v0.51.0 — P0 .apr hotfix + PMAT-877..888 wave

[0.51.0] - 2026-06-21

Hotfix-driven release (brought forward from the Friday cadence by a P0). Each fix ships a named
proof-obligation + a RED-on-bug / GREEN-on-fix falsifier + a pv-validated contract.

Fixed

• P0 — non-Gemma2 .apr inference produced garbage (PMAT-888, regressed in 0.50.0 via PMAT-810b)
  — every non-Gemma2 .apr (qwen2/llama/mistral/phi/deepseek/qwen3 — the majority of models) generated
  garbage on inference (CPU and GPU) while the same model as GGUF was coherent. PMAT-810b added a
  Gemma2 post-attention-norm load keyed on the HF name post_attention_layernorm.weight — which is the
  FFN norm for all those architectures — un-gated by architecture, so a spurious extra RMSNorm
  was applied. Now gated on config.is_gemma2(), mirroring the GGUF loader. GGUF was never affected.
• BatchNorm1d never updated running_mean/running_var (PMAT-877, Pillar-2) — they stayed at
  init (0/1) forever, so eval-mode normalization was wrong vs PyTorch. Now EMA-updated each training
  forward (running = (1-momentum)·running + momentum·batch).
• Linear bias initialized to zeros (PMAT-878, Pillar-2) — PyTorch uses U(±1/√fan_in); now matches
  (seed-deterministic).
• LoRA dropout never applied (PMAT-879, Pillar-3) — LoRALayer::forward ignored the configured
  dropout, so fine-tuning trained with zero regularization. Now applies dropout to the input
  (y = Wx + s·B(A(dropout(x))), train-only), matching HF PEFT.
• Batched-GPU GQA fail-closed (PMAT-880, Pillar-4) — attention_with_cache_gqa did not validate
  kv_dim == num_kv_heads·head_dim/cache consistency, silently reading wrong memory on a corrupt config;
  now returns a clear error (zero false-positives on valid models), where llama.cpp/Ollama run garbage.

Performance — GPU (Blackwell / GB10)

• First pure-Rust cuda-oxide #[kernel] to BEAT hand-PTX (PMAT-882) — the incremental KV-cache
  attention kernel: bit-exact (cos = 1.0) and 1.7–2.9× faster than the production hand-PTX kernel on
  GB10 (true on-device A/B). FMA/softmax kernels are not DP4A-bound, so pure-Rust competes and wins.
• Blackwell CUDA-graph replay fixed + re-enabled (PMAT-886a) — the default sm_121 Q4K GEMV variant
  was not recorded into the manual graph, so graph replay dropped ~6 GEMVs/layer → stale buffers →
  garbage (cosine 0.53). Now recorded; parity 0.53→0.9934 (== eager, token-for-token), graph decode
  re-defaulted ON for Blackwell, +16% decode (96→112 tok/s).
• Blackwell decode throughput-floor guard (PMAT-885) — a stale-binary / F2-false-fallback that
  silently drops the GPU path to ~10 tok/s CPU is now a falsifiable invariant (≥100 tok/s on GB10).

Infrastructure

• Pre-release Gate 11 (cargo publish -p aprender --dry-run) — catches the two classes that broke
  the 0.50.0 cascade mid-publish (sibling path-deps missing a version; version-pinned sibling dev-deps
  forming publish cycles) which cargo metadata does not detect.
• Dogfood Gate 18 (fresh-convert .apr inference parity vs GGUF, CPU+GPU) — catches the PMAT-888
  class that inspect/validate/tensors and a stale pre-existing .apr all pass through.