v0.2.3 — Faster source-BF16 kernels and PyPI installation

This release speeds up source-BF16 projection and attention while preserving the recorded rounding and accumulation results. It also makes installation available directly from PyPI with the pinned public medmlx-core==0.1.2 dependency.

uv tool install --python 3.12 mlx-reason-ct==0.2.3

On one M1 Max synthetic short-response case, cached full-model execution decreased from 17:54 in the preceding source build to 14:49 (17.2% less time; 1.21× faster). The result was 1.80× faster than the older projection control. Recorded native tensors and logits were bitwise unchanged; comparisons with the CUDA source passed the existing tolerances.

Full original-BF16 qualification remains incomplete. The additional uncached pass encountered a system GPU reset; a fresh-model recovery reached EOS with matching logits. These single-case timings do not establish long-generation throughput or clinical accuracy. See measurement details and limits.

Validation: make env, make qa (115 tests, strict type checks and lint), and make build; clean package installation and a synthetic FP32 full-model smoke run. The dependency's existing v0.1.2 wheel passed all 219 tagged tests in a fresh environment.

The release includes the wheel, source archive and SHA-256 checksums. The pinned converted weights and model-material licenses are unchanged. FP32 remains the default; inference requires macOS arm64 and Metal.