[devopsfortunato]

What this changes

The Qwen4 host gate and the CUDA host family's host gates ran their synthetic-graph arithmetic in Debug, so their checks spent minutes in an unoptimized path while their compile took seconds. Both now build in ReleaseSafe, which keeps every runtime safety check (integer overflow, bounds, unreachable, std.debug.assert) and only lets the optimizer work. No test source, no engine code and no served behaviour changes.

• Issue: Refs CI: zig and zig-windows spend most of their time rerunning tests that can't have changed #2242 — its plan item 3, "Build the heaviest test binaries in ReleaseSafe, where running them costs more than compiling them"; CI: zig and zig-windows keep main's build cache, and the quick checks run first #2140 named this gate as the next thing to fix ("zig/qwen4_family.zig alone takes over 6 minutes on Linux and over 3 on Windows").
• Stacked on: main, 610c67bf7.
• atc overlap --json: 3 changed files; the snapshot is complete: false (its own invalidPulls and pullFilesLimit limitations, so atc claim refuses every write repo-wide). One entry reported conflicts with qwen3-asr: host slice, the shared tower and Qwen3 schedules over a backend binding (#2105) #2239, and git merge-tree --write-tree confirmed it was the CHANGELOG.md bullet alone — both pull requests prepend one under ## Unreleased, and zig/build/cuda.zig auto-merged. Resolved by rebasing onto 610c67bf7 with both bullets kept. No zig/ file conflicts with any open pull request.

Checks on this head

• Head commit: 5b68662ca
• Lean checker, python3 tools/zig/lean_check.py: base 610c67bf7 5501 files / 425 problems, this head 5501 files / 425 problems, 0 new findings.
• Tests run on this head, all on one aarch64-linux host (Zig 0.17.0, 20 cores, no GPU), cold cache:
  • zig build test --summary all: 545/545 steps, 7066/7125 tests passed, 59 skipped, 0 failed.
  • Unmodified main on the same host, same command: 543/545 steps, 7065/7125 tests, 1 failed — roce_test: "a peer that stops makes the waiting rank time out, and abort ends a wait at once", a wait-time test that the box's load disturbs; it passes on this head.
  • zig build test-kernel-contract test-golden test-server-security test-pull-security fuzz-cortheon: exit 0. Goldens: 322/335 and 55/55 equal to the frozen corpus, 0 unexpected differences; server-security: 24 passed, 0 failed, 0 skipped.
  • zig fmt --check build.zig zig: clean.
• CI: green on this head — zig 22m55s, zig-windows 29m45s, release-paths 5m25s, zig-macos-dev-tools 1m55s, fmt 14s, benchmark-clients 14s and 34s.
  The zig job's own zig build test step took 19m12s here. Same-window pull requests on this repo ranged from 9m23s to 16m52s for that same step, so runner spread and the cache a pull request cannot share (until CI: zig and zig-windows keep main's build cache, and the quick checks run first #2140 lands) dominate this signal: this run recompiles every module that imports core, lanes or a family, which is what the mode change costs once. Read CI's job time as unproven either way, and the measurements above as what the change does to the tests themselves.
• Not run, and why: macOS Metal device checks, the CUDA/HIP/XPU device gates and the Windows job — this host has no such GPU and no macOS or Windows machine. The change only selects a host test binary's optimize mode and does not touch kernels, device code or the served path; the Linux and Windows unit jobs both build these same two steps.

Receipt

The change cannot reach inference (it is a build-graph and changelog change), so the inference part of the receipt is "n/a".

• Environment: 0311c9280 (the base the numbers were taken on; the rebase onto 610c67bf7 moved only the changelog bullet) + this change, Zig 0.17.0, aarch64-linux (Debian 13), 20 cores, no GPU, no checkpoint.
• Paths and platforms the change reaches: zig build test and zig build test-cuda-host / test-qwen4-host, host-only binaries, run on Linux aarch64. Windows and macOS build the same steps from the same scripts; not run here.
• Exactness: n/a — no arithmetic, kernel or request path changes; the same 7,125 tests run as before.
• Decode speed / prompt speed: n/a, no served path.

Run times from zig build test --summary all on the machine above (cold cache, wall clock per step):

Step	Binary	Debug	ReleaseSafe
test-qwen4-host	113 checks	240 s run + 15 s compile	18 s run + 50 s compile
test-cuda-host	546 checks	43 s	4 s
test-cuda-host	335 checks	42 s	3 s
test-cuda-host	495 checks	60 s	10 s
test-cuda-host	118 checks	18 s	1 s
test-cuda-host	67 checks	12 s	1 s
test-cuda-host	139 checks	10 s	0.7 s

Whole suite on the same host: 5m01 s wall, 1517 s compile CPU and 426 s measured run time before; 4m24 s wall, 1661 s compile CPU and 297 s run time after. The extra compile CPU is the price of optimizing those modules once; it is paid once per cache key, and the tests themselves no longer take minutes. On a runner with fewer cores the 4-minute serial floor is what CI's zig and zig-windows jobs were paying, so the wall-clock gain there should be larger than the 37 s seen on 20 cores — the pull request's own job times are the check.

The one exception, named: zig/src/families/nemotron/kv_pages.zig's holds() is a builtin.mode-gated std.debug.assert precondition, so inside these ReleaseSafe family modules it no longer runs. It is a developer precondition, not an assertion any test makes: the test totals are unchanged (7,125 both sides) and no test fails. The Debug family copies in cuda.zig (check_mods, and the copies loop) still compile it in Debug for their own filtered steps. A follow-up could widen that guard to skip only ReleaseFast/ReleaseSmall, which would change shipped-engine behaviour and does not belong in this CI-only change.

Checklist

• The title and description match the head commit.
• The suite fails on nothing this change introduced; the one failure seen on unmodified main is a load-dependent wait test that passes here.
• Comments are one line, within 120 columns. 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.

[ashhart]

Thanks @devopsfortunato. Same safety checks, same tests, and the Qwen4 gate at 18 s instead of 4 minutes is a clear win for every contributor's local run. Queued for after the 1.2.3 cut.

[ashhart]

Thank you @devopsfortunato. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.