simd_avx2.rs — 3 precision tricks, all AVX2-accelerated (additive only):

Trick 1: Double-f16 (Error-Free Split)
  f16_double_encode/decode: store value as hi+lo f16 pair
  ~20-bit effective precision (vs 10-bit single f16)
  f16_double_encode/decode_batch: AVX2 F16C + f32x8 addition
  Error: ≤2^{-21} × |value| (vs ≤2^{-11} for single f16)

Trick 2: Kahan-compensated accumulation
  f16_kahan_sum: O(ε) error instead of O(N·ε) — independent of count
  f16_kahan_dot: AVX2 f32x8 multiply + Kahan-accumulate partial sums

Trick 3: Exponent-aligned scaling (F16Scaler)
  from_range/from_data: auto-compute scale factor for value range
  encode/decode_batch: AVX2 f32x8 scale + F16C convert
  Up to ~128× precision improvement for narrow-range data

⚠️  NOT FOR GGUF CALIBRATION — BF16 pipeline is separate

.claude/agents/arm-neon-specialist.md:
  Complete ARM SBC knowledge: Pi Zero 2W through Pi 5, Orange Pi 3-5
  Per-CPU microarchitecture (A53/A72/A76 pipeline differences)
  big.LITTLE awareness (RK3399, RK3588)
  F16 inline asm trick, codebook strategy per tier, memory budgets

6 new tests passing. No existing code modified.

https://claude.ai/code/session_017ZN5PNEf8boFBgorUZVrFU