[swbiggart]

What this changes

cluster check and cluster status printed not planned but returned success when a model could not be planned. Return 1 for that failure. serve, including --dry-run, now stops at this preflight before describing or attempting a launch.

The existing preflight checks every configured model unless --model filters it: serve A can therefore now refuse because a different model B cannot be planned. A non-fitting B already blocked A; this change leaves model selection unchanged.

• Issue: n/a; a standalone host planner bug, with no model arithmetic change.
• Rebased onto main at 023def309dd1d5534c740fb7465e2835cee3ee2c to resolve the changelog conflict. Current head: a25ffe1cdb691b633e5a0f973b846f06551ea1a3.
• atc overlap --json: incomplete (pullFilesLimit, pullFilesUnavailable, timeLimit). A separate paginated audit of all 173 open PRs on 2026-11-03 at 19:55 UTC found no PR changing these files; related cluster discussions and merged changes were also checked.

Checks

• Tested implementation commit: e8089a852c95043fe679f7db9601500c88746486, before the rebase and changelog-only follow-up. Both changed Zig files are byte-identical to that tested commit.
• Changelog follow-up: documents all three preflight exit codes under Unreleased. git diff --check passes; the rebased branch merges cleanly with main. No tests rerun for the documentation-only follow-up.
• Lean checker, python3 tools/zig/lean_check.py: 424 problems on base and head, 0 new.
• New behavior regression: on main, serve --dry-run fails the expected-exit-1 assertion with actual 0; the filtered run has 7 passes and 1 failure before the fix and 8 passes afterward (7 module test blocks and 1 behavior regression; 0 skipped). It also checks check and status for missing file lists and unsupported estimates.
• Full host CLI suite: 72 passed, 6 skipped, 0 failed.
• zig build test test-kernel-contract test-golden test-server-security test-pull-security fuzz-cortheon -Dcpu=apple_m1 -j1 --summary all: exit 0; 675/675 steps succeeded, 372/378 tests passed, 6 skipped. This final-head run was mostly cached.
• Plain zig build -Dcpu=apple_m1 -j1 --summary all: exit 0; 250/250 steps succeeded.
• zig fmt --check on both changed files and git diff --check: pass.
• CI: pending on this PR.
• Not run: Linux host execution, CUDA device tests and real-cluster/model tests. This is a host preflight change, with no inference claim.

Filtered regression command:

zig test -lc --dep sys --dep core --dep fabric --dep lanes \
  -Mroot=zig/src/cluster/cli.zig -Msys=zig/src/platform/sys.zig \
  --dep sys -Mcore=zig/src/core/root.zig \
  --dep sys -Mfabric=zig/src/fabric/fabric.zig \
  --dep sys -Mlanes=zig/src/core/lanes/lanes.zig \
  --test-filter 'check, status and serve'

Receipt

n/a: host planner exit codes only; no inference path changes. Checks ran on Apple M5 Max, macOS 27.2, Zig 0.17.0.

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.
• 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 @swbiggart. Printing "not planned" and exiting 0 was a real trap for scripts, and the fix puts it on the same footing as a model that doesn't fit, with a test for each command and both refusal kinds. One small thing: could you add a CHANGELOG line saying cluster check, cluster status and serve's preflight now exit 1 when a configured model can't be planned? Then it can land.

[ashhart]

Thanks @swbiggart. The CHANGELOG line is exactly what we asked for, and the code commit is the one we reviewed, so this is ready to land.

[ashhart]

Thank you @swbiggart. 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.