[Wildebeest]

Summary

Upgrades mlx from 0.32.0 to 0.32.4 and nanobind from 2.13.0 to 3.1.0:

1. mlx==0.32.4 (engine/pyproject.toml):
   • Picks up Apple Silicon Metal (g13–g16 + M5 g17 NAX) kernel improvements between 0.32.0 and 0.32.4, including faster quantized MoE (SwitchGLU) / gather_qmm / SDPA prefill kernels, the Metal concurrent fence deadlock fix (ml-explore/mlx#3469), and the NAX >32K prompt token integer-overflow fix (ml-explore/mlx#3492).
2. nanobind==3.1.0 (engine/pyproject.toml):
   • mlx==0.32.4 wheels are built against nanobind 3.1.0 (ml-explore/mlx#3512). Bumping the build-system requirement from 2.13.0 to 3.1.0 keeps engine/native/grammar_native.cpp's nb::ndarray type-caster ABI matched with libmlx.
3. Immutable mx.random.state compatibility (engine/src/rec_mlx/kernels/masked_sampling.py, engine/tests/validate_resident.py):
   • In mlx>=0.32.2 (ml-explore/mlx#3495), mx.random.state returns a read-only tuple rather than a mutable list, so mx.random.state[0] = key raises TypeError: 'tuple' object does not support item assignment.
   • Restoring a 2×uint32 Threefry key [hi, lo] via mx.random.seed((int(hi) << 32) | int(lo)) constructs the identical mx.random.key state (mx.array([hi, lo], dtype=mx.uint32)) and preserves exact EOS RNG truncation parity.

Benchmarks & Testing

Measured on flint-int4 (Apple Silicon Metal, 64 GB unified memory, 128-token greedy decode after cold prefill):

Prompt Tokens	mlx==0.32.0 TTFT	mlx==0.32.4 TTFT	mlx==0.32.0 Prefill	mlx==0.32.4 Prefill	Prefill Speedup
1,024 (warm)	1.89 s	1.46 s	546.6 tok/s	716.2 tok/s	+31.0%
4,096 (rep 0)	7.10 s	5.71 s	578.1 tok/s	719.5 tok/s	+24.5%
4,096 (rep 1)	7.07 s	5.79 s	580.7 tok/s	708.7 tok/s	+22.0%
8,192 (rep 0)	15.14 s	12.24 s	541.6 tok/s	670.1 tok/s	+23.7%
8,192 (rep 1)	15.03 s	12.27 s	545.5 tok/s	668.9 tok/s	+22.6%

Tested:

• cd engine && .venv/bin/pytest -v -rs: 66 passed, 1 skipped (including native _grammar_native C++ extension rebuild and m4_max_prefill Metal kernel compilation check)
• make test-engine runtime packaging (packaging/build-mlx-runtime.py on standalone CPython 3.12.12, tests/smoke_test.py, and tests/protocol_test.py): all passed
• engine/tests/validate_resident.py: 61 / 61 numerical parity cases, 9 / 9 EOS truncation RNG checks, and 3 / 3 cancellation checks pass
• uvx ruff@0.17.1 check and uvx ruff@0.17.1 format --check: clean