[Bizuayeu]

What this changes

test-glm-snapshots-host was the one GLM host step left out of test-cuda-host: its run was wired only to its own step
in zig/build/glm_family_host.zig, so zig build test never built it, and the break from #2226 stayed red on main
until #2403 (#2095 lists it among the steps red on main, under #2031). The run now also joins the family's aggregate
with step.dependOn, as the family's other host steps do, so CI builds it and runs its 21 tests. Its host-cache test
declared libc's setenv and unsetenv itself, which Windows' C library lacks, so the zig-windows job failed to link
it; it now sets TF_GLM_HOST_CACHE through sys.env, as cuda_snaps.zig's own test does.

• Issue: refs CI never reaches 246 of the 393 named test steps, and nine of them fail on main #2031, one of the named steps CI did not reach; the follow-up suggested in
  glm cuda: test-glm-snapshots-host's copies module imports core, which cuda_snaps.zig reads for TF_GLM_PROMPT_TAIL #2403 (comment). ci: collect native Linux host checks and guard exclusions #2095's test-host would reach the step too;
  with glm cuda: test-glm-snapshots-host's copies module imports core, which cuda_snaps.zig reads for TF_GLM_PROMPT_TAIL #2403 in, its exclusion entry for this step can go. This PR leaves ci: collect native Linux host checks and guard exclusions #2095's list alone.
• Stacked on: main 1b58f2e98, with glm cuda: test-glm-snapshots-host's copies module imports core, which cuda_snaps.zig reads for TF_GLM_PROMPT_TAIL #2403 in.
• atc overlap --json: at 2026-11-07 06:47 UTC, on this head against main 9b15a958d: glm cuda: EXL3 mcg experts at any whole-bit width, mixed by projection #2094 and glm cuda: TF_GLM_GRAPHS=1 replays verify windows as CUDA graphs (#1490) #2279 overlap and
  merge cleanly, no conflicts; incomplete (pullFilesUnavailable, timeLimit). glm cuda: a chat's next turn can resume at the previous turn's history (opt-in, #2396) #2404, which adds a step beside this
  one in glm_family_host.zig, merges with it cleanly in that file (git merge-tree).

Checks on this head

• Head commit: b04769ece
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 116 and 116 problems, none new.
• Tests run, with the command and pass, fail and skip counts: zig build test-cuda-host 327/327 steps (5,332 of 5,357
  tests, 25 skipped) on this head and 325/325 steps (5,311 of 5,336 tests, 25 skipped) on main 1b58f2e98; with CI's
  flags, zig build -Dfixed-cpu test 561/561 steps (7,647 of 7,708 tests, 61 skipped) on this head and 559/559 steps
  (7,626 of 7,687 tests, 61 skipped) on main. The 21 more in each are the snapshots run's.
  zig fmt --check build.zig zig.
• CI: all seven checks pass on this head (zig, zig-windows, zig-macos-dev-tools, release-paths, fmt and both
  benchmark-clients), and both zig and zig-windows run the snapshots tests in zig build -Dfixed-cpu test (their
  GLM snapshot sizing lines are in both logs). On the first head, 0790b76d0, zig-windows failed to link them
  (setenv, unsetenv); the second commit fixes that.
• Not run, and why: the GPU suites, which a build file's list of host steps does not reach; Windows, which only CI's
  zig-windows job runs here.

Receipt

n/a: a build file's list of host test steps; it cannot reach inference.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new tests: the
  step's 21 tests now run under test-cuda-host and test.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.