Pre-1.0 minor bump signaling meaningful new public surface area.
All additions are backwards-compatible — no API breaks.

ADR-022 — kernel coverage parity with llama.cpp (5563e83 merge):
new GGUF format coverage and dense matmul completions.
- Q5_1 + IQ4_NL: loader + host dequant + dot-product helpers,
  mv_id Metal kernels + dispatch, dense mv kernels, mm_id
  template instantiations, dense mm + mm_tensor + mm_id parity,
  mul_mv_ext r1 family.
- Q5_K: mm_id + mm_id_tensor, dense mv + dense mm + dense mm_tensor.
- Q4_K: dense mm + mm_tensor.
- Q8_0: perm021 transpose-fused matmul.
- mv_ext r1 family across Q4_0/Q8_0/Q4_K/Q5_K/Q6_K.
- AC-4 (81af181): env-gated HF2Q_LOG_MM_ID_ROUTE engagement trace.

ADR-020 iter-11h — additional backward / training kernels:
- exp_f32 + backward (b64f1e1)
- divide_f32 + backward (c0aafce)
- sqrt_f32 + backward (d4bdacb)
- conv1d_depthwise_causal forward + backward (d85c14a)
- outer_product forward + backward (d40ee5f)
- take_along_axis (gather + scatter-backward) (1ed1d14)
- moe_weighted_sum_seq backward (243e10e)
- moe_swiglu_seq fused backward (c3ce412)

ADR-020 iter-15c — simdgroup-MMA qmm_affine variants:
- qmm_affine_t_f32_simd (feb1e0e)
- qmm_affine_simd4 (501c664) — closes §8.3 AC #4
- qmm_affine_simd4 gs=64 variant (fd4944e)

Plus: README refresh from v0.6.4–v0.6.7 surface (37670df) lands
on crates.io / docs.rs with this release.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>