Refresh MLX and add H3 MPP projection lab

[a-github-name]

Summary

• refresh the pinned mlx-swift fork and vendored Metal provenance after merging the current low-risk upstream MLX changes
• add a lab-only BF16 MPP projection kernel candidate for MiniMax H3 on current Metal 4 Macs
• keep production H3 dispatch unchanged until clean practical-shape timing proves a gain
• document source attribution, capability boundaries, and the guarded benchmark command

Dependency refs

• sawfwair/mlx:codex/h3-metal-low-risk-refresh at 31af89c4c21642236b8a2bc1358438512d9521e3
• sawfwair/mlx-swift:codex/h3-metal-refresh at 5bf3e46fecfb69cd3b559025fa99885ddd188731

Validation

• ./scripts/check.sh
  • SwiftLint: 0 violations across 1,241 files
  • XCTest: 2,878 tests, 209 skipped, 0 failures
  • Swift Testing: 30 tests in 4 suites passed
  • build, metallib provenance, CLI help sweep, and hygiene scans passed
• MERERUN_TEST_MLX_DEVICE=gpu swift test --filter MiniMaxH3MPPProjectionTests
  • both kernel tile configurations are bit-exact against MLX on the GPU canary

Performance boundary

The production-shape benchmark was correctly refused by the lab guard because the host had unrelated workloads and about 4.97 GiB swap in use. No speedup is claimed, and the candidate is not wired into production dispatch. Rerun scripts/h3-kernel-lab.sh mpp-projections on a clean host before promotion.

The standalone mlx-swift package compiled, but its tests cannot launch without a colocated default metallib. The downstream mere.run gate validates the pinned sources with the rebuilt vendored metallib.