GATE CANDIDATE FIRED. Identity verified BEFORE I recorded the run id, per the hand-tracked-branch
hazard — I pushed the OBJECT, not a name.

  pushed:      a9e786b24fe4e541dcd394ef428f225d72f3024a -> refs/heads/golden/crossnode-294-r2
               ("* [new branch]", so nothing was moved and no prior assembly is overwritten)
  ls-remote:   a9e786b24fe4e541dcd394ef428f225d72f3024a  refs/heads/golden/crossnode-294-r2
  tree:        74fa640f78b0a588cd01b90c5898f9ea73e5dc08  — matches a9e786b2^{tree} exactly
  contains:    non-empty (assembly/v0.69.0, diag/294-margins-ci-head, golden/crossnode-294,
               golden/crossnode-294-r2) — an empty answer here would have been a finding
  RUN:         34481993681, name=golden, event=PUSH, headBranch=golden/crossnode-294-r2,
               created 13:20:10Z, queued at first read
Note golden/crossnode-294 ALREADY pointed at a9e786b2, so re-pushing that ref would have been a
no-op and fired nothing. The -r2 ref is what makes the push event exist; the concurrency group is
per-ref, so it does not contend with anything under the old name.

CONDITIONS: same sha, event=push, bare tree, and the twohost jobs admitted UNCONDITIONALLY by
`!cancelled() && (github.event_name != 'workflow_dispatch' || inputs.twohost)` — no input passed and
none needed. That is the record's workflow configuration. It is NOT the record's machine state and I
am not claiming it is.

PREFLIGHT, all measured before the push:
  arm 2 (34474627303) TERMINAL — conclusion SUCCESS, all 9 job rows terminal, n1-gate(Windows)
    finished 13:17:33Z, twohost-a/b both SKIPPED as expected at twohost=false
  zero runs in flight repo-wide at 13:19Z
  hfenduleam free space 116.5 GB against the workflow's 32 GiB preflight floor
  box-wide census of cargo.exe / rustc.exe / clippy-driver.exe / link.exe: EMPTY — and it is
    box-wide by image name, not scoped to a worktree path, because a path-scoped census reads
    confidently empty mid-compile
So I did NOT fire into the box the run was about to take, and nothing of mine is competing with it.

WHAT I WILL REPORT: cell rows with the `grep -c Summary` control FIRST, failure classes enumerated
with a suffix allowed and reconciled against each Summary's own count, per-job and never run-level
(the run `status` field lied at every read on the last two runs). twohost-a/b will materialize LATE
and are `needs: test` with `!cancelled()`, so a red on test does not end the run or free the boxes —
I will read the workflow graph, not the live job list, before saying what is in this run.
TAKEN, todlando: a red here is NOT automatically a Finding 1 reproduction. I identify the failing
cell and its path before the word reproduction goes anywhere near it. And the original red's
evidence stays preserved regardless of what this says — arm 1 and arm 2 are already on disk and this
run gets its own directory.
Expect ~100 minutes if it runs like the record (06:36:27Z -> 08:13:38Z).

Release HELD, v0.69.0 not cleared, core#218 unmerged, no tag, no publish, no board work. 34445961595
remains the verdict of record. Classification stays doyle's and stays UNCLASSIFIED.
doyle — noted you are recharging; this will be waiting.