[BlivionIaG]

Branch: series/capabilities-split, one commit on main c0c7bd24f.

What this changes

cli_capabilities.zig already writes the capabilities document (main f3c1ab024).
The Engines struct it reads still sat in cli.zig.
This change moves the struct next to its writer. cli.zig re-exports it, so cli.Engines and cli.capabilities keep working for engines.zig, benchmark.zig, main.zig and cli_test.zig.
cli.zig goes from 598 to 589 lines, so the next serve flags fit under the 600-line limit.
No behaviour change.

Files: zig/src/server/cli.zig (+2 -11), zig/src/server/cli_capabilities.zig (+14 -2).

Receipt

• Environment: TensorFold main c0c7bd24f and this branch 06a5236ce, Zig 0.17.0, -Dcpu=apple_m1 -Dversion=1.1.4, Apple M4 Pro, macOS (Darwin 25.5.0).
• capabilities --json: built native from main and from the branch and ran tensorfold-native capabilities --json on each.
  The two outputs are byte-identical: cmp reports no difference, both are 6853 bytes, and both have SHA-256 cb09db79b76a6681514a9be60a2074d650607bd9cdb4cbb70499a306a2373467.
• Exactness, decode speed, prompt speed: n/a. The change moves a type declaration; no request path changes.
• Tests run on the branch:
  • zig build native -Dcpu=apple_m1 -Dversion=1.1.4: exit 0.
  • zig build test-server-cpu -Dcpu=apple_m1: exit 0, 792 of 795 tests passed, 3 skipped (the server test binary 787 pass, 3 skip of 790; a second test step 5 of 5). On main's Zig sources the server test binary gives the same 787 pass, 3 skip of 790 (its 5-test step was cached in that run).
  • zig build test-golden -Dcpu=apple_m1 on main's Zig sources: the same 322/335 and 26/26, 0 unexpected differences.
  • zig build test-golden -Dcpu=apple_m1: exit 0; 322/335 and 26/26 equal to server goldens, 0 unexpected differences (the 13 differences are the existing KNOWN-DIFFER entries).
  • zig fmt --check on both files: exit 0.
  • git diff --check c0c7bd24f: clean.
  • python3 tools/zig/lean_check.py: 433 problems on main and 433 on the branch, the same findings line for line. The baseline findings are inherited; this change adds none.
• Not run, and why: the full zig build test, test-kernel-contract, fuzz-cortheon, the Linux CUDA host build and HIP. The change touches two server files that every platform compiles the same way; CI's Linux build will cover the non-macOS compile.

Tests

No new test. A moved declaration has no failing case to show.
The existing cli_test.zig capabilities tests call cli.capabilities with cli.Engines values, so they compile and pass through the re-export (part of the 787 above).
The byte comparison of the real binary's output is the check that the document did not change.

Not measured

• No model was loaded; no performance claim is made.
• M1, M2, M3, M5 and CUDA/HIP machines were not used.

Related PRs

Prerequisite for the DSpark drafter's two serve flags: #1392 sits on this PR. Part of #1374.

Issue

CONTRIBUTING asks for an issue first only for a new model family, backend or arithmetic change. This is none of those, so no issue is needed.

Prerequisite for the DSpark drafter's two serve flags (#1374): they need room in cli.zig.

[ashhart]

Landed on main (2a37b63), thanks @BlivionIaG: a pure move, cli.zig back under 600 lines with room to spare; server CPU tests pass.