feat(kernels): add blitz-int8-matmul + blitz-bf16-matmul — 11 kernels, 162 tests

Wave 4 (INT8 quantized matmul): 4× memory bandwidth reduction, symmetric
per-tensor and per-channel quantization, production-grade INT8 inference.

Wave 5 (BF16 matmul): H100/A100/TPU native dtype, same dynamic range as
FP32, trivial conversion, LLaMA/Mistral/Gemma default training dtype.

Catalog: 11 kernels (10 workspace + cc-faculty-wasm). 162 tests passing.
Pricing corrected: Full catalog $6,500 (matches landing page).

-- NEXUS | 2026-03-11

Author: spectre

Cargo.toml

@@ -8,6 +8,8 @@ members = [
     "kernels/blitz-fused-mlp",
     "kernels/blitz-swiglu",
     "kernels/blitz-rmsnorm",
+    "kernels/blitz-int8-matmul",
+    "kernels/blitz-bf16-matmul",
 ]
 
 [workspace.package]

README.md

@@ -8,7 +8,7 @@
 [![built with Rust](https://img.shields.io/badge/built_with-Rust-orange)](https://www.rust-lang.org)
 [![target: wasm32-wasip2](https://img.shields.io/badge/target-wasm32--wasip2-purple)](https://wasi.dev)
 
-## Kernel Catalog — 9 Kernels Available
+## Kernel Catalog — 11 Kernels Available
 
 | Kernel | Description | Use Case |
 |--------|-------------|----------|
@@ -20,6 +20,8 @@
 | `blitz-fused-mlp` | Fused Linear→LayerNorm→GELU→Linear | Full FFN block (GPT-style) |
 | `blitz-swiglu` | SwiGLU gated activation | LLaMA 2/3, Mistral FFN |
 | `blitz-rmsnorm` | RMS LayerNorm (no mean subtraction) | LLaMA/Mistral/Gemma normalization |
+| `blitz-int8-matmul` | INT8 quantized matmul · 4× memory bandwidth | Quantized weight inference |
+| `blitz-bf16-matmul` | BF16 matmul · H100/TPU native dtype | Mixed-precision LLM inference |
 | `cc-faculty-wasm` | Claude Code cognitive substrate | Agent memory + reasoning integration |
 
 **[→ View full catalog and pricing](https://blitzkernels.pages.dev)**
@@ -91,6 +93,7 @@ rms_norm(&mut hidden, &weight, hidden_size, 1e-6);
 │  blitz-embedding → blitz-attention → blitz-kv-cache  │
 │  blitz-rope → blitz-rmsnorm → blitz-swiglu           │
 │  blitz-fused-mlp → blitz-layernorm-gelu              │
+│  blitz-int8-matmul → blitz-bf16-matmul               │
 │  cc-faculty-wasm                                      │
 └──────────────────────────────────────────────────────┘
          Pure Rust · No allocator required
@@ -109,7 +112,7 @@ rms_norm(&mut hidden, &weight, hidden_size, 1e-6);
 | Option | Price | Includes |
 |--------|-------|---------|
 | Single kernel | **$1,500** | Pre-compiled WASM + source + 30-min integration call |
-| Full catalog (9) | **$8,500** | All 9 kernels + dedicated integration support + priority updates |
+| Full catalog (11) | **$6,500** | All 11 kernels + dedicated integration support + priority updates |
 | Support add-on | **$200/mo** | Priority email, patch releases, architecture review |
 
 **[→ Purchase at blitzkernels.pages.dev](https://blitzkernels.pages.dev)**  

kernels/blitz-bf16-matmul/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "blitz-bf16-matmul"
+version = "0.1.0"
+edition = "2021"
+description = "BF16 matrix multiplication — H100/A100 inference dtype, pure Rust, WASM-portable"
+license = "MIT"
+
+[lib]
+crate-type = ["cdylib", "rlib"]
+
+[features]
+default = ["std"]
+std = []