[jishnuvenugopal]

Thank you, @ashhart, for settling how #2031 splits, and @garvrao80 for carrying the CI side in #2095, which is what keeps this step on a reviewed list instead of forgotten. Thanks too, @BobClawblaw, for the routed-expert path in 10eb897: it's why the recording root has one more name to export.

What this changes

zig/src/families/qwen27/moe_gpu.zig:6 reads core.affine_mm. In test-metal-residency-owners-host, core is the recording root zig/residency_owner_tests.zig, which zig/build/residency_host.zig hands to the Qwen27 engine modules. That root exports affine4 and affine4_lane but not affine_mm, so the step stops with root source file struct 'residency_owner_tests' has no member named 'affine_mm'. One line adds the export, in the same form as the lines beside it. Only a test root changes.

• Issue: Refs CI never reaches 246 of the 393 named test steps, and nine of them fail on main #2031. This is the fixture repair I said I'd take there. Once it lands, test-metal-residency-owners-host can come off ci: collect native Linux host checks and guard exclusions #2095's exclusion list. This does not put named steps on CI.
• Stacked on: main at 3e37f94, one commit, 238ace8. Through ac470c0, zig/residency_owner_tests.zig has not changed on main, and zig/src/core/affine_mm.zig has had comment-only changes. On 34d42fc the step still fails without this commit and passes 68/68 with it.
• atc overlap --json: ATC 0.8.1 found no overlaps, but the run reports complete: false (pullFilesUnavailable, timeLimit), so it is not a full result on its own. The same check by hand: the changed-file lists of all 187 open pull requests, and none touches zig/residency_owner_tests.zig. atc check 2031 says takenOutside because ci: collect native Linux host checks and guard exclusions #2095 references the issue; the split between ci: collect native Linux host checks and guard exclusions #2095's CI work and these fixture repairs was agreed on CI never reaches 246 of the 393 named test steps, and nine of them fail on main #2031.

Verified outcome

• Why: on main, the check that compares actual allocations with the recorded residency inventory does not compile, so none of its 68 tests run.
• What was done: the test's stand-in for the shared library now provides the one name the newer expert code asks for.
• Result: all 68 tests pass again, and every other check I compared gives the same result with and without this change.

Checks on this head

• Head commit: 238ace8.
• Lean checker, python3 tools/zig/lean_check.py, then tools/zig/lean_new.py: 425 problems on 3e37f94 and 425 on this commit, 0 new.

	main 3e37f94	this commit 238ace8
test-metal-residency-owners-host	does not compile: no member named affine_mm	5/5 steps, 68/68 tests passed
test-flashnext-vision-memory-host	16/16	same
test-flashnext-cleanup-host	16/16	same
test-qwen-dense-route-host	1/1	same
zig build test, empty cache on each side	597/711 steps succeeded, 63 failed; 9384/9447 tests passed, 62 skipped, 1 failed	same counts
test-golden	5/5 steps	same
test-kernel-contract	22/22 steps, 198/198 tests	same
fuzz-cortheon	3/3 steps, 20/20 tests	same
test-server-security	5/5 steps, 20/20 tests	same

zig fmt --check build.zig zig and git diff --check against 3e37f94 are clean on this commit.

• Cause pinned to one commit: on 10eb897's parent 206e629 the step passes 68/68; at 10eb897 it stops with this compile error.
• The other three rows share this test root and pass without the change, so this repairs one step, not four.
• The one failed test in zig build test is the same on main and on this commit: exl3_load's "a scan that fails an allocation frees every group prefix it held", which fails without output. This change does not touch it, and I have not looked into its cause.
• Of the 63 failed steps inside zig build test, 62 stop because xcrun cannot find metal, and the other is that test. The names match on both sides.
• CI: all 7 checks passed on this commit. No CI job reaches this step, which is how it went red on main unseen, so this green run cannot show this fix. The local mirror is the table above.
• Not run: plain zig build, native and dev-tools (no Xcode here, and this file is in none of them), Metal and CUDA devices, Linux hosts, and Windows hosts.

Receipt

n/a: this changes a test root that no shipped binary builds, so it cannot reach inference.

• Environment: 238ace8 on 3e37f94, Zig 0.17.0, Apple M1 Pro, macOS 26.6. Command Line Tools only. No checkpoint and no GPU run.
• Paths and platforms the change reaches: the recording test root's host compile. No kernel is new or changed.
• Licence: n/a. No new weights.

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 test. The step fails to compile on 3e37f94 and passes here.)
• Comments and docstrings are one line. No measurements or history in the source. (The diff adds no comment.)
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @jishnuvenugopal. Confirmed on an Apple M3 Ultra: zig build test-metal-residency-owners-host passes 5/5 steps and 68/68 tests with this commit. Without it, it fails on exactly the missing member: moe_gpu.zig:6:16: root source file struct 'residency_owner_tests' has no member named 'affine_mm'.

It's in the landing queue, with a CHANGELOG line that thanks you. Once it lands, the step can come off #2095's exclusion list, as you say.

[ashhart]

/atc exchange 3e17c53089ba0120568ae4318704fa5ade6905ef70a6c936104c9eb7aaff60c8

[ashhart]

Thank you @jishnuvenugopal. 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.

[ashhart]

Thanks @jishnuvenugopal! This is in TensorFold 1.2.4, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.4