fix(metal): build with Xcode 27 / Metal 4.1 (explicit thread on member functions)

[solderzzc]

Problem

Xcode 27 ships Metal 4.1 (metalfe-32023.921). In Metal 4.1, this inside a member function no longer defaults to the thread address space; it now defaults to __metal_generic. As a result, a member function that returns thread T& to one of its members fails to compile:

steel/gemm/mma.h:245: error: reference to type 'thread vec' could not bind to an lvalue of type 'vec'

17 of the 40 kernels fail. That includes every quantized matmul kernel (quantized, quantized_nax, fp_quantized*) and all steel GEMM and attention kernels. ./build.sh in SwiftLM stops at mlx-metallib, so nothing can be built with Xcode 27. Found while setting up an M6 Mac mini on macOS 27.0.

Fix

• Backport ml-explore/mlx#3963 (ef5fc0fab2), the upstream fix for this exact issue, into the vendored mlx 0.30.6. Some of its hunks don't apply because upstream has since rewritten that code (sort.h LessThan, the NAXFrag-based nax.h, the Conv3D loaders). Those hunks aren't needed here.
• Apply the same rule to the older NAXSubTile/NAXTile structs in steel/gemm/nax.h and steel/attn/nax.h: add an explicit thread qualifier to each member function signature (32 per file). #3963 doesn't cover these because upstream replaced them.
• Regenerate mlx-generated/. Both the JIT .cpp sources and the metal/ header copies are regenerated from the patched kernels. I ran the generator twice, on unpatched and on patched sources, and applied only the difference between the two runs. The existing hand edits in mlx-generated/ (the compiled_preamble bool fix, conv.cpp, and the hand-written steel_conv_3d.cpp stub) are left as they were.

On older toolchains this changes nothing, because thread was already the implicit default for this.

Verification

• With Metal 32023.921 (Xcode 27A266a), all 39 kernels in the metallib build list compile. steel_conv_3d still fails on its own, but it isn't in the build list and it failed before this change too.
• SwiftLM ./build.sh completes: a 128 MB mlx.metallib, and the Swift 6.4 build succeeds.
• Inference smoke test on M6 with mlx-community/Qwen3.8-27B-4bit (it runs the quantized, GEMM, attention and NAX kernels). Results: correct answers on a short factual question and on arithmetic, and a planted code word retrieved correctly from a 7.9K-token prompt; decode 9.0 tok/s; no Metal errors.
• Build check on Xcode 26

🤖 Generated with Claude Code