Stage M4 — fp16/bf16 on the Metal path *(reopens the Stage 8 reject; new)*

[thanos]

Stage M4 — fp16/bf16 on the Metal path (reopens the Stage 8 reject; new)

Goal: halve weight and KV bytes moved per token before touching
quantization.

• fp16 (and bf16 where the device supports it usefully) weight storage
  and compute for the Qwen Metal path; f32 accumulation where numerics
  require it (teach why: softmax, RMSNorm reductions, logits).
• KV cache in fp16 with documented tolerance impact.
• Extend .zynfer artifact + converter for half-precision payloads
  (dev-time conversion; runtime stays Python-free).
• Differential tests with dtype-justified tolerances — do not loosen
  tolerances to pass; explain every widened bound numerically.
• Report bytes/token before/after and the new roofline fraction.

Gate: fp16 path matches golden tokens for greedy decoding on fixture
prompts (or the deviation is characterized and accepted explicitly);
measured decode speedup consistent with the bandwidth model.
Tutorial: docs/tutorials/18-half-precision-inference.md.