[Whq02]

What this changes

zig build server-cpu (the CPU-only fake server and the engine-less CLI that zig/tests/server/build.sh installs) does
not compile on main at 53311e7:

$ zig build server-cpu
server-cpu
+- install fake_serve
   +- compile exe fake_serve fast native 1 errors
zig/src/server/slide_model.zig:4:20: error: no module named 'core' available within module 'server'
const st = @import("core").safetensors;
                   ^~~~~~
referenced by:
    readShard: zig/src/server/slide_model.zig:72:19
    scan: zig/src/server/slide_model.zig:50:18
    23 reference(s) hidden; use '-freference-trace=25' to see all references
error: 1 compilation errors
...
Build Summary: 4/7 steps succeeded (1 failed)
$ echo $?
1

Since 0a83dd5, slide_model.zig imports core, and that commit gave the core module to three of the four build
graphs that root zig/src/server/root.zig: the Metal build's server tests (zig/build/metal_targets.zig), the CUDA
host tests (zig/build/cuda.zig) and fuzz-cortheon (zig/build/cortheon_fuzz.zig). The fourth, plain in
zig/build/server.zig, builds the fake server's server module for server-cpu and was left without it, so
fake_serve no longer compiles; the engine-less CLI beside it roots main.zig directly with its own core and
still builds.

After this change plain carries core/root.zig over its own tokenizer module, the shape metal_targets.zig gives
the server tests, and cpuArtefacts takes core from the import table as it already takes engine_api, tokenizer
and template, so the fake server and the engine-less CLI share one core instead of the CLI's separate one. One
file, 3 insertions, 1 deletion.

$ zig build server-cpu --summary all
Build Summary: 7/7 steps succeeded
server-cpu success
+- install fake_serve success
|  +- compile exe fake_serve fast native success 57s MaxRSS:2G
+- install tensorfold-native success
   +- compile exe tensorfold-native fast native success 1m MaxRSS:2G
      +- options cached
      +- WriteFile exlr8_alias.zig cached
$ echo $?
0

Reproduced with Zig 0.17.0 on macOS arm64, without the Metal toolchain; server-cpu needs no Metal compilation.

• Issue: n/a: no issue filed; the failing build and its output are above.
• Stacked on: main (53311e7), one commit.
• atc overlap --json: ran atc overlap --json --base origin/main with ATC 0.9.0 at 2026-11-06T12:36:02Z on this head against origin/main at 53311e7; exit 0. changedFiles 1; 17 overlaps, each sharing zig/build/server.zig: server: an explicit thinking off that a template can't honour is refused #2170, Tests: pins for the documented limits, defaults and flag dependencies, a replay of the prompt cache's eviction, and a timing of its index #2305, lanes: one fixed code for every wait and refusal (#2135) #2196, /v1/models: each model lists what it supports, from the same facts the refusals read #2157 and metrics: an opt-in histogram of how long a request queued #2158 clean; EXL3 Flash Next on CUDA: one rank, offloaded experts, sm_120 gate (1/8) #1441, EXL3 Flash Next: expert-parallel TP=2 (same tokens as TP=1) and MTP drafts (2/8) #1442, EXL3 Flash Next: deep prefill (3/8) #1443, EXL3 Flash Next: KV pages from the core pool (4/8) #1444 and EXL3 Flash Next: host-resident K/V on the tier (6/8) #1522 conflicts; Unlimited-OCR-2 on native Metal: the publisher's BF16 release, page images to structured text #2074, Server: tool calls completed just before the thinking close are returned when no answer follows (#2310) #2313, EXL3 Flash Next: draft lane guard for offloaded experts (8/8) #1446, EXL3 Flash Next: CUDA graphs and the device planner (7/8) #1445, Metal: serve the dense Qwen3.8-27B straight from a GGUF file (Q4_K_M, Q8_0) (#1920) #2078, Unlimited-OCR-2 on CUDA: the family's steps on NVIDIA GPUs, the router's long-context speedup (stacked on #2074) #2077 and TensorFold Web: a small local web app, and tensorfold web to open it #1781 unknown. No local worktree overlaps.
  • The five conflicting EXL3 Flash Next PRs were checked at heads EXL3 Flash Next on CUDA: one rank, offloaded experts, sm_120 gate (1/8) #1441 eb31d6e, EXL3 Flash Next: expert-parallel TP=2 (same tokens as TP=1) and MTP drafts (2/8) #1442 3a1e883, EXL3 Flash Next: deep prefill (3/8) #1443 17b545b, EXL3 Flash Next: KV pages from the core pool (4/8) #1444 8643fd9 and EXL3 Flash Next: host-resident K/V on the tier (6/8) #1522 63bebd6. Their common base, 116a0d4, is 196 commits behind the main used here. git merge-tree --write-tree origin/main <their head> merges server.zig cleanly for every one, although it finds conflicts in other files. Against this head, every one adds the same direct conflict in cpuArtefacts: this change reads core from server_mod.import_table, while their patch keeps the separate core creation and adds server_mod.addImport("core", core). Those five overlaps are between two fixes for the same missing import, not inherited server.zig conflicts with main.
  • Resolution for each of EXL3 Flash Next on CUDA: one rank, offloaded experts, sm_120 gate (1/8) #1441, EXL3 Flash Next: expert-parallel TP=2 (same tokens as TP=1) and MTP drafts (2/8) #1442, EXL3 Flash Next: deep prefill (3/8) #1443, EXL3 Flash Next: KV pages from the core pool (4/8) #1444 and EXL3 Flash Next: host-resident K/V on the tier (6/8) #1522 when combining the changes: retain this change's core creation/import in plain and const core = server_mod.import_table.get("core").?; in cpuArtefacts; omit their separate core creation and redundant server_mod.addImport("core", core). Applying that single-hunk choice to each merge result gives exactly this tested head's server.zig. This is the resolution to coordinate before combining, not a claim that their branches have been updated or a combined tree tested; no source change was made here for it.
  • The result says complete: false (otherPullRequestsLimit, pullFilesLimit): there are 214 open PRs, more than the 200 the check reads, and it reports partial file-list coverage. The overlaps above are the ones it reached; the seven unknown results and omitted coverage are not cleared by exit 0.

Checks on this head

• Head commit: 2b92ced
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: both exit 1 with 5616 files, 252 problems, using the identical checker. The 252 findings are inherited; python3 tools/zig/lean_new.py BASE.txt HEAD.txt exits 0 with lean: 252 problems, 252 on the base, 0 new.
• Tests run, with the command and pass, fail and skip counts:
  • The build step is the regression check: zig build server-cpu exits 1 on 53311e7 with the error above, 4/7 successful steps and 1 failed compile. zig build server-cpu --summary all exits 0 on this head, 7/7 steps, 0 failures. No new test file: these commands directly exercise the affected build graph.
  • zig build test-golden --summary all: exit 0, 5/5 steps; 322/335 and 55/55 equal to the server goldens, 13 known differences, 0 unexpected differences, 0 skipped cases. It builds its own fake_serve from the macOS server-tests module, not plain, so this is the tree's golden gate rather than the regression check for this change.
  • zig fmt --check build.zig zig: exit 0, no files reported. git diff --check and git diff --check origin/main..HEAD: exit 0, no whitespace errors.
  • zig-out/server/fake_serve serve zig/tests/server/fixtures --port 0 selected free port 59858. GET /health returned 200 (curl exit 0, status: "ok", model: "fixtures"); SIGTERM drained and stopped the server with exit 0. Three assertions passed: health, fixture identity and graceful stop; 0 failures, 0 skips.
  • zig-out/server/tensorfold-native --version: exit 0, tensorfold-native 1.2.3.
• CI: not run on this head; the branch has not been pushed. The local checks above are not a hosted CI result.
• Not run, and why: zig build test and zig build native on macOS compile Metal shaders; no Metal compiler is available (xcrun -f metal exits 72). zig build test-grammar-tools -Dxgrammar=DIR, which builds tool_grammar_xgrammar_test through the other call to plain, needs an xgrammar checkout, which was not supplied; it gains the same core import but was not exercised. zig build grammar -Dxgrammar=DIR was not run either, and the golden and health runs warned that libtfgrammar.dylib is absent. No Linux or Windows build was run. ATC's direct overlaps and incomplete coverage are recorded above; no combined EXL3-series tree was built.

Receipt

n/a: this changes module wiring for the fake server, the engine-less CLI and the grammar tool tests, not an inference
engine or kernel. The built CPU server's /health and SIGTERM checks above exercise the changed graph's artifact;
no checkpoint, token-exactness or performance result is claimed.

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 file: the build step itself is red on the base and green here; it requires no optional backend.)
• 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.

[ashhart]

Thanks @Whq02, that's a real break on main, and the fix is the right shape: the plain graph now builds core once with tokenizer and cpuArtefacts reuses it, as the other graphs do. zig build server-cpu passes 7 of 7 steps here on your head. It's in the landing queue, right behind another build fix (#2362), with a CHANGELOG line we'll add at landing.

[ashhart]

Thank you @Whq02. 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.