名留青册，虚位以待 · Reserve the AMD NPU provider contract

[eric8810]

Summary

• Adds the no-hardware Phase 0 deliverable from the AMD NPU pre-study: the reserved amdnpu execution provider name across the C++ enum, the Node addon, TypeScript types, CLI flags (--provider), the document CLI, and the server EXECUTION_MODE.
• No build ships an AMD NPU backend, so the contract fails fast and structurally: any runtime policy that lists amdnpu fails with provider_abi_mismatch (fatal in Auto as well, per D112 — a descriptor the Core cannot honor is a package error, never silently skipped), and explicit provider: "amdnpu" on current packages returns unsupported_capability.
• The precision contract keeps only the default (automatic); the INT8/BF16 choice is pre-study open question feat: add encoded JPEG/PNG image input #2 and deliberately not encoded yet. deviceId stays rejected.
• Includes the AMD NPU pre-study document (constraints: Ryzen AI SW is Windows-first, Vitis AI EP has no standalone C API, model compilation is Windows-only, CNN models require INT8/BF16 derived artifacts) and links it from linux-device-acceleration.md §6.
• CHANGELOG records the reservation; cli-design.md and the package README document the value as reserved.

Verification

• C++: fresh CMake configure + build (build/amdnpu-dev, Release), light_ocr_unit_tests 90/90 passed including 4 new amdnpu tests (precision contract, bundled-provider gate, abi mismatch, Auto fail-fast); ctest 4/4 passed (unit, integration, memory gate, tiled).
• Node: packages/runtime 35/35 passed, packages/light-ocr-server 20/20 passed.
• bindings/node adapter tests: 15/19 passed locally; the 3 failures require a prebuilt addon via LIGHT_OCR_NODE_BINARY and are environment preconditions, not regressions (CI builds the addon first).

Checklist

• The change is focused; relevant tests were added or updated when behavior changed, otherwise this is N/A.
• Documentation and CHANGELOG.md were updated for user-visible changes, otherwise this is N/A.
• No credentials, private OCR inputs, generated build trees, or unrelated artifacts are included.

[eric8810]

自审报告

已核实正确

• 接入点对称完整：enum（types.hpp）→ addon parse/string → TS 类型 → CLI/帮助文本 → document CLI → server EXECUTION_MODE → benchmark → docs/CHANGELOG，无遗漏。
• 失败链路完整且被测试锁定：known_provider 加 amdnpu 后，runtime descriptor 可合法声明它（valid_runtime_policy 通过），Core 候选处返回 provider_abi_mismatch；该原因不在 D112 skippable 列表，Auto 下同样 fatal——auto_policy_with_amdnpu_fails_fast_instead_of_silently_skipping 专门防止将来被改成静默跳过。
• 精度契约：仅 automatic（fp16/fp32 均拒绝），与预研待决问题 #2（INT8 vs BF16）不预设的决策一致；deviceId 拒绝。
• capability 语义：ProviderCapabilityInfo{"amdnpu", false, false, false} —— 当 policy available_providers 含 amdnpu 而 ordered_candidates 不含时创建可走 cpu 成功，此时 capability 准确反映"Core 无法兑现"；ordered 含 amdnpu 时创建必败、capability 不可达。与 openvino 的 package_included=true 差异是有意的：无后端构建不应声称包内包含。
• 测试结构与 openvino 系列完全对称（同款 test policy 形态）；本地验证：unit 90/90（含 4 个新测试）、ctest 4/4、runtime 35/35、server 20/20。
• PR 描述已按模板填写 Summary/Verification/Checklist。

小问题（非阻塞，可本 PR 修或留 follow-up）

1. docs/amd-npu-acceleration.md §7 第 1 条："D113 未覆盖的 新维度"——"的"与"新"之间多一个空格。
2. amdnpu_accepts_only_automatic_precision 只覆盖 fp16；fp32 同样应拒绝，对称测试各测了一个精度，建议本测试两个都断言。
3. tests/integration/main.cpp 的 accelerators 列表只查 webgpu/openvino：当前 builtin policy 永不含 amdnpu，正确；将来引入 LIGHT_OCR_HAS_AMDNPU 构建时需同步（预研文档 Phase B 已覆盖此边界）。
4. engine.cpp capability push 的单行初始化格式与相邻条目的多行格式略不一致，纯风格。

[eric8810]

自审报告中的问题 1（文档空格）与问题 2（fp32 断言）已在 c3b7c93 修复；补测时顺带修正了重用已 move 的 bundle 的笔误，每段断言独立构造 bundle。本地重跑 unit 90/90 通过。问题 3、4 留作记录，不需要本 PR 处理。