diff --git a/.github/workflows/npm-release.yml b/.github/workflows/npm-release.yml
index d782e94..72160e2 100644
--- a/.github/workflows/npm-release.yml
+++ b/.github/workflows/npm-release.yml
@@ -6,7 +6,7 @@ on:
       version:
         description: Core and stable Small facade version
         required: true
-        default: 0.5.7
+        default: 0.5.9
         type: string
       publish_to_registry:
         description: Publish the tested candidate to npm under next
@@ -14,7 +14,7 @@ on:
         default: false
         type: boolean
       npu_sdk_run_id:
-        description: Optional run containing npu-release-sdks (AMD requires explicit provider configuration)
+        description: Optional override containing both locked Intel and AMD deployment SDKs
         required: false
         default: ""
         type: string
@@ -47,6 +47,10 @@ jobs:
           python tools/npm_release.py ensure-unpublished --version "$RELEASE_VERSION"
 
   build-native:
+    # The locked AMD asset is attached to a draft until release publication.
+    permissions:
+      contents: write
+      actions: read
     needs: validate
     name: build ${{ matrix.id }}
     strategy:
@@ -97,9 +101,22 @@ jobs:
           path: dist/npu-sdk
           run-id: ${{ inputs.npu_sdk_run_id }}
           github-token: ${{ github.token }}
-      - name: Verify NPU SDK inventory and runtime identities
-        if: matrix.id == 'linux-x64' && inputs.npu_sdk_run_id != ''
-        run: python tools/npu/prepare_release.py --sdk-root dist/npu-sdk --github-env "$GITHUB_ENV"
+      - name: Prepare packaged Intel and opt-in AMD NPU payloads
+        if: matrix.id == 'linux-x64'
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          if [[ ! -d dist/npu-sdk/openvino ]]; then
+            python tools/npu/build_openvino.py \
+              --package-cache .cache/npu-packages --output-dir dist/npu-sdk/openvino
+          fi
+          if [[ ! -d dist/npu-sdk/amdnpu ]]; then
+            python tools/npu/fetch_amdnpu.py \
+              --package-cache .cache/npu-packages --output-dir dist/npu-sdk/amdnpu
+          fi
+          python tools/npu/prepare_release.py --sdk-root dist/npu-sdk --github-env "$GITHUB_ENV" \
+            --require-provider openvino --require-provider amdnpu
       - name: Fetch the pinned PDF renderer for npm bundling (Unix)
         if: runner.os != 'Windows'
         shell: bash
@@ -343,7 +360,7 @@ jobs:
           npm install --offline --ignore-scripts --no-audit --no-fund --package-lock=false \
             ../dist/release/arcships-light-ocr-model-ppocrv6-small-0.3.4.tgz \
             ../dist/release/arcships-light-ocr-${{ matrix.native }}-${RELEASE_VERSION}.tgz \
-            ../dist/release/arcships-light-ocr-runtime-0.1.8.tgz \
+            ../dist/release/arcships-light-ocr-runtime-0.1.9.tgz \
             ../dist/release/arcships-light-ocr-${RELEASE_VERSION}.tgz
           cp ../tools/npm/smoke.cjs .
       - name: Run stable Small OCR
@@ -377,8 +394,8 @@ jobs:
           npm install --offline --ignore-scripts --no-audit --no-fund --package-lock=false \
             ../dist/release/arcships-light-ocr-model-ppocrv6-tiny-0.1.0.tgz \
             ../dist/release/arcships-light-ocr-model-ppocrv6-medium-0.1.0.tgz \
-            ../dist/release/arcships-light-ocr-tiny-0.1.7.tgz \
-            ../dist/release/arcships-light-ocr-medium-0.1.7.tgz
+            ../dist/release/arcships-light-ocr-tiny-0.1.8.tgz \
+            ../dist/release/arcships-light-ocr-medium-0.1.8.tgz
           node ../tools/npm/smoke-tier.cjs @arcships/light-ocr-tiny light-ocr-tiny \
             ppocrv6-tiny-onnx-20260722.1
           node ../tools/npm/smoke-tier.cjs @arcships/light-ocr-medium light-ocr-medium \
@@ -412,7 +429,7 @@ jobs:
           docker run --rm -v "$PWD:/work" -w /work
           -e LIGHT_OCR_SMOKE_FIXTURE=/work/corpus/fixtures/generated-hello-123
           alpine:3.22 sh .github/scripts/smoke-musl.sh
-          "${{ matrix.id }}" "${{ inputs.version }}" "0.1.8"
+          "${{ matrix.id }}" "${{ inputs.version }}" "0.1.9"
 
   publish:
     needs: [test-packages, smoke-musl]
@@ -463,7 +480,7 @@ jobs:
           cd registry-document
           npm init --yes
           npm install --ignore-scripts --no-audit --no-fund --package-lock=false \
-            "@arcships/light-ocr-document@0.1.4"
+            "@arcships/light-ocr-document@0.1.5"
           node ../tools/npm/document-smoke.cjs
       - name: Verify the published stable package offline
         shell: bash
diff --git a/.github/workflows/npu-native.yml b/.github/workflows/npu-native.yml
index 89556dc..94c25dd 100644
--- a/.github/workflows/npu-native.yml
+++ b/.github/workflows/npu-native.yml
@@ -20,6 +20,10 @@ permissions:
 
 jobs:
   build:
+    # Draft release assets are visible only to tokens with push access.
+    permissions:
+      contents: write
+      actions: read
     runs-on: ubuntu-24.04
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
@@ -39,6 +43,11 @@ jobs:
           path: dist/npu-sdk/amdnpu
           run-id: ${{ inputs.amd_sdk_run_id }}
           github-token: ${{ github.token }}
+      - name: Fetch locked AMD deployment runtime and compiled recognition models
+        if: inputs.amd_sdk_run_id == ''
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: python tools/npu/fetch_amdnpu.py --package-cache .cache/npu --output-dir dist/npu-sdk/amdnpu
       - name: Build the native addon with packaged NPU SDKs
         shell: bash
         run: |
diff --git a/CHANGELOG.md b/CHANGELOG.md
index fb732fc..398b4c6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,27 +4,32 @@ This file records user-visible changes to `light-ocr`. Published artifact detail
 
 ## [Unreleased]
 
+## [0.5.9] - Unreleased
+
+Release preparation; npm publication and GitHub Release are pending.
+
 ### Added
 
-- Added a source-level Intel NPU backend (`provider: "openvino"`) for Linux x64.
+- Added a bundled Intel NPU backend (`provider: "openvino"`) for Linux x64.
   Builds configured with `LIGHT_OCR_OPENVINO_SDK_DIR` load the OpenVINO C
   runtime at run time, run detection and 20-bucket recognition on the NPU, and
   place `openvino` first in the package Auto policy
   (`openvino → webgpu → cpu`); hosts without an NPU skip it with
   `adapter_unavailable`. Added a pinned native SDK extractor, package-local
   runtime descriptor 2.1, numeric driver/compiler checks, and reviewed-evidence
-  release staging with optional device reports. Published npm packages do not include OpenVINO yet, so
-  `provider: "openvino"` returns `unsupported_capability` there. See
+  release staging with optional device reports. The 0.5.9 Linux x64 build includes
+  pinned OpenVINO 2026.4.0 libraries. See
   [Intel NPU 加速技术方案](docs/intel-npu-acceleration.md).
-- Added a source-level AMD NPU backend (`provider: "amdnpu"`) for Linux x64
+- Added a bundled AMD NPU backend (`provider: "amdnpu"`) for Linux x64
   glibc and Ryzen AI STX/KRK devices. It isolates the vendor ORT runtime,
   loads 20 immutable BF16 recognition contexts, and runs detection on CPU.
   Added model compilation and SDK import tools, descriptor validation,
   license/SBOM staging, and candidate/release workflows. AMD is excluded
   from Auto and enabled only through `execution.provider: "amdnpu"` (or
   `--provider amdnpu`). NPU device reports are optional for release;
-  `deviceValidated` remains false. Current published packages
-  return `unsupported_capability`. See [AMD NPU 加速技术方案](docs/amd-npu-acceleration.md)
+  `deviceValidated` remains false. The 0.5.9 Linux x64 build includes Ryzen AI 1.8
+  deployment libraries and real compiled contexts bound to Small 0.3.4.
+  Other platforms retain their existing providers. See [AMD NPU 加速技术方案](docs/amd-npu-acceleration.md)
   and [NPU SDK 构建与发布](docs/npu-runtime-release.md).
 
 ## [0.5.8] - 2026-09-24
diff --git a/README.md b/README.md
index cf4dec3..e17cfd9 100644
--- a/README.md
+++ b/README.md
@@ -235,6 +235,8 @@ selection guidance, and exit code reference:
 - [Build and release](docs/build-and-release.md)
 - [Roadmap](docs/roadmap.md)
 - [Changelog](CHANGELOG.md)
+- [npm 0.5.9 release preparation — English](docs/releases/npm-0.5.9.en.md)
+- [npm 0.5.9 发布准备 — 中文](docs/releases/npm-0.5.9.md)
 - [npm 0.5.8 release record — English](docs/releases/npm-0.5.8.en.md)
 - [npm 0.5.8 发布记录 — 中文](docs/releases/npm-0.5.8.md)
 - [npm 0.5.7 release record — English](docs/releases/npm-0.5.7.en.md)
diff --git a/README.zh-CN.md b/README.zh-CN.md
index fd18bc9..3585f2b 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -226,6 +226,8 @@ AI Agent 使用。它提供场景驱动的工作流、命令选择建议和退
 - [构建与发布](docs/build-and-release.md)
 - [路线图](docs/roadmap.md)
 - [更新日志](CHANGELOG.md)
+- [npm 0.5.9 release preparation — English](docs/releases/npm-0.5.9.en.md)
+- [npm 0.5.9 发布准备 — 中文](docs/releases/npm-0.5.9.md)
 - [npm 0.5.8 release record — English](docs/releases/npm-0.5.8.en.md)
 - [npm 0.5.8 发布记录 — 中文](docs/releases/npm-0.5.8.md)
 - [npm 0.5.7 release record — English](docs/releases/npm-0.5.7.en.md)
diff --git a/VERSION b/VERSION
index b35b9dd..416bfb0 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-0.5.8
\ No newline at end of file
+0.5.9
diff --git a/bindings/node/test/cli.test.cjs b/bindings/node/test/cli.test.cjs
index 67f2089..bca89c2 100644
--- a/bindings/node/test/cli.test.cjs
+++ b/bindings/node/test/cli.test.cjs
@@ -108,7 +108,7 @@ test('info: --version is metadata-only and reports the Small tier', async () =>
   assert.equal(code, EXIT.success);
   assert.equal(stderr, '');
   const info = JSON.parse(stdout);
-  assert.equal(info.core, '0.5.8');
+  assert.equal(info.core, '0.5.9');
   assert.equal(info.tier, 'small');
   assert.equal(info.maturity, 'stable');
   assert.equal(info.model, 'ppocrv6-small-native-20260719.1');
diff --git a/docs/amd-npu-acceleration.md b/docs/amd-npu-acceleration.md
index 12875ab..b78b34c 100644
--- a/docs/amd-npu-acceleration.md
+++ b/docs/amd-npu-acceleration.md
@@ -1,6 +1,6 @@
 # AMD NPU 加速技术方案
 
-更新时间：2026-10-06。状态：Linux x64 glibc 后端与打包改造已实现，默认关闭、配置显式开启；Ryzen AI SDK 与编译产物尚缺，真机验收不再作为发布前置。当前已发布 npm 包不含 AMD NPU 运行时。
+更新时间：2026-10-06。状态：Linux x64 glibc 后端与打包改造已实现，默认关闭、配置显式开启；0.5.9 发布输入已包含 Ryzen AI 1.8 的原生部署库与 Small 0.3.4 的 20 个真实 BF16 context。无需真机即可完成编译，尚未执行 AMD 设备推理。正式 npm 发布仍待完成。
 
 ## 范围与模型路线
 
@@ -16,22 +16,21 @@ BF16 路线由 VAIML 编译 FP32 ONNX，生成可嵌入权重与微码的 EPCont
 - 校验源识别模型哈希、VAIML 配置和 20 个静态宽度桶；识别形状为 `[1,3,48,width]`，批次为 1。编译结果的字典类别数必须与 bundle 一致。
 - 设备检查要求可读写的 AMD STX/KRK `/dev/accel` 节点；无设备返回 `adapter_unavailable`。库或模型哈希不符属于不可恢复错误，不能静默跳到 CPU。
 - detector 使用 CPU，recognition 使用 `VitisAIExecutionProvider` 和其允许的 CPU 分区。只有首次会话及对应桶实际创建成功后才选中该候选；后续推理失败直接报告错误。
-- `tools/npu/compile_amdnpu.py`：在 vendor SDK Python 环境编译所有桶，要求存在嵌入式 VitisAI EPContext，并重新打开部署产物。
+- `tools/npu/compile_amdnpu.py`：在 vendor SDK Python 环境编译所有桶，要求存在嵌入式 VitisAI EPContext，并检查输入输出类型与形状；编译过程不执行模型推理，不要求真机验收。
 - `tools/npu/import_amdnpu.py`：从已取得的 SDK 部署目录导入原生库，设置包内 RPATH、检查 ELF 依赖闭包、记录变更前后哈希，并锁定模型、配置及许可证。
 - CMake、Node descriptor 2.1、JS loader、npm staging、SBOM 和候选/正式发布流程均已接入。构建入口与报告格式见 [NPU SDK 构建与发布](npu-runtime-release.md)。
 
 ## 默认关闭与发布策略
 
-AMD 始终排除在 Auto 候选列表之外；用户通过 `execution.provider: "amdnpu"`、CLI `--provider amdnpu` 或 C++ 的 `ExecutionProvider::amdnpu` 显式开启。构建必须含 vendor SDK 与匹配的识别模型；运行时未配置 AMD 时不装载 vendor ORT。
+AMD 始终排除在 Auto 候选列表之外；用户通过 `execution.provider: "amdnpu"`、CLI `--provider amdnpu` 或 C++ 的 `ExecutionProvider::amdnpu` 显式开启。0.5.9 Linux x64 发布构建默认附带 vendor SDK 与匹配的 Small 识别模型；运行时未配置 AMD 时不装载 vendor ORT。
 
 按维护者决定，NPU 真机验收不再阻塞发布。SDK 导入的 `qualificationOnly: true` 仅记录尚无设备证据，平台包可随其他已满足条件的运行时一起发布；`deviceValidated` 保持 `false`。库存、模型哈希、依赖闭包和许可证校验继续执行，设备报告可选。
 
 ## 尚未验证的设备行为（可选后续工作）
 
-1. 取得 Ryzen AI 1.8 Linux SDK 和 STX/KRK 主机；确认独立 namespace 装载、XRT/amdxdna 驱动及 VitisAI C API 20 能协同工作。
-2. 编译真实 PP-OCRv6 模型，核对 20 桶的 NPU 分区与 FP32 CPU goldens；当前没有生成或伪造这些产物。
-3. 对完整 locked corpus 记录逐行文本、框与置信度差异，审阅 BF16 误差；测量冷/热延迟、CPU 时间和 RSS。
-4. 覆盖 20 次关闭/重建、驱动兼容、权限不足、无设备、损坏产物，以及与 WebGPU 共存和回退。
-5. 核实 SDK/模型再分发条款和完整第三方 notices，再创建 accepted SDK。
+1. 在 STX/KRK 主机确认独立 namespace 装载、XRT/amdxdna 驱动及 VitisAI C API 20 协同工作。
+2. 比较已编译 20 桶模型与 FP32 CPU goldens，记录文本、框、置信度及 BF16 误差。
+3. 测量冷/热延迟、CPU 时间和 RSS，覆盖关闭/重建、权限不足、驱动兼容及 WebGPU 共存。
 
+SDK 的原始与打包后哈希、依赖闭包、许可证/第三方 notices 和源模型绑定已随部署包记录；这不等同于真机验收或独立法律审阅。
 这些设备行为尚无实测结论，不再作为源码合入或 AMD 显式开启支持发布的前置条件。
diff --git a/docs/implementation-status.md b/docs/implementation-status.md
index 3615059..25bc695 100644
--- a/docs/implementation-status.md
+++ b/docs/implementation-status.md
@@ -1,8 +1,10 @@
 # C++ Core 与 Node-API 实施状态
 
-更新时间：2026-09-24<br>
+更新时间：2026-10-06<br>
 结论：npm `0.5.8` 已发布并晋升 `latest`。本版新增 musl（Alpine）Linux 的 CPU-only 预编译平台包（`@arcships/light-ocr-linux-x64-musl`/`-arm64-musl`，平台包总数八），其 onnxruntime 1.22.0 musl 产物从源码构建、由 `models/deps.lock.json` 以 SHA-256 锁定于 [`musl-runtime-1.22.0`](https://github.com/arcships/light-ocr/releases/tag/musl-runtime-1.22.0)；同时所有包的 `engines.node` 放宽为下限式 `>=22.0.0`（Node-API 稳定 ABI，Tier 1 仍为 22/24 LTS）。实现经 [PR #63](https://github.com/arcships/light-ocr/pull/63) 两轮独立 review；[发布 run 36013862673](https://github.com/arcships/light-ocr/actions/runs/36013862673) 完成八平台构建、离线 smoke、registry 发布与回装，[晋升 run 36018294655](https://github.com/arcships/light-ocr/actions/runs/36018294655) 只提升稳定闭包。完整证据见 [npm 0.5.8 发布记录](releases/npm-0.5.8.md)（[English](releases/npm-0.5.8.en.md)）与 [`v0.5.8` GitHub Release](https://github.com/arcships/light-ocr/releases/tag/v0.5.8)。上一版 `0.5.7`（macOS 下游重签名兼容）的证据链见 [npm 0.5.7 发布记录](releases/npm-0.5.7.md)（[English](releases/npm-0.5.7.en.md)）与 [`v0.5.7` GitHub Release](https://github.com/arcships/light-ocr/releases/tag/v0.5.7)。
 
+发布准备：`0.5.9` 已同步 Core/native、runtime `0.1.9`、Document `0.1.5`、Tiny/Medium `0.1.8`。Linux x64 默认带 Intel NPU SDK；AMD 默认关闭且需要真实 SDK/模型输入。NPU 真机报告不作为发布前置；npm 和公开 GitHub Release 尚未发布。见 [0.5.9 发布准备](releases/npm-0.5.9.md)（[English](releases/npm-0.5.9.en.md)）。
+
 状态含义：
 
 - **Done**：代码存在，并有本地实际运行证据。
@@ -88,6 +90,10 @@
 | 高分辨率峰值内存 | Done | Release 原生独立进程本机参考：2048² 空白 `318.8 MiB ≤ 384 MiB`；xfund 密集表单 116 框 `400.5 MiB ≤ 640 MiB`。四平台 release jobs 的真实模型与 RSS gates 均通过。 |
 | Tiled 高分辨率准确模式 | Done / `0.2.0` published | 1280 tile、2048→4-pass row-major、全局 candidate ceiling、IoU/IOS greedy merge、原图 recognition、C++/Node contract、8-fixture/196-line corpus、独立 oracle、四平台 36-entry accepted baseline 与 package smoke 均已完成。 |
 
+## 0.5.9 NPU 发布准备（2026-10-06）
+
+Linux x64 glibc 双 NPU Core/Node 构建与 npm native staging 已完成。默认发布输入包含 OpenVINO 2026.4.0，以及 Ryzen AI 1.8 的 20 个原生部署库和绑定 Small 0.3.4 的 20 个真实 BF16 context。Intel 进入 Auto；AMD 始终默认关闭，由配置显式开启。未执行 AMD 设备推理，`deviceValidated` 保持 false；真机验收不作为发布前置。版本、归档哈希与后续发布步骤见 [0.5.9 发布记录](releases/npm-0.5.9.md)。
+
 ## 本机最终验证快照
 
 环境：macOS arm64 Apple M4 Max，macOS 26.5.1，Apple Clang 21.0.0，CMake 4.2.1，macOS deployment target 13.3；CPU 使用 ONNX Runtime，Apple 候选使用系统 Core ML。
diff --git a/docs/npu-runtime-release.md b/docs/npu-runtime-release.md
index edfe99a..01fc162 100644
--- a/docs/npu-runtime-release.md
+++ b/docs/npu-runtime-release.md
@@ -1,6 +1,6 @@
 # NPU SDK 构建与发布
 
-当前实现覆盖 Linux x64 glibc。Intel 使用 OpenVINO C API；AMD 使用独立装载的 Ryzen AI ORT 与预编译 BF16 识别模型。安装后的 OCR 进程无需 Python，系统仍需厂商 NPU 驱动。当前已发布 npm 包未包含这些候选运行时。
+当前实现覆盖 Linux x64 glibc。Intel 使用 OpenVINO C API；AMD 使用独立装载的 Ryzen AI ORT 与预编译 BF16 识别模型。安装后的 OCR 进程无需 Python，系统仍需厂商 NPU 驱动，AMD 路线使用 Ryzen AI 1.8 支持的 Ubuntu 24.04/glibc 环境。0.5.9 的 Linux x64 发布构建默认附带两套运行时，AMD 默认关闭。
 
 ## Intel 候选 SDK
 
@@ -15,7 +15,18 @@ python tools/npu/build_openvino.py \
 
 ## AMD 候选 SDK
 
-先自行取得 Ryzen AI 1.8 Linux SDK，并在该 SDK 的 Python 环境执行：
+常规发布直接取得已锁定的部署包，不需要安装完整 SDK：
+
+```bash
+python tools/npu/fetch_amdnpu.py \
+  --package-cache .cache/npu-packages --output-dir dist/npu-sdk/amdnpu
+```
+
+`amdnpu.lock.json` 锁定归档大小、SHA-256 与完整 artifact set。发布草稿阶段，工具通过已授权的 `GH_TOKEN` 和 `gh release download` 取得附件；公开后直接下载。缓存可使用 `--offline`。官方来源及其原始库哈希保存在 `amdnpu-source.lock.json`。
+
+随包提供 Ryzen AI 1.8/ORT 1.27 的 20 个原生库，以及绑定 Small 0.3.4 的 20 个 EPContext。大型库会增加 Linux x64 包体积；Tiny/Medium 不能复用这些识别 context。附带的 `libpython3.12` 是 vendor 库的原生依赖，终端用户无需安装 Python 解释器。
+
+需要重新编译模型时，从官方 SDK 建立隔离 Python 环境后执行：
 
 ```bash
 python tools/npu/compile_amdnpu.py \
@@ -24,7 +35,7 @@ python tools/npu/compile_amdnpu.py \
   --output-dir dist/amdnpu-models
 ```
 
-默认配置为 `tools/npu/vaip_config.json`（VAIML BF16）；可用 `--configuration` 指定审阅后的配置。编译和重新打开 context 需要可用的 vendor 栈及设备，尚未在本项目的 AMD 真机上执行。
+默认配置为 `tools/npu/vaip_config.json`（VAIML BF16）；可用 `--configuration` 指定审阅后的配置。编译仅创建 vendor 编译会话并检查嵌入式 EPContext，不执行推理。20 桶已在无 AMD 设备的主机完成编译；设备推理尚未验证。
 
 在安装了 `patchelf` 和 `readelf` 的 Linux 主机导入完整部署目录：
 
@@ -37,7 +48,7 @@ python tools/npu/import_amdnpu.py \
   --output-dir dist/npu-sdk/amdnpu
 ```
 
-部署目录需含 vendor ORT、VitisAI provider 及所有配套原生库。导入工具只允许 OS 和已安装的 NPU 驱动组件留在包外；不安装驱动、不修改系统库或 `LD_LIBRARY_PATH`。应先确认提供的许可证允许再分发。
+部署目录需含 vendor ORT、VitisAI provider 及所有配套原生库。导入工具只允许 OS 和已安装的 NPU 驱动组件留在包外；设置 `$ORIGIN:/opt/xilinx/xrt/lib` RPATH 并清除 vendor ELF 的 executable-stack 标记，原始/变更后哈希均记录；不安装驱动、不修改系统库或 `LD_LIBRARY_PATH`。应先确认提供的许可证允许再分发。
 
 ## 编译 Core 与 Node
 
@@ -68,7 +79,7 @@ const engine = await createEngine({
 
 模型无关的 `@arcships/light-ocr-runtime` 还需传入 `bundlePath`；模型 facade 自动提供 bundle。CLI 使用 `--provider amdnpu`；C++ 使用 `options.execution.provider = ExecutionProvider::amdnpu`。省略该配置或指定 `auto` 会走默认候选列表。显式 AMD 请求不静默回退；缺少包内 AMD SDK/模型、设备或驱动时报告对应错误。
 
-发布不要求 NPU 真机报告；`deviceValidated` 保持 `false`。SDK 和真实编译模型仍必须取得并随构建交付，此设置不会生成缺失的 vendor 产物。
+发布不要求 NPU 真机报告；`deviceValidated` 保持 `false`。SDK 和真实编译模型通过锁定的部署包自动随构建交付。
 
 ## 可选设备报告
 
@@ -108,9 +119,9 @@ python tools/npu/accept_qualification.py \
 
 ## CI 与 npm staging
 
-`NPU native candidates` workflow 生成 Intel SDK 与 Node 候选构建。可输入另一个 run 的 `amdnpu-candidate-sdk` artifact，内容应为 AMD SDK 目录。始终上传可分发的 `npu-release-sdks`，无需设备报告。可选的 `npu-reviewed-reports` artifact 按 `openvino/*.json`、`amdnpu/*.json` 放置；仅在主动提供报告时才检查其绑定关系。
+`NPU native candidates` workflow 默认取得 Intel 与 AMD SDK 并生成双 NPU Node 候选构建。可输入另一个 run 的 `amdnpu-candidate-sdk` artifact，内容应为 AMD SDK 目录。始终上传可分发的 `npu-release-sdks`，无需设备报告。可选的 `npu-reviewed-reports` artifact 按 `openvino/*.json`、`amdnpu/*.json` 放置；仅在主动提供报告时才检查其绑定关系。
 
-`npm release` workflow 的可选 `npu_sdk_run_id` 读取同仓库的 `npu-release-sdks`（根目录下为 `openvino/`、`amdnpu/`），仅 Linux x64 构建使用。CI 在配置 CMake 前验证 SDK 库存、模型、配置和哈希；已附带的接受报告仍会校验，但不要求提供报告。
+从 0.5.9 发布准备起，`npm release` 默认为 Linux x64 准备锁定的 Intel 与 AMD SDK，缺少任一套均阻止发布构建。可选 `npu_sdk_run_id` 读取同仓库的 `npu-release-sdks`（根目录下为 `openvino/`、`amdnpu/`），仅 Linux x64 构建使用。CI 在配置 CMake 前验证 SDK 库存、模型、配置和哈希；已附带的接受报告仍会校验，但不要求提供报告。
 
 本地 `tools/npm_release.py stage-native` 接受 `--openvino-sdk-dir` 和 `--amdnpu-sdk-dir`；SDK 需与构建 addon 使用的版本、资格 ID 和包含的 provider 一致。NPU SDK 的 `qualificationOnly` 记录其设备证据状态，不再使整个平台包变成 qualification build；无需真机报告即可 staging。WebGPU 自身的既有资格门槛仍适用，正式 `assemble` 仍拒绝基础运行时的 qualification build。许可证、SBOM、SDK manifest 和可选报告随平台包保存。
 
diff --git a/docs/releases/npm-0.5.9.en.md b/docs/releases/npm-0.5.9.en.md
new file mode 100644
index 0000000..73226e8
--- /dev/null
+++ b/docs/releases/npm-0.5.9.en.md
@@ -0,0 +1,57 @@
+# npm 0.5.9 Release Preparation
+
+[中文版](npm-0.5.9.md)
+
+Status: preparation only. npm publication, a public tag, artifact hashes and registry promotion are pending. The GitHub Release is a draft; previous release measurements are not reused.
+
+## Versions
+
+NPU implementation: [PR #67](https://github.com/arcships/light-ocr/pull/67), merged. Preparation: [PR #68](https://github.com/arcships/light-ocr/pull/68); the GitHub draft exists. The release source must be the main commit after merging version preparation.
+
+| Package | Prepared version | Planned tags |
+| --- | --- | --- |
+| Small facade | `0.5.9` | `next`, then `latest` |
+| Model-free runtime | `0.1.9` | `next`, then `latest` |
+| Eight native packages | `0.5.9` | `next`, then `latest` |
+| Document compatibility facade | `0.1.5` | `next` |
+| Tiny/Medium preview facades | `0.1.8` | `next` |
+| Small model | `0.3.4` | Reuse immutable package |
+| Tiny/Medium models | `0.1.0` | Reuse immutable packages |
+
+## Changes
+
+- Linux x64 glibc release builds include pinned OpenVINO 2026.4.0 by default. Auto tries Intel NPU before WebGPU and CPU, running detection and 20-bucket recognition when supported. Creation-time device/model/driver exclusions advance to the next provider.
+- Linux x64 glibc builds include Ryzen AI 1.8 native deployment libraries and 20 real BF16 recognition contexts bound to Small 0.3.4. AMD stays outside Auto; explicitly select `execution: { provider: "amdnpu" }` or `--provider amdnpu`. Detection uses CPU; STX/KRK devices are the target. Tiny/Medium require separately compiled contexts.
+- The workflow fetches the immutable AMD deployment archive automatically. End users need vendor drivers, without installing the full SDK or Python.
+- Hardware qualification reports are optional for NPU shipping. Unvalidated devices remain `deviceValidated: false`; inventory, hashes, model binding, licenses and existing base-runtime checks remain enforced.
+- NPU packages use runtime descriptor 2.1; other packages retain 2.0. Windows, arm64 and musl retain their existing provider routes. Model versions, the `>=22.0.0` Node floor and public API contracts are unchanged.
+
+## Release sequence
+
+1. Merge preparation and select main with Core `0.5.9`.
+2. Run `npm release` with `version=0.5.9` and `publish_to_registry=false`. Both SDKs are prepared automatically on Linux x64. An optional `npu_sdk_run_id` overrides both SDKs and must contain `openvino/` and `amdnpu/`. Draft asset downloads use the workflow GitHub token. AMD stays opt-in.
+3. Review eight-platform builds, offline image/PDF smoke and the tarball manifest. Record the source SHA, run ID, 13 new package identities and artifact hashes.
+4. Publish to `next` with `publish_to_registry=true` and record registry reinstall results.
+5. Run `npm promote` with the publication run ID and `tag=latest`; promote only Small/runtime/native. Document and preview facades stay on `next`.
+6. Update the draft release target to the final source commit, attach the manifest/artifact links, and fill publication dates and both release records before publishing `v0.5.9`.
+
+Preparation does not publish npm packages, promote tags or publish the GitHub draft. PR #67 CI is implementation evidence, not the eight-platform 0.5.9 release result. All 13 new package identities are vacant (the complete manifest also reuses two preview model packages). Intel+AMD Core/Node were reconfigured and built with version 0.5.9; version closure, lockfile and script syntax checks passed. No new local tests were run; artifact integrity and publication runs remain pending. All 20 AMD recognition contexts were actually compiled; 20 native libraries and 56 inventory files were imported. No AMD inference or NPU hardware qualification was performed.
+
+## Rollback
+
+Restore stable tags using the archived 0.5.8 publication artifact (run `36013862673`): Small/native `0.5.8`, runtime `0.1.8`. Handle compatibility/preview tags separately; never overwrite or delete published package versions.
+
+## AMD deployment artifact
+
+- Archive: `amdnpu-1.8.0-linux-x64-ppocrv6-small-0.3.4-opt-in-v1.tar.gz`
+- Size: `532,631,585` bytes (~508 MiB); unpacked SDK inventory: `1,356,635,270` bytes (~1.26 GiB). This substantially increases the Linux x64 npm package size.
+- Archive SHA-256: `bee70b00f3254943751b9460628295696d9901f22887f11b0c791191847fb016`
+- Artifact set SHA-256: `5df515842656974be5517ae40b408427f0ffdbb63473c1ced1b610f2a0afc8a4`
+- Small source recognition SHA-256: `5435fd747c9e0efe15a96d0b378d5bd157e9492ed8fd80edf08f30d02fa24634`
+- Original SDK provenance is locked in `tools/npu/amdnpu-source.lock.json`; deployment archive identity is in `tools/npu/amdnpu.lock.json`. The archive is attached to the draft release. Final npm tarball identities are generated separately by the release workflow.
+
+## GitHub Release
+
+Tag: `v0.5.9` (not public while draft).
+
+Name: `双芯入卷，轻舟再发 · Ship Intel NPU and opt-in AMD support in 0.5.9`
diff --git a/docs/releases/npm-0.5.9.md b/docs/releases/npm-0.5.9.md
new file mode 100644
index 0000000..83a5964
--- /dev/null
+++ b/docs/releases/npm-0.5.9.md
@@ -0,0 +1,69 @@
+# npm 0.5.9 发布准备
+
+[English version](npm-0.5.9.en.md)
+
+状态：准备中，尚未发布 npm、尚未创建公开 tag。GitHub Release 使用草稿；制品大小、SHA-256、发布 run 和晋升结果在实际生成后补录，不沿用上一版本数据。
+
+## 发布身份与版本
+
+- NPU 实现：[PR #67](https://github.com/arcships/light-ocr/pull/67)，已合并。
+- 发布准备：[PR #68](https://github.com/arcships/light-ocr/pull/68)；`v0.5.9` GitHub Release 草稿已创建，尚未公开。
+- 版本采用补丁递增：Core/Small/native `0.5.8 → 0.5.9`。
+- 本版本的正式源码应指向版本准备 PR 合并后的 main 提交，不能绑定旧的 0.5.8 提交。
+
+| 包 | 准备版本 | 发布标签计划 |
+| --- | --- | --- |
+| Small `@arcships/light-ocr` | `0.5.9` | 先 `next`，再 `latest` |
+| model-free runtime | `0.1.9` | 先 `next`，再 `latest` |
+| 八个 native 平台包 | `0.5.9` | 先 `next`，再 `latest` |
+| Document compatibility facade | `0.1.5` | `next` |
+| Tiny/Medium preview facade | `0.1.8` | `next` |
+| Small model | `0.3.4` | 复用已有不可变制品 |
+| Tiny/Medium model | `0.1.0` | 复用已有不可变制品 |
+
+## 用户可见变化
+
+- Linux x64 glibc 平台包默认附带锁定的 OpenVINO 2026.4.0 原生运行时；Intel NPU 在 Auto 中位于 WebGPU/CPU 前。有可用 NPU 时运行检测和 20 桶识别；无设备、模型不支持或驱动过旧时按创建阶段策略继续选择后续 provider。
+- Linux x64 glibc 发布构建默认附带 Ryzen AI 1.8 原生部署库及 Small 0.3.4 的 20 个真实 BF16 识别 context。AMD 始终默认关闭，不进入 Auto；通过 `execution: { provider: "amdnpu" }` 或 `--provider amdnpu` 显式开启。检测使用 CPU，目标为 STX/KRK；Tiny/Medium 需另行编译，不能复用 Small context。
+- 发布工作流自动下载哈希锁定的 AMD 部署归档，不需要维护者另行提供 SDK。终端用户仍需系统厂商驱动，无需 SDK/Python 安装。
+- NPU 真机验收不作为发布前置；未验证的 NPU 会话保持 `deviceValidated: false`。库存、哈希、模型绑定、许可证和基础运行时的既有发布检查保留。
+- 包内 NPU 描述采用 schema 2.1；不含 NPU 的平台保持 2.0。新增模型配置、版本与产物信息，Node loader 在装载前完成验证。
+- 其他平台继续使用现有 Apple/WebGPU/CPU 路线；Windows、arm64 和 musl 没有本次 NPU 后端。模型包版本、Node 安装下限 `>=22.0.0` 及 public API 保持既有契约。
+
+## Release 操作
+
+1. 合并版本准备 PR，确认 main 中 Core 为 `0.5.9`。
+2. 运行 `npm release` workflow，输入 `version=0.5.9`、`publish_to_registry=false`。Linux x64 默认从锁文件取得 Intel 与 AMD SDK，无需填写 NPU run ID。
+3. 可选 `npu_sdk_run_id` 仅用于覆盖默认 SDK 输入；须同时包含 `openvino/` 与 `amdnpu/`。草稿附件下载使用 workflow 的 GitHub token；保持 Release 草稿及附件可用。
+4. 审阅八平台构建、离线安装/图片/PDF smoke 与 tarball manifest；将运行 ID、精确来源提交、13 个新 package identity 和制品哈希补录本文件。
+5. 用 `publish_to_registry=true` 发布到 `next`，记录发布 run，并确认 registry 回装结果。
+6. 运行 `npm promote`：`version=0.5.9`、`release_run_id=<正式发布run>`、`tag=latest`。仅晋升 stable Small/runtime/native；Document 与 Tiny/Medium 保留 `next`。
+7. 将 GitHub Release 草稿 target 更新为正式 main 来源提交，附上 release-manifest 和公开制品链接；确认发布结果后更新 Changelog 日期、两份发布记录和 Release 状态，再发布 `v0.5.9`。
+
+准备阶段不会主动调用 registry publish、晋升 dist-tag 或公开发布 GitHub Release。
+
+## 当前证据
+
+- PR #67 的 Linux native 与 workspace CI 均通过；这属于实现来源，不是 0.5.9 八平台发布结果。
+- 13 个新 package identity 的 npm 空位检查通过；加上复用的两个 preview 模型，release manifest 仍包含 15 包。
+- Intel+AMD Core/Node addon 已按 `LIGHT_OCR_VERSION=0.5.9` 重新配置并构建通过；版本闭包、lockfile 和脚本语法检查通过，未运行新的本地测试。实际 tarball、发布/晋升 run 和 registry integrity 尚未生成。
+- AMD 20 桶已实际编译，20 个原生库与 56 个库存文件已导入；原始 SDK、打包后运行库及模型哈希均已记录。本次没有执行 AMD 推理，不安排 NPU 真机验收。
+
+## 回滚
+
+已发布 npm 版本不可覆盖。若 0.5.9 晋升后需要回滚，使用 0.5.8 的已归档正式发布 artifact（run `36013862673`）恢复 stable Small/native `0.5.8` 与 runtime `0.1.8` 的 `latest` 标签。Document 与 preview 标签单独处理；不删除用户已安装的 0.5.9 产物。
+
+## AMD 部署制品
+
+- 归档：`amdnpu-1.8.0-linux-x64-ppocrv6-small-0.3.4-opt-in-v1.tar.gz`
+- 大小：`532,631,585` bytes（约 508 MiB）；SDK 库存共 `1,356,635,270` bytes（约 1.26 GiB），Linux x64 npm 包因此显著增大。
+- 归档 SHA-256：`bee70b00f3254943751b9460628295696d9901f22887f11b0c791191847fb016`
+- Artifact set SHA-256：`5df515842656974be5517ae40b408427f0ffdbb63473c1ced1b610f2a0afc8a4`
+- Small 识别源 SHA-256：`5435fd747c9e0efe15a96d0b378d5bd157e9492ed8fd80edf08f30d02fa24634`
+- 官方 SDK 原始归档身份见 `tools/npu/amdnpu-source.lock.json`，最终部署归档身份见 `tools/npu/amdnpu.lock.json`。归档已附于 GitHub 草稿，正式 npm tarball 身份须由发布工作流另行生成。
+
+## GitHub Release
+
+Tag：`v0.5.9`（草稿阶段未公开创建）。
+
+Release name：`双芯入卷，轻舟再发 · Ship Intel NPU and opt-in AMD support in 0.5.9`
diff --git a/package-lock.json b/package-lock.json
index 1059150..ab67abc 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1093,11 +1093,11 @@
     },
     "packages/light-ocr": {
       "name": "@arcships/light-ocr",
-      "version": "0.5.8",
+      "version": "0.5.9",
       "license": "Apache-2.0",
       "dependencies": {
         "@arcships/light-ocr-model-ppocrv6-small": "0.3.4",
-        "@arcships/light-ocr-runtime": "0.1.7"
+        "@arcships/light-ocr-runtime": "0.1.9"
       },
       "bin": {
         "light-ocr": "src/cli.cjs"
@@ -1108,10 +1108,10 @@
     },
     "packages/light-ocr-document": {
       "name": "@arcships/light-ocr-document",
-      "version": "0.1.4",
+      "version": "0.1.5",
       "license": "Apache-2.0",
       "dependencies": {
-        "@arcships/light-ocr": "0.5.7"
+        "@arcships/light-ocr": "0.5.9"
       },
       "bin": {
         "light-ocr-document": "src/cli.cjs"
@@ -1122,11 +1122,11 @@
     },
     "packages/light-ocr-medium": {
       "name": "@arcships/light-ocr-medium",
-      "version": "0.1.7",
+      "version": "0.1.8",
       "license": "Apache-2.0",
       "dependencies": {
         "@arcships/light-ocr-model-ppocrv6-medium": "0.1.0",
-        "@arcships/light-ocr-runtime": "0.1.7"
+        "@arcships/light-ocr-runtime": "0.1.9"
       },
       "bin": {
         "light-ocr-medium": "src/cli.cjs"
@@ -1279,11 +1279,11 @@
     },
     "packages/light-ocr-tiny": {
       "name": "@arcships/light-ocr-tiny",
-      "version": "0.1.7",
+      "version": "0.1.8",
       "license": "Apache-2.0",
       "dependencies": {
         "@arcships/light-ocr-model-ppocrv6-tiny": "0.1.0",
-        "@arcships/light-ocr-runtime": "0.1.7"
+        "@arcships/light-ocr-runtime": "0.1.9"
       },
       "bin": {
         "light-ocr-tiny": "src/cli.cjs"
@@ -1304,20 +1304,20 @@
     },
     "packages/runtime": {
       "name": "@arcships/light-ocr-runtime",
-      "version": "0.1.8",
+      "version": "0.1.9",
       "license": "Apache-2.0",
       "engines": {
         "node": ">=22.0.0"
       },
       "optionalDependencies": {
-        "@arcships/light-ocr-darwin-arm64": "0.5.8",
-        "@arcships/light-ocr-darwin-x64": "0.5.8",
-        "@arcships/light-ocr-linux-arm64-gnu": "0.5.8",
-        "@arcships/light-ocr-linux-arm64-musl": "0.5.8",
-        "@arcships/light-ocr-linux-x64-gnu": "0.5.8",
-        "@arcships/light-ocr-linux-x64-musl": "0.5.8",
-        "@arcships/light-ocr-win32-arm64": "0.5.8",
-        "@arcships/light-ocr-win32-x64": "0.5.8"
+        "@arcships/light-ocr-darwin-arm64": "0.5.9",
+        "@arcships/light-ocr-darwin-x64": "0.5.9",
+        "@arcships/light-ocr-linux-arm64-gnu": "0.5.9",
+        "@arcships/light-ocr-linux-arm64-musl": "0.5.9",
+        "@arcships/light-ocr-linux-x64-gnu": "0.5.9",
+        "@arcships/light-ocr-linux-x64-musl": "0.5.9",
+        "@arcships/light-ocr-win32-arm64": "0.5.9",
+        "@arcships/light-ocr-win32-x64": "0.5.9"
       }
     }
   }
diff --git a/packages/light-ocr-document/package.json b/packages/light-ocr-document/package.json
index 38b97f9..d5a7e6a 100644
--- a/packages/light-ocr-document/package.json
+++ b/packages/light-ocr-document/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@arcships/light-ocr-document",
-  "version": "0.1.4",
+  "version": "0.1.5",
   "private": true,
   "description": "Compatibility entry for the document API built into light-ocr",
   "license": "Apache-2.0",
@@ -33,7 +33,7 @@
     "node": ">=22.0.0"
   },
   "dependencies": {
-    "@arcships/light-ocr": "0.5.8"
+    "@arcships/light-ocr": "0.5.9"
   },
   "scripts": {
     "test": "node --test test/*.test.cjs"
diff --git a/packages/light-ocr-medium/package.json b/packages/light-ocr-medium/package.json
index adec51b..5157a74 100644
--- a/packages/light-ocr-medium/package.json
+++ b/packages/light-ocr-medium/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@arcships/light-ocr-medium",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "private": true,
   "description": "Offline PP-OCRv6 Medium OCR for Node.js — preview quality tier",
   "license": "Apache-2.0",
@@ -29,7 +29,7 @@
   },
   "dependencies": {
     "@arcships/light-ocr-model-ppocrv6-medium": "0.1.0",
-    "@arcships/light-ocr-runtime": "0.1.8"
+    "@arcships/light-ocr-runtime": "0.1.9"
   },
   "scripts": {
     "test": "node --test test/*.test.cjs"
diff --git a/packages/light-ocr-tiny/package.json b/packages/light-ocr-tiny/package.json
index ef3dda6..98beb92 100644
--- a/packages/light-ocr-tiny/package.json
+++ b/packages/light-ocr-tiny/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@arcships/light-ocr-tiny",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "private": true,
   "description": "Offline PP-OCRv6 Tiny OCR for Node.js — preview size-first tier",
   "license": "Apache-2.0",
@@ -29,7 +29,7 @@
   },
   "dependencies": {
     "@arcships/light-ocr-model-ppocrv6-tiny": "0.1.0",
-    "@arcships/light-ocr-runtime": "0.1.8"
+    "@arcships/light-ocr-runtime": "0.1.9"
   },
   "scripts": {
     "test": "node --test test/*.test.cjs"
diff --git a/packages/light-ocr/package.json b/packages/light-ocr/package.json
index 68ecd2e..5fcf527 100644
--- a/packages/light-ocr/package.json
+++ b/packages/light-ocr/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@arcships/light-ocr",
-  "version": "0.5.8",
+  "version": "0.5.9",
   "private": true,
   "description": "Offline image and PDF OCR for Node.js — the stable PP-OCRv6 Small tier",
   "license": "Apache-2.0",
@@ -30,7 +30,7 @@
   },
   "dependencies": {
     "@arcships/light-ocr-model-ppocrv6-small": "0.3.4",
-    "@arcships/light-ocr-runtime": "0.1.8"
+    "@arcships/light-ocr-runtime": "0.1.9"
   },
   "scripts": {
     "test": "node --test test/*.test.cjs"
diff --git a/packages/runtime/README.md b/packages/runtime/README.md
index d3c316c..91b0db5 100644
--- a/packages/runtime/README.md
+++ b/packages/runtime/README.md
@@ -11,3 +11,16 @@ const engine = await createEngine({ bundlePath: '/absolute/model/bundle' });
 ```
 
 Use `@arcships/light-ocr` for stable Small, `@arcships/light-ocr-tiny` for the size-first preview, or `@arcships/light-ocr-medium` for the quality-first preview. The runtime never downloads a model.
+
+## NPU configuration
+
+Starting with the prepared 0.5.9 native release, Linux x64 glibc packages include Intel OpenVINO in Auto ahead of WebGPU/CPU. The same Linux x64 release includes Ryzen AI 1.8 native libraries and 20 compiled BF16 contexts for Small 0.3.4. AMD is excluded from Auto; enable it explicitly:
+
+```js
+const engine = await createEngine({
+  bundlePath: '/absolute/model/bundle',
+  execution: { provider: 'amdnpu' },
+});
+```
+
+These AMD contexts require the Small 0.3.4 recognition model and an installed Ryzen AI 1.8 STX/KRK driver stack. Other tiers require separately compiled contexts. Platforms without AMD payload report `unsupported_capability`. Explicit selection does not silently fall back. NPU shipping does not assert hardware qualification; `deviceValidated` stays false.
diff --git a/packages/runtime/package.json b/packages/runtime/package.json
index 7234ec9..06fe74d 100644
--- a/packages/runtime/package.json
+++ b/packages/runtime/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@arcships/light-ocr-runtime",
-  "version": "0.1.8",
+  "version": "0.1.9",
   "private": true,
   "description": "Model-free Node.js runtime for light-ocr",
   "license": "Apache-2.0",
@@ -28,14 +28,14 @@
     "node": ">=22.0.0"
   },
   "optionalDependencies": {
-    "@arcships/light-ocr-darwin-arm64": "0.5.8",
-    "@arcships/light-ocr-darwin-x64": "0.5.8",
-    "@arcships/light-ocr-linux-arm64-gnu": "0.5.8",
-    "@arcships/light-ocr-linux-arm64-musl": "0.5.8",
-    "@arcships/light-ocr-linux-x64-gnu": "0.5.8",
-    "@arcships/light-ocr-linux-x64-musl": "0.5.8",
-    "@arcships/light-ocr-win32-arm64": "0.5.8",
-    "@arcships/light-ocr-win32-x64": "0.5.8"
+    "@arcships/light-ocr-darwin-arm64": "0.5.9",
+    "@arcships/light-ocr-darwin-x64": "0.5.9",
+    "@arcships/light-ocr-linux-arm64-gnu": "0.5.9",
+    "@arcships/light-ocr-linux-arm64-musl": "0.5.9",
+    "@arcships/light-ocr-linux-x64-gnu": "0.5.9",
+    "@arcships/light-ocr-linux-x64-musl": "0.5.9",
+    "@arcships/light-ocr-win32-arm64": "0.5.9",
+    "@arcships/light-ocr-win32-x64": "0.5.9"
   },
   "scripts": {
     "test": "node --test test/*.test.cjs"
diff --git a/packages/runtime/src/load-native.cjs b/packages/runtime/src/load-native.cjs
index acad02e..09aab2b 100644
--- a/packages/runtime/src/load-native.cjs
+++ b/packages/runtime/src/load-native.cjs
@@ -209,7 +209,18 @@ function safeArtifactPath(root, value, field) {
 }
 
 function sha256(filename) {
-  return crypto.createHash('sha256').update(fs.readFileSync(filename)).digest('hex');
+  const hash = crypto.createHash('sha256');
+  const buffer = Buffer.allocUnsafe(1024 * 1024);
+  const file = fs.openSync(filename, 'r');
+  try {
+    let count;
+    while ((count = fs.readSync(file, buffer, 0, buffer.length, null)) !== 0) {
+      hash.update(buffer.subarray(0, count));
+    }
+  } finally {
+    fs.closeSync(file);
+  }
+  return hash.digest('hex');
 }
 
 function verifyArtifact(root, artifact, field) {
diff --git a/tools/npm/smoke.cjs b/tools/npm/smoke.cjs
index 113d2c9..e1f4956 100644
--- a/tools/npm/smoke.cjs
+++ b/tools/npm/smoke.cjs
@@ -112,15 +112,17 @@ async function main() {
         'cpu',
       );
     } else {
+      const candidates = process.platform === 'linux'
+        ? ['openvino', 'webgpu', 'cpu'] : ['webgpu', 'cpu'];
       assert.deepEqual(
         engine.info.execution.selectionTrace.orderedCandidates,
-        ['webgpu', 'cpu'],
+        candidates,
       );
       assert.ok(engine.info.execution.providerCapabilities.some(
         (capability) => capability.provider === 'webgpu'
           && capability.packageIncluded,
       ));
-      assert.ok(['webgpu', 'cpu'].includes(
+      assert.ok(candidates.includes(
         engine.info.execution.selectionTrace.selectedProvider,
       ));
     }
diff --git a/tools/npm_release.py b/tools/npm_release.py
index e3b8d9a..fc9ec39 100644
--- a/tools/npm_release.py
+++ b/tools/npm_release.py
@@ -63,12 +63,12 @@
     },
     "tiny": {
         "name": "@arcships/light-ocr-tiny",
-        "version": "0.1.7",
+        "version": "0.1.8",
         "workspace": "light-ocr-tiny",
     },
     "medium": {
         "name": "@arcships/light-ocr-medium",
-        "version": "0.1.7",
+        "version": "0.1.8",
         "workspace": "light-ocr-medium",
     },
 }
@@ -1229,6 +1229,8 @@ def assemble(arguments: argparse.Namespace) -> None:
         copy_tree(source / "native", package / "native")
         copy_tree(source / "pdfium", package / "pdfium")
         copy_tree(source / "licenses", package / "licenses")
+        if (source / "qualification").is_dir():
+            copy_tree(source / "qualification", package / "qualification")
         copy_file(source / "license-inventory.json", package / "license-inventory.json")
         copy_file(source / "sbom.spdx.json", package / "sbom.spdx.json")
         package_json = common_package(
@@ -1249,6 +1251,7 @@ def assemble(arguments: argparse.Namespace) -> None:
                     "native/",
                     "pdfium/",
                     "licenses/",
+                    "qualification/",
                     "license-inventory.json",
                     "sbom.spdx.json",
                     "artifact-hashes.json",
diff --git a/tools/npu/amdnpu-source.lock.json b/tools/npu/amdnpu-source.lock.json
new file mode 100644
index 0000000..8b54833
--- /dev/null
+++ b/tools/npu/amdnpu-source.lock.json
@@ -0,0 +1,171 @@
+{
+  "schemaVersion": "1.0",
+  "version": "1.8.0",
+  "package": {
+    "filename": "ryzen_ai-1.8.0.tgz",
+    "bytes": 7819506855,
+    "sha256": "7532d511532dd8842995bc8dc2edd8a45fbf261312ea9abe8dd01b4d502ac98d",
+    "url": "https://download.amd.com/opendownload/RyzenAI/1.8.0/ryzen_ai-1.8.0.tgz"
+  },
+  "nativeDependencies": [
+    {
+      "filename": "libncurses6_6.4+20240113-1ubuntu2_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/n/ncurses/libncurses6_6.4+20240113-1ubuntu2_amd64.deb",
+      "bytes": 111902,
+      "sha256": "f2b8efe3a3e2f622c5266d5a1301048306441885d2bbebc02685cb8c2d097537"
+    },
+    {
+      "filename": "libtinfo6_6.4+20240113-1ubuntu2_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/n/ncurses/libtinfo6_6.4+20240113-1ubuntu2_amd64.deb",
+      "bytes": 107794,
+      "sha256": "53e1e1753729d04cf65b05e6e58abe06e2bb76cc07eff0e1b2a638a638ca209b"
+    },
+    {
+      "filename": "libpython3.12t64_3.12.3-1_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/p/python3.12/libpython3.12t64_3.12.3-1_amd64.deb",
+      "bytes": 2338824,
+      "sha256": "31df8af8521fb3fdf198d7507765d4ebf435c7165a50013b39d2761fd6c0d98c"
+    },
+    {
+      "filename": "libuuid1_2.39.3-9ubuntu6_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/u/util-linux/libuuid1_2.39.3-9ubuntu6_amd64.deb",
+      "bytes": 35248,
+      "sha256": "4ba2c3a10469d1ab8e01a9af8e39093ed43addada658c04737287c5f91a1645b"
+    },
+    {
+      "filename": "libzstd1_1.5.5+dfsg2-2build1_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/libz/libzstd/libzstd1_1.5.5+dfsg2-2build1_amd64.deb",
+      "bytes": 299490,
+      "sha256": "4dcc3ba9606a5837a55248d08162a5b426dd621f75c0eb24d86409e0d0e71d52"
+    },
+    {
+      "filename": "libexpat1_2.6.1-2build1_amd64.deb",
+      "url": "https://archive.ubuntu.com/ubuntu/pool/main/e/expat/libexpat1_2.6.1-2build1_amd64.deb",
+      "bytes": 86992,
+      "sha256": "730b29497212289cda8a3139e693a3538af0cb57846c24f807d1bb293a591c5e"
+    }
+  ],
+  "sourceModelSha256": "5435fd747c9e0efe15a96d0b378d5bd157e9492ed8fd80edf08f30d02fa24634",
+  "libraries": [
+    {
+      "library": "libaie_codegen.so.3",
+      "path": "native-source/deployment/lib/libaie_codegen.so.3.5",
+      "bytes": 765056,
+      "sha256": "c690396adac16751fc7d99e5e6b91497c7293eab2a7027a53f083b0372289767"
+    },
+    {
+      "library": "libaiecompiler_client.so",
+      "path": "native-source/deployment/lib/libaiecompiler_client.so",
+      "bytes": 223400960,
+      "sha256": "e94e0d2136ca84f91936ba33f07c66a1b0ab3d5a503e511d7b9d234459f019c8"
+    },
+    {
+      "library": "libdyn_dispatch_core.so",
+      "path": "native-source/deployment/lib/libdyn_dispatch_core.so",
+      "bytes": 187806488,
+      "sha256": "af71fa98497bce04d3e549c620706195aa45837ceabe1f903101f1b26339c95e"
+    },
+    {
+      "library": "libexpat.so.1",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libexpat.so.1.9.1",
+      "bytes": 170240,
+      "sha256": "e62066b1ba04145a07f03de30ea91254559ac4ca22389ea4cd7f1f700d9f2ff5"
+    },
+    {
+      "library": "libflexmlrt.so",
+      "path": "native-source/deployment/lib/libflexmlrt.so",
+      "bytes": 12623576,
+      "sha256": "0fca2344557642b23b7fb9369781b782d412a227864ffa760934c01747ca7dad"
+    },
+    {
+      "library": "libncurses.so.6",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libncurses.so.6.4",
+      "bytes": 165968,
+      "sha256": "c090994a8fd216fafa8b4f1f4e1998e4b8bcce82dd28bfa2549acbd91fd561bd"
+    },
+    {
+      "library": "libonnxruntime.so",
+      "path": "native-source/deployment/lib/libonnxruntime.so.1.27.0",
+      "bytes": 28316944,
+      "sha256": "b26e093d750d8c066ff4d3daebe986b37d949de542c453ccdb578db442265555"
+    },
+    {
+      "library": "libonnxruntime_providers_shared.so",
+      "path": "native-source/deployment/lib/libonnxruntime_providers_shared.so",
+      "bytes": 15480,
+      "sha256": "f0e61e401eed9466d56da57b28f7ddd842f546d405a0077255c74b26a81ed044"
+    },
+    {
+      "library": "libonnxruntime_providers_vitisai.so",
+      "path": "native-source/deployment/lib/libonnxruntime_providers_vitisai.so",
+      "bytes": 273680,
+      "sha256": "e92a6ad323c8a426f45c1b3211b202e1f598ad031face7e402332768ed846539"
+    },
+    {
+      "library": "libonnxruntime_vitisai_ep.so",
+      "path": "native-source/deployment/lib/libonnxruntime_vitisai_ep.so",
+      "bytes": 71403752,
+      "sha256": "cf5206d4810e503e017dfac1fb0cfe89789ff9c3eae57d6db1682a040d28a09b"
+    },
+    {
+      "library": "libpeano-lib.so.21.0git",
+      "path": "venv/lib/python3.12/site-packages/lnx64.o/tools/peano/lib/libpeano-lib.so.21.0git",
+      "bytes": 186631968,
+      "sha256": "19701a4944dacf516846e42c8b0a466db4201d7a59fd385c3d3aba48603768be"
+    },
+    {
+      "library": "libpython3.12.so.1.0",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libpython3.12.so.1.0",
+      "bytes": 9042824,
+      "sha256": "a6d543df2a20bff421d6cd00eca3243042140829827751f60f55b05495a79c2f"
+    },
+    {
+      "library": "libtinfo.so.6",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libtinfo.so.6.4",
+      "bytes": 208328,
+      "sha256": "61f693cd7f8e55ad3ab518b20a00c01adb46ef829cf90773c97615a73f99e09b"
+    },
+    {
+      "library": "libuuid.so.1",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libuuid.so.1.3.0",
+      "bytes": 35032,
+      "sha256": "65b9396655f948e1eeb5d20d25829389b9d513c34840a24002dd5d417b60c507"
+    },
+    {
+      "library": "libvaiml2.so",
+      "path": "native-source/deployment/lib/libvaiml2.so",
+      "bytes": 2769136,
+      "sha256": "38636c3b141b31bd0dcb73ae86307d4a256e2aa2da0765dc6711be4e1c4b8658"
+    },
+    {
+      "library": "libxcompiler-core-without-symbol.so",
+      "path": "native-source/deployment/lib/libxcompiler-core-without-symbol.so",
+      "bytes": 64630424,
+      "sha256": "a25960aba0c8828e9bb6494c50a727790ae8172997ad2380f20a989dbc8da3f1"
+    },
+    {
+      "library": "libxcompiler-mlir-e2e.so",
+      "path": "native-source/deployment/lib/libxcompiler-mlir-e2e.so",
+      "bytes": 306756952,
+      "sha256": "b5f57ce492a1bd0a791354cc4be2cc169cd7c765c9e094b3a01f6b1486badbb4"
+    },
+    {
+      "library": "libxcompiler-xcompiler-core.so.1",
+      "path": "native-source/deployment/lib/libxcompiler-xcompiler-core.so.1.6.0",
+      "bytes": 77568,
+      "sha256": "a0956717669f8dd91ff7775fde8c97e01a65640c429ae10cdcdd56835b3c8467"
+    },
+    {
+      "library": "libz.so.1",
+      "path": "native-source/deployment/lib/libz.so.1.3.2",
+      "bytes": 117552,
+      "sha256": "08dd93987c367ec2d928ce9148821d0ced5521050bc3e3e9c6384b69d991936a"
+    },
+    {
+      "library": "libzstd.so.1",
+      "path": "ubuntu-libs/usr/lib/x86_64-linux-gnu/libzstd.so.1.5.5",
+      "bytes": 755864,
+      "sha256": "3c8a5cd1ce4008afa91da80db302b6987b2a75bb053f3a80e203068c11f7ca49"
+    }
+  ]
+}
diff --git a/tools/npu/amdnpu.lock.json b/tools/npu/amdnpu.lock.json
new file mode 100644
index 0000000..a371a30
--- /dev/null
+++ b/tools/npu/amdnpu.lock.json
@@ -0,0 +1,14 @@
+{
+  "schemaVersion": "1.0",
+  "version": "1.8.0",
+  "artifactSetSha256": "5df515842656974be5517ae40b408427f0ffdbb63473c1ced1b610f2a0afc8a4",
+  "sourceModelSha256": "5435fd747c9e0efe15a96d0b378d5bd157e9492ed8fd80edf08f30d02fa24634",
+  "package": {
+    "filename": "amdnpu-1.8.0-linux-x64-ppocrv6-small-0.3.4-opt-in-v1.tar.gz",
+    "bytes": 532631585,
+    "sha256": "bee70b00f3254943751b9460628295696d9901f22887f11b0c791191847fb016",
+    "url": "https://github.com/arcships/light-ocr/releases/download/v0.5.9/amdnpu-1.8.0-linux-x64-ppocrv6-small-0.3.4-opt-in-v1.tar.gz",
+    "releaseTag": "v0.5.9",
+    "repository": "arcships/light-ocr"
+  }
+}
diff --git a/tools/npu/compile_amdnpu.py b/tools/npu/compile_amdnpu.py
index 0ba7397..e9c7885 100644
--- a/tools/npu/compile_amdnpu.py
+++ b/tools/npu/compile_amdnpu.py
@@ -8,6 +8,7 @@
 
 import argparse
 import json
+import os
 from pathlib import Path
 import shutil
 
@@ -33,14 +34,17 @@ def validate_context(path: Path) -> int:
     return len(contexts)
 
 
-def compile_models(source: Path, expected_sha256: str, configuration: Path, output: Path) -> dict:
-    import numpy as np
+def compile_models(source: Path, expected_sha256: str, configuration: Path, output: Path,
+                   widths: list[int] | None = None) -> dict:
     import onnx
     import onnxruntime as ort
     if digest(source) != expected_sha256:
         raise ValueError("source recognition model does not match its immutable bundle hash")
     if "VitisAIExecutionProvider" not in ort.get_available_providers():
         raise ValueError("run this tool with the Ryzen AI SDK Python interpreter")
+    selected_widths = recognition_widths() if widths is None else sorted(set(widths))
+    if not selected_widths or any(width not in recognition_widths() for width in selected_widths):
+        raise ValueError("recognition widths must belong to the locked 20-bucket contract")
     config = json.loads(configuration.read_text("utf-8"))
     if (config.get("target") != "VAIML" or
             not any(item.get("name") == "vaiml_partition" for item in config.get("passes", []))):
@@ -57,7 +61,7 @@ def compile_models(source: Path, expected_sha256: str, configuration: Path, outp
     if len(source_shape) != 4:
         raise ValueError("recognition input must be NCHW")
     models = []
-    for width in recognition_widths():
+    for width in selected_widths:
         shaped = onnx.ModelProto()
         shaped.CopyFrom(model)
         for dimension, value in zip(shaped.graph.input[0].type.tensor_type.shape.dim, [1, 3, 48, width]):
@@ -67,28 +71,32 @@ def compile_models(source: Path, expected_sha256: str, configuration: Path, outp
         context_path = output / f"rec-{width}-context.onnx"
         onnx.save_model(shaped, static_path)
         options = ort.SessionOptions()
+        options.intra_op_num_threads = 1
+        options.inter_op_num_threads = 1
         options.add_session_config_entry("ep.context_enable", "1")
         options.add_session_config_entry("ep.context_file_path", str(context_path.absolute()))
         options.add_session_config_entry("ep.context_embed_mode", "1")
         session = ort.InferenceSession(str(static_path), sess_options=options,
             providers=["VitisAIExecutionProvider", "CPUExecutionProvider"],
-            provider_options=[{"config_file": str(config_path.absolute())}, {}])
-        result = session.run(None, {session.get_inputs()[0].name: np.zeros([1, 3, 48, width], dtype=np.float32)})
-        if len(result) != 1 or result[0].dtype != np.float32 or result[0].ndim != 3:
+            provider_options=[{"config_file": str(config_path.absolute()),
+                               "cache_dir": str((output / "compiler-cache").absolute()),
+                               "cache_key": f"rec-{width}",
+                               "enable_cache_file_io_in_mem": "0"}, {}])
+        inputs, outputs = session.get_inputs(), session.get_outputs()
+        if (len(inputs) != 1 or inputs[0].type != "tensor(float)" or
+                inputs[0].shape != [1, 3, 48, width] or len(outputs) != 1 or
+                outputs[0].type != "tensor(float)" or len(outputs[0].shape) != 3):
             raise ValueError("AMD compiler produced an incompatible recognition tensor")
         del session
         count = validate_context(context_path)
-        # Reopen the exact deployment artifact: compilation success alone is
-        # insufficient evidence that the embedded context can be loaded.
-        deployed = ort.InferenceSession(str(context_path), providers=["VitisAIExecutionProvider", "CPUExecutionProvider"],
-            provider_options=[{"config_file": str(config_path.absolute())}, {}])
-        deployed.run(None, {deployed.get_inputs()[0].name: np.zeros([1, 3, 48, width], dtype=np.float32)})
-        del deployed
+        # Compilation and structural validation do not execute the model.
+        # Hardware inference is a separate, optional activity.
         models.append({"width": width, "artifact": record(context_path, output), "contextCount": count})
         static_path.unlink()
-        print(f"Compiled and reopened recognition bucket {width}", flush=True)
+        print(f"Compiled recognition bucket {width} (inference not executed)", flush=True)
     manifest = {"schemaVersion": "1.0", "sourceModelSha256": expected_sha256,
                 "precision": "bf16", "runtimeVersion": ort.__version__,
+                "inferenceValidated": False,
                 "compilerConfiguration": record(config_path, output), "recognitionModels": models}
     (output / "compile-manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", "utf-8")
     return manifest
@@ -100,8 +108,19 @@ def main() -> None:
     parser.add_argument("--source-sha256", required=True)
     parser.add_argument("--configuration", type=Path, default=Path(__file__).with_name("vaip_config.json"))
     parser.add_argument("--output-dir", type=Path, required=True)
+    parser.add_argument("--width", type=int, action="append",
+                        help="Compile selected buckets; SDK import still requires all 20")
     args = parser.parse_args()
-    compile_models(args.source_model, args.source_sha256, args.configuration, args.output_dir)
+    # Vendor compiler components write signatures relative to the current
+    # directory. Keep those files out of the repository and isolate buckets
+    # when independent compiler processes run concurrently.
+    args.source_model = args.source_model.resolve()
+    args.configuration = args.configuration.resolve()
+    args.output_dir = args.output_dir.resolve()
+    working = args.output_dir.with_name(args.output_dir.name + ".work")
+    working.mkdir(parents=True, exist_ok=True)
+    os.chdir(working)
+    compile_models(args.source_model, args.source_sha256, args.configuration, args.output_dir, args.width)
 
 
 if __name__ == "__main__":
diff --git a/tools/npu/fetch_amdnpu.py b/tools/npu/fetch_amdnpu.py
new file mode 100644
index 0000000..6cafe05
--- /dev/null
+++ b/tools/npu/fetch_amdnpu.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+"""Fetch the immutable AMD deployment payload, without installing the compiler."""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+from pathlib import Path
+import shutil
+import subprocess
+import tarfile
+import urllib.error
+import urllib.request
+
+try:
+    from .sdk import digest, relative_path, validate_sdk
+except ImportError:
+    from sdk import digest, relative_path, validate_sdk
+
+LOCK = Path(__file__).with_name("amdnpu.lock.json")
+
+
+def fetch(cache: Path, output: Path, offline: bool = False) -> dict:
+    lock = json.loads(LOCK.read_text("utf-8"))
+    package = lock["package"]
+    cache.mkdir(parents=True, exist_ok=True)
+    archive = cache / package["filename"]
+    if not archive.is_file():
+        if offline:
+            raise ValueError("locked AMD deployment archive is absent from the offline cache")
+        temporary = cache / (package["filename"] + ".partial")
+        try:
+            with urllib.request.urlopen(package["url"], timeout=60) as source, temporary.open("wb") as target:
+                shutil.copyfileobj(source, target)
+        except urllib.error.HTTPError as error:
+            # The payload is attached to the release draft during preparation.
+            # gh keeps the credential on GitHub API requests, including when
+            # asset downloads redirect to a separate storage host.
+            if error.code != 404 or not (os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")):
+                raise
+            temporary.unlink(missing_ok=True)
+            subprocess.run(["gh", "release", "download", package["releaseTag"],
+                            "--repo", package["repository"], "--pattern", package["filename"],
+                            "--output", str(temporary)], check=True)
+        temporary.replace(archive)
+    if archive.stat().st_size != package["bytes"] or digest(archive) != package["sha256"]:
+        raise ValueError("AMD deployment archive identity mismatch")
+    if output.exists():
+        raise ValueError("output directory already exists; choose a fresh SDK directory")
+    output.mkdir(parents=True)
+    with tarfile.open(archive, "r:gz") as bundle:
+        seen = set()
+        for member in bundle:
+            path = relative_path(member.name)
+            if member.name in seen or not (member.isfile() or member.isdir()):
+                raise ValueError(f"invalid AMD deployment archive entry: {member.name}")
+            seen.add(member.name)
+            destination = output.joinpath(*path.parts)
+            if member.isdir():
+                destination.mkdir(parents=True, exist_ok=True)
+                continue
+            destination.parent.mkdir(parents=True, exist_ok=True)
+            with bundle.extractfile(member) as source, destination.open("wb") as target:
+                shutil.copyfileobj(source, target)
+    value = validate_sdk(output, "amdnpu")
+    if value["artifactSetSha256"] != lock["artifactSetSha256"]:
+        raise ValueError("AMD deployment artifact set does not match the lock")
+    return value
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--package-cache", type=Path, required=True)
+    parser.add_argument("--output-dir", type=Path, required=True)
+    parser.add_argument("--offline", action="store_true")
+    args = parser.parse_args()
+    value = fetch(args.package_cache, args.output_dir, args.offline)
+    print(f"AMD deployment SDK: {value['artifactSetSha256']} (hardware inference not validated)")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tools/npu/import_amdnpu.py b/tools/npu/import_amdnpu.py
index 5182135..42e05c2 100644
--- a/tools/npu/import_amdnpu.py
+++ b/tools/npu/import_amdnpu.py
@@ -19,7 +19,8 @@
     from sdk import artifact_set, digest, record, recognition_widths, validate_sdk, verify_record
 
 
-def import_sdk(runtime: Path, models: Path, licenses: list[Path], output: Path, version: str) -> dict:
+def import_sdk(runtime: Path, models: Path, licenses: list[Path], output: Path, version: str,
+               source_manifest: Path | None = None) -> dict:
     compiled = json.loads((models / "compile-manifest.json").read_text("utf-8"))
     if (compiled.get("schemaVersion") != "1.0" or compiled.get("precision") != "bf16" or
             [item["width"] for item in compiled["recognitionModels"]] != recognition_widths() or
@@ -37,14 +38,20 @@ def import_sdk(runtime: Path, models: Path, licenses: list[Path], output: Path,
     lib.mkdir()
     provenance = []
     for source in sorted(runtime.iterdir()):
-        if not re.fullmatch(r"lib[^/]+\.so(?:\.[0-9]+)*", source.name):
+        if not re.fullmatch(r"lib[^/]+\.so(?:\.[A-Za-z0-9_-]+)*", source.name):
             continue
         if not source.resolve().is_file():
             raise ValueError(f"AMD runtime entry is not a file: {source}")
         destination = lib / source.name
         shutil.copyfile(source.resolve(), destination)
         source_hash = digest(destination)
-        subprocess.run([patchelf, "--set-rpath", "$ORIGIN", str(destination)], check=True, capture_output=True)
+        # AMD's Linux wheels contain ELF files marked with an executable
+        # stack. Recent glibc rejects those files in dlopen/dlmopen. ELF
+        # metadata is adjusted alongside the package-local search path, with
+        # both the original and packaged hashes in provenance.
+        subprocess.run([patchelf, "--clear-execstack", str(destination)], check=True, capture_output=True)
+        subprocess.run([patchelf, "--set-rpath", "$ORIGIN:/opt/xilinx/xrt/lib", str(destination)],
+                       check=True, capture_output=True)
         provenance.append({"path": destination.relative_to(output).as_posix(), "sourceSha256": source_hash,
                            "packagedSha256": digest(destination)})
     names = {item.name for item in lib.iterdir()}
@@ -92,7 +99,8 @@ def import_sdk(runtime: Path, models: Path, licenses: list[Path], output: Path,
                 "files": [record(path, output) for path in sorted(output.rglob("*")) if path.is_file()],
                 "licenses": [record(path, output) for path in sorted(license_dir.iterdir())],
                 "provenance": provenance,
-                "source": {"url": "https://ryzenai.docs.amd.com/en/latest/linux.html"}}
+                "source": (json.loads(source_manifest.read_text("utf-8")) if source_manifest else
+                           {"url": "https://ryzenai.docs.amd.com/en/latest/linux.html"})}
     (output / "sdk-manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", "utf-8")
     return validate_sdk(output, "amdnpu")
 
@@ -104,9 +112,12 @@ def main() -> None:
     parser.add_argument("--license-file", type=Path, action="append", required=True)
     parser.add_argument("--output-dir", type=Path, required=True)
     parser.add_argument("--version", default="1.8.0")
+    parser.add_argument("--source-manifest", type=Path,
+                        help="Archive identities and compiler provenance for the acquired deployment payload")
     args = parser.parse_args()
-    value = import_sdk(args.runtime_dir, args.models_dir, args.license_file, args.output_dir, args.version)
-    print(f"AMD NPU SDK: {value['artifactSetSha256']} (qualification-only)")
+    value = import_sdk(args.runtime_dir, args.models_dir, args.license_file, args.output_dir, args.version,
+                       args.source_manifest)
+    print(f"AMD NPU SDK: {value['artifactSetSha256']} (hardware inference not validated)")
 
 
 if __name__ == "__main__":
diff --git a/tools/npu/prepare_release.py b/tools/npu/prepare_release.py
index f229032..7ac5f85 100644
--- a/tools/npu/prepare_release.py
+++ b/tools/npu/prepare_release.py
@@ -13,7 +13,13 @@ def main() -> None:
     parser = argparse.ArgumentParser()
     parser.add_argument("--sdk-root", type=Path, required=True)
     parser.add_argument("--github-env", type=Path, required=True)
+    parser.add_argument("--require-provider", choices=("openvino", "amdnpu"),
+                        action="append", default=[])
     args = parser.parse_args()
+    missing = [provider for provider in args.require_provider
+               if not (args.sdk_root / provider).is_dir()]
+    if missing:
+        raise ValueError(f"release is missing required NPU payloads: {', '.join(missing)}")
     found = False
     lines = []
     for provider in ("openvino", "amdnpu"):
diff --git a/tools/npu/vaip_config.json b/tools/npu/vaip_config.json
index 6dba3aa..b3be01b 100644
--- a/tools/npu/vaip_config.json
+++ b/tools/npu/vaip_config.json
@@ -4,7 +4,8 @@
     {
       "name": "vaiml_partition",
       "plugin": "vaip-pass_vaiml_partition",
-      "vaiml_config": {"optimize_level": 1, "preferred_data_storage": "vectorized"}
+      "vaiml_config": {"optimize_level": 1, "preferred_data_storage": "auto",
+                       "enable_f32_to_bf16_conversion": true}
     }
   ],
   "target": "VAIML",