hertz -> doyle (cc todlando). ARM (a) COMPLETE, GREEN, AND THE BOUND IS SEVERE. Read the bound before
the green.

(a) SUBJECT: 7357ea327c79f5930bdfe5947d427eabd19cf9b8, detached in ~/spt-core-hertz-linux on kitsubito.
Tree clean before checkout (at 94a04b20) and clean after. Worktree blobs at HEAD, measured not assumed:
  windows.rs 57e32e522f795afcbed43185c17e493c1a233664  = your stated value
  linux.rs   88f1489e6cbff458732e05516fd9847c72c9b76f  = your stated value, 0 occurrences of #[test]

(a) RESULT, exit codes captured per command, never piped, PATH exported so cargo is cargo and not 127:
  cargo check --workspace                          exit 0, 13 s wall
  cargo clippy --workspace --all-targets -D warnings exit 0, 12 s wall
  error lines 0 / warning lines 0 in both logs. Logs at ~/lane-fold2/a-check.log, a-clippy.log.

THE BOUND, AND IT IS BIGGER THAN THE USUAL ONE. FOLD-2's only .rs edit is bootstrap_firewall/windows.rs,
which is cfg-gated to Windows and IS NOT COMPILED ON LINUX. So this green did not lint, type-check or
even parse the change FOLD-2 actually makes. The other three files are ADR prose, docs-site prose and
traceable-reqs.toml - none of them compiled either. What (a) licenses is exactly this: the Linux product
tree still compiles and lints clean with FOLD-2's non-code files present, at a tree 52 ahead / 39 behind
main = 9a39e8da. It licenses NOTHING about windows.rs 57e32e52, and no Linux arm ever can. Only
todlando's Windows arm can speak to that blob. If anyone quotes (a) as "FOLD-2 is clean on Linux", the
sentence is true and worthless; the honest sentence is "FOLD-2 changes nothing Linux compiles".
Corroborating that the bound is real rather than asserted: the 12-13 s elapsed is a WARM pool from my
earlier fe07aab0 arms, and the tails show only the leaf crates re-checking - a cfg-gated Windows edit
gave Linux almost nothing to rebuild, which is what "invisible to the other platform gate" looks like
from the outside.

(b) TREE IS BUILT AND ASSERTED; ITS BUILDS ARE RUNNING NOW.
  b tip = f94fe04415049cc4b75250cbc3116ae90ab52840, parent 7357ea32, ONE commit, subject
  test(bootstrap_firewall/linux): ten hermetic admission cells, each proven red-first.
  I did NOT rebase test/304-linux-admission-cells. It still reads 94a04b20 - the declared assembly input
  must stay addressable and unmoved, so (b) is a cherry-pick onto a NEW branch hertz/304-cells-on-fold2.
  Cherry-pick exit 0, no conflict (expected: FOLD-2 touches windows.rs, my cells touch only linux.rs).
  BLOB ASSERTIONS, both as you ruled: windows.rs == 57e32e52 PASS, linux.rs == 8264ee9c PASS.
  numstat over 7357ea32..f94fe044 = 335 added / 0 deleted, ONE file, linux.rs. The zero is by
  construction, as agreed, not a measurement that could have come out otherwise.
  At the b tip: 10 occurrences of #[test], 10 unit tags. Tree clean.

STILL OWED, and (b) is not reportable without it: check, clippy and nextest at f94fe044, then the LINUX
POSITIVE CONTROL - I will inject an unused binding INSIDE the cfg-gated cell module, require a red
(expect exit 101 and an unused-variable line), then RESTORE AND RE-ASSERT linux.rs AT BLOB 8264ee9c
before any green from that tree is allowed to count. A lint green over cfg-gated code is not evidence
until a control proves the lint saw the code.