docs: update for v2.3.0 release — changelog, sponsors, roadmap, zh-CN/zh-TW docs

Author: Michael-A-Kuykendall

CHANGELOG.md

@@ -5,6 +5,29 @@ All notable changes to Shimmy will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.3.0] - 2026-07-20
+
+### Added
+- **Grammar control hooks** — schoolmarm+grammar module for structured generation
+- **PPT invariant cage (B1-B3)** — golden-vault-based regression detection. Contracts verify per-layer RMS/checksum against vault oracles for 12 populated models. Run with `cargo test --test test_invariants -- --test-threads=1`.
+- **Discrete GPU preference** — adapter selection now enumerates all GPUs and prefers `DiscreteGpu` over `IntegratedGpu` (commit `6ed0349`). Fixes multi-GPU laptops picking the Intel iGPU.
+
+### Fixed
+- **batch_count: 0 → batch_count: 1** in `server_inference.rs`. The QKV shader guard `if (global_id.y >= params.batch_count)` killed ALL threads when `batch_count=0` (since `0>=0` is true), producing no Q/K/V output. Root cause of the server-side gibberish regression (bisected to commit `43027d3`).
+- **Rust/WGSL struct layout mismatch** — per-field quant types (Q/QK/K/V) aligned between Rust `LayerParams` and WGSL shader structs.
+
+### Changed
+- Airframe dependency updated to 0.2.9.
+- Shimmy version bumped to 2.3.0.
+
+### Library API
+- `tokenizer_arc()` — shared tokenizer access
+- `eos_token()`, `im_end_token()` — token accessors
+- `fse_control_from_patterns()` — grammar pattern compiler
+- `trace_callback()` — capture hook for invariant probes
+
+---
+
 ## [2.2.1] - 2026-06-30
 
 ### Fixed
@@ -1027,9 +1050,12 @@ The systematic engineering discipline established in v1.7.3 creates a **BULLETPR
 ### Free Forever Commitment
 Shimmy is committed to being free forever with no asterisks, no "free for now" periods, and no pivot to paid services. The MIT license ensures this commitment is legally binding.
 
-[Unreleased]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.0.0...HEAD
-[2.0.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.10.0...v2.0.0
-[1.10.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.9.0...v1.10.0
+[Unreleased]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.3.0...HEAD
+[2.3.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.2.1...v2.3.0
+[2.2.1]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.2.0...v2.2.1
+[2.2.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.1.0...v2.2.0
+[2.1.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v2.0.0...v2.1.0
+[2.0.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.9.0...v2.0.0
 [1.9.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.8.1...v1.9.0
 [1.8.1]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.8.0...v1.8.1
 [1.8.0]: https://github.com/Michael-A-Kuykendall/shimmy/compare/v1.7.4...v1.8.0

README.md

@@ -62,9 +62,11 @@ Shimmy is a **single-binary** that provides **100% OpenAI-compatible endpoints**
 
 **⚡ NEW in v2.1.0**: [TurboShimmy INT4 KV](#-turboshimmy-int4-kv) — ~7× less KV cache VRAM with one flag. Run Llama-3.2-3B on 4 GB GPUs.
 
-## 🔥 Airframe Engine (v0.2.7)
+**⚡ NEW in v2.3.0**: Adapter selection fix (prefers discrete GPU over integrated), grammar control hooks, 357 passing tests.
 
-Starting in v2.0.0, Shimmy's default inference engine is **Airframe** — a pure-Rust WebGPU (WGSL) transformer runtime built from scratch. **v0.2.7** brings the Inference Saturation Fabric (ISF) refit and TDR transport integration to production.
+## 🔥 Airframe Engine (v0.2.9)
+
+Starting in v2.0.0, Shimmy's default inference engine is **Airframe** — a pure-Rust WebGPU (WGSL) transformer runtime built from scratch. **v0.2.9** brings the production batch_count fix, GPU adapter selection improvement, grammar hooks, and the PPT invariant cage.
 
 **See [airframe CHANGELOG](https://github.com/Michael-A-Kuykendall/airframe/blob/master/CHANGELOG.md) for full release notes.**
 
@@ -328,7 +330,7 @@ cd shimmy
 cargo build --release
 ```
 
-> **Note**: The Airframe GPU engine is a private dependency and **cannot be built from source** by public users. The [pre-built release binaries](#-download-pre-built-binaries-recommended) already include Airframe compiled in — download those to get full GPU acceleration. `cargo install shimmy` installs the huggingface engine variant from crates.io.
+> **Note**: The Airframe GPU engine is a public crate on [crates.io](https://crates.io/crates/airframe) and builds from source automatically. `cargo install shimmy` installs the huggingface engine variant from crates.io; for the full GPU build use the [pre-built release binaries](#-download-pre-built-binaries-recommended) or clone and build with `cargo build --release`.
 
 ### GPU Acceleration
 

ROADMAP.md

@@ -34,8 +34,8 @@ Shimmy is a zero-config, OpenAI-compatible inference server with a native WebGPU
   - No CUDA toolkit or Vulkan SDK required; wgpu handles adapter selection
 
 ## 🎯 Q2–Q3 2026 Milestones
-- [ ] **Stop tokens from GGUF metadata** — read `tokenizer.ggml.eos_token_id` natively
-- [ ] **Quantization in Airframe** — Q4_K_M and Q8_0 inference on the WebGPU pipeline
+- [x] **Stop tokens from GGUF metadata** — read `tokenizer.ggml.eos_token_id` natively
+- [x] **Quantization in Airframe** — Q4_K_M and Q8_0 inference on the WebGPU pipeline
 - [ ] **SafeTensors support** — ingest `.safetensors` model checkpoints directly
 - [ ] **Multi-model serving** — load balancing across multiple active models
 - [ ] **Enterprise Embeddings** — `/v1/embeddings` endpoint targeting RAG workloads

SPONSORS.md

@@ -1,15 +1,14 @@
 # These amazing people make Shimmy possible 🙏
 
-## Infrastructure Partners ($500+/month)
-*Your logo could be here*
-
 ## Corporate Backers ($100+/month)
-*Your logo could be here*
+- **[ZephyrCloudIO](https://github.com/ZephyrCloudIO)**
 
 ## Bug Prioritizers ($25+/month)
+*Your name could be here — [sponsor](https://github.com/sponsors/Michael-A-Kuykendall)*
 
 ## Coffee Tier Heroes ($5+/month)
-*Your support could be here*
+- **[gqf2008](https://github.com/gqf2008)** (高庆丰)
+- **[alistairheath](https://github.com/alistairheath)**
 
 ---
 

docs/USER_MANUAL.zh-CN.md

@@ -6,7 +6,7 @@
 
 [📚 中文文档中心](zh-CN/README.md) · **简体中文手册** · [繁體中文](USER_MANUAL.zh-TW.md) · [English](../README.md)
 
-版本：v2.0.0 及以上 · 最后更新：2026 年 5 月
+版本：v2.3.0 及以上 · 最后更新：2026 年 7 月
 
 </div>
 
@@ -51,12 +51,17 @@ Shimmy 是一个**单一二进制文件**的本地 AI 推理服务器，提供
 | 🦀 **纯 Rust 实现** | 单一可执行文件，无 Python 运行时，无 C++ 依赖 |
 | 🖥️ **跨平台 GPU 加速** | 通过 WebGPU（wgpu）支持 NVIDIA、AMD、Intel 及 Apple Silicon |
 
-### v2.0.0 新特性
+### v2.3.0 新特性
 
-- **Airframe 引擎**：全新纯 Rust WGSL 推理引擎，彻底取代 llama.cpp
-- **WebGPU 加速**：通过 wgpu 自动选择最佳 GPU 适配器
-- **YaRN RoPE 缩放**：支持超长上下文窗口（最高 16384+ tokens）
-- **零工具链依赖**：无需 CUDA 工具包、Vulkan SDK 或任何 C++ 编译器
+- **独立 GPU 自动优先选择**：多 GPU 系统（如笔记本）自动优先选择独立显卡而非集成显卡
+- **Grammar 控制钩子**：支持结构化生成（schoolmarm+grammar 集成）
+- **PPT 不变性测试笼**：基于黄金 vault 的回归检测，覆盖 12 个模型的逐层 RMS/校验和验证
+- **batch_count 修复**：修复了 QKV 着色器中 batch_count=0 导致所有线程被终止的严重问题
+- **357 个库测试全部通过**
+
+### 历史版本
+
+**v2.0.0**: Airframe 引擎首发，纯 Rust WGSL 推理引擎，彻底取代 llama.cpp
 
 ---
 
@@ -177,7 +182,7 @@ cargo install shimmy
 ### 方式三：从源码构建（含 Airframe 引擎）
 
 ```bash
-git clone https://github.com/Michael-A-Kuykendall/shimmy --recurse-submodules
+git clone https://github.com/Michael-A-Kuykendall/shimmy
 cd shimmy
 cargo build --release --features airframe,huggingface
 

docs/USER_MANUAL.zh-TW.md

@@ -6,7 +6,7 @@
 
 [📚 中文文件中心](zh-TW/README.md) · [简体中文](USER_MANUAL.zh-CN.md) · **繁體中文** · [English](../README.md)
 
-版本：v2.0.0 及以上 · 最後更新：2026 年 5 月
+版本：v2.3.0 及以上 · 最後更新：2026 年 7 月
 
 </div>
 
@@ -51,12 +51,17 @@ Shimmy 是一個**單一執行檔**的本地 AI 推論伺服器，提供與 Open
 | 🦀 **純 Rust 實作** | 單一可執行檔，無 Python 執行環境，無 C++ 相依套件 |
 | 🖥️ **跨平台 GPU 加速** | 透過 WebGPU（wgpu）支援 NVIDIA、AMD、Intel 及 Apple Silicon |
 
-### v2.0.0 新功能
+### v2.3.0 新功能
 
-- **Airframe 引擎**：全新純 Rust WGSL 推論引擎，完全取代 llama.cpp
-- **WebGPU 加速**：透過 wgpu 自動選擇最佳 GPU 適配器
-- **YaRN RoPE 縮放**：支援超長上下文視窗（最高 16384+ tokens）
-- **零工具鏈相依**：無需 CUDA 工具包、Vulkan SDK 或任何 C++ 編譯器
+- **獨立 GPU 自動優先選擇**：多 GPU 系統（如筆電）自動優先選擇獨立顯示卡而非內建顯示卡
+- **Grammar 控制鉤子**：支援結構化生成（schoolmarm+grammar 整合）
+- **PPT 不變性測試籠**：基於黃金 vault 的迴歸檢測，覆蓋 12 個模型的逐層 RMS/校驗和驗證
+- **batch_count 修復**：修復了 QKV 著色器中 batch_count=0 導致所有執行緒被終止的嚴重問題
+- **357 個庫測試全部通過**
+
+### 歷史版本
+
+**v2.0.0**: Airframe 引擎首發，純 Rust WGSL 推論引擎，完全取代 llama.cpp
 
 ---
 
@@ -177,7 +182,7 @@ cargo install shimmy
 ### 方式三：從原始碼建置（含 Airframe 引擎）
 
 ```bash
-git clone https://github.com/Michael-A-Kuykendall/shimmy --recurse-submodules
+git clone https://github.com/Michael-A-Kuykendall/shimmy
 cd shimmy
 cargo build --release --features airframe,huggingface