refactor(engine): 将 engine 从 submodule 内嵌为仓内一等组件

`vendor/MetasequoiaImeEngine` submodule 改为 `engine/` 目录，导入的是
product-lock.json 原本钉住的提交 c810d201，逐文件核对过 blob，因此这是一次
搬迁而不是引擎升级。引擎自己的嵌套 submodule（googlepinyinime-rev、utfcpp、
miniaudio）按各自 gitlink 一并展开，来源与提交见 engine/UPSTREAM.md。

Windows 专门化裁掉了不属于本产品的部分：whisper.cpp 与本地推理 provider、
macOS 示例宿主、整个 dictionary/ 建库流水线、上游 CI 与 release-please 配置。
词库仍按 product-lock.json 在构建时取发布产物。

引擎不再进 product-lock.json：没有 gitlink 就没有可比对的东西，锁里写一个上游
commit 等于宣称这棵树仍等于上游，而第一个 Windows 专门化改动落地时这句话就假了。
verify-contracts / verify-checkout 随之移除，verify-published 保留。

CI 与工具链：引擎源码并入根 scripts/format.sh（删掉引擎自己的那份，避免两套排除
列表分叉）；引擎 voice.yml 里的 silero_vad.onnx 摘要校验移到本仓 lock job；
engine-bump-triage 工作流与其测试删除；gitsubmodule 的 dependabot 频率由 daily
改为 monthly。

README、README.en、AGENTS.md、CONTRIBUTING、docs/product-release.md 及各子目录
文档同步说明新的位置与边界。

.github/ISSUE_TEMPLATE/config.yml

@@ -7,9 +7,12 @@ contact_links:
   - name: 安全漏洞 / Security vulnerability
     url: https://github.com/metasequoiaime/MSIME-Windows/security/advisories/new
     about: 疑似漏洞请私下上报，不要提交公开 issue。Report privately; do not open a public issue.
-  - name: 组词引擎与词库问题 / Engine and dictionary
+  # The engine moved in-tree, so candidate-order bugs are now this repository's own bugs and belong
+  # in the normal bug form. The dictionary data is still built and released elsewhere, so a missing
+  # word is still someone else's to fix -- that is the only part this row still routes away.
+  - name: 词库缺词 / Missing dictionary words
     url: https://github.com/metasequoiaime/MSIME-Engine/issues
-    about: 选字顺序、词库缺词等属于共享引擎与词库，请到对应仓库反馈。Candidate order and missing words belong to the shared engine.
+    about: 词库数据在仓外构建发布，缺词、错词请到对应仓库反馈；选字顺序等引擎问题请直接在本仓提 bug。Dictionary data is built elsewhere; engine behaviour belongs here.
   - name: Telegram 群组 / Telegram Group
     url: https://t.me/msimegroup
     about: 使用问题、讨论与日常交流。Questions and discussions.

.github/dependabot.yml

@@ -12,13 +12,14 @@ updates:
       prefix: "chore(deps)"
     open-pull-requests-limit: 5
 
-  # Daily rather than monthly, which is the cadence Server used before the consolidation. The
-  # engine moves far faster than the frontends and engine-bump-triage.yml exists to absorb that
-  # rate: it auto-merges the provably inert bumps and holds the rest for review.
+  # Daily was the engine's cadence: it moved far faster than the frontends, and engine-bump-triage.yml
+  # absorbed that rate by auto-merging the provably inert bumps. The engine is vendored in-tree now,
+  # so what is left here is opencc and cpp-pinyin -- third-party libraries that release a few times a
+  # year. A daily poll on those is a daily no-op.
   - package-ecosystem: "gitsubmodule"
     directory: "/"
     schedule:
-      interval: "daily"
+      interval: "monthly"
     commit-message:
       prefix: "chore(deps)"
     open-pull-requests-limit: 5

.github/workflows/ci.yml

@@ -24,30 +24,28 @@ jobs:
     runs-on: ubuntu-24.04
     timeout-minutes: 60
     steps:
-      # The dictionary validator is compared against the engine's copy of it, so this job needs the
-      # submodule's contents and not just its gitlink.
+      # The engine is vendored in-tree, so a plain checkout already carries the contracts this job
+      # compares against. opencc and cpp-pinyin are still submodules but nothing here builds.
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
-        with:
-          submodules: true
       - name: Check the shared dictionary validator
         run: python3 scripts/check-dictionary-contract.py
       - name: Check product-lock contract
         run: python3 scripts/check-product-lock-contract.py
+      # The voice activity detector is committed rather than downloaded, and it is executed inside
+      # the IME process. The engine's own voice.yml used to re-check its digest; that workflow left
+      # with the submodule, so the check moves here rather than disappearing.
+      - name: Check the committed voice model against its recorded digest
+        run: |
+          echo '1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3  engine/voice/assets/models/silero_vad.onnx' | sha256sum --check --strict
       - name: Test rejection of changed or mismatched inputs
         run: python3 -m unittest discover -s tests -p 'test_product_lock.py'
       - name: Test local environment setup in a moved checkout
         run: python3 -m unittest discover -s tests -p 'test_prepare_env.py'
       - name: Test the automatic release CI gate
         run: python3 -m unittest discover -s tests -p 'test_release_pr_ci.py'
-      - name: Test the engine-bump auto-merge guards
-        run: |
-          python3 -m pip install --quiet pyyaml
-          python3 -m unittest discover -s tests -p 'test_engine_bump_triage.py'
       - name: Read the reviewed lock
         id: lock
-        run: |
-          python3 scripts/product_lock.py verify-contracts .
-          python3 scripts/product_lock.py outputs --github-output "$GITHUB_OUTPUT"
+        run: python3 scripts/product_lock.py outputs --github-output "$GITHUB_OUTPUT"
 
   windows:
     name: Windows ${{ matrix.name }}
@@ -231,12 +229,12 @@ jobs:
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
-          submodules: true
 
       # The pages ship a copy of the engine's web contract. Before the consolidation the product
-      # lock kept that copy honest by pinning UiHtml and the engine together; inside one repository
-      # the engine submodule is the pin, so the copy is checked against it directly.
-      - name: Check the bundled contract against the engine submodule
+      # lock kept that copy honest by pinning UiHtml and the engine together; now that the engine is
+      # a directory of this repository, the copy is checked against engine/contracts/webview
+      # directly and a plain checkout is enough to do it.
+      - name: Check the bundled contract against the engine
         run: python3 ui-html/scripts/sync-contracts.py --check
 
       - name: Set up pnpm

.github/workflows/engine-bump-triage.yml

@@ -1,169 +0,0 @@
-name: Engine bump triage
-
-# Dependabot opens the engine submodule bump; this decides whether it can merge itself.
-#
-# The rule is an allowlist, not a blocklist: the bump merges on green CI only when every file the
-# engine changed across the range is provably inert. Anything else waits for a human. A blocklist
-# would have let the utfcpp bump through, because it appears in the range as a bare submodule
-# pointer rather than as a source file.
-
-on:
-  pull_request:
-    paths:
-      - 'vendor/MetasequoiaImeEngine'
-
-# Read-only by default, matching every other workflow here. The write scopes this job needs to merge
-# the bump are granted on the job instead, so a second job added later starts from the same floor
-# rather than inheriting write.
-permissions:
-  contents: read
-
-jobs:
-  triage:
-    if: github.actor == 'dependabot[bot]'
-    name: Classify the engine range
-    runs-on: ubuntu-24.04
-    timeout-minutes: 10
-    permissions:
-      contents: write
-      pull-requests: write
-    env:
-      SUBMODULE: vendor/MetasequoiaImeEngine
-      ENGINE: metasequoiaime/MSIME-Engine
-      GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      PR_URL: ${{ github.event.pull_request.html_url }}
-    steps:
-      - name: Check out source
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
-        with:
-          fetch-depth: 0
-          persist-credentials: false
-
-      - name: Resolve the engine range
-        id: range
-        env:
-          BASE_REF: ${{ github.base_ref }}
-        run: |
-          set -euo pipefail
-          base=$(git rev-parse "origin/$BASE_REF:$SUBMODULE")
-          head=$(git rev-parse "HEAD:$SUBMODULE")
-          echo "base=$base" >> "$GITHUB_OUTPUT"
-          echo "head=$head" >> "$GITHUB_OUTPUT"
-          echo "引擎 $base -> $head"
-
-      - name: Classify the changed files
-        id: classify
-        env:
-          BASE: ${{ steps.range.outputs.base }}
-          HEAD: ${{ steps.range.outputs.head }}
-        run: |
-          set -euo pipefail
-          base="$BASE"
-          head="$HEAD"
-          if [ "$base" = "$head" ]; then
-            echo "verdict=inert" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-
-          # Fail closed on anything but a plain fast-forward. The compare API answers with an empty
-          # file list whenever head is already an ancestor of base, which the emptiness test below
-          # would otherwise read as "nothing suspicious changed" and auto-merge — silently accepting
-          # a downgrade that reverts engine source the frontend depends on.
-          status=$(gh api "repos/$ENGINE/compare/$base...$head" --jq '.status')
-          echo "compare status: $status"
-          if [ "$status" != "ahead" ]; then
-            echo "verdict=review" >> "$GITHUB_OUTPUT"
-            delimiter="EOF_$(openssl rand -hex 16)"
-            {
-              echo "reason<<$delimiter"
-              echo "引擎范围不是单纯的快进（compare status: $status），需要人工确认。"
-              echo "$delimiter"
-            } >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-
-          files=$(gh api --paginate "repos/$ENGINE/compare/$base...$head" --jq '.files[].filename')
-          echo "改动文件:"; echo "$files" | sed 's/^/  /'
-          if [ -z "$files" ]; then
-            echo "verdict=review" >> "$GITHUB_OUTPUT"
-            delimiter="EOF_$(openssl rand -hex 16)"
-            {
-              echo "reason<<$delimiter"
-              echo "compare 没有返回任何文件，无法判定这次 bump 是否安全。"
-              echo "$delimiter"
-            } >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
-
-          # Provably inert: documentation, CI wiring, dependency manifests, licence text. Everything
-          # else - sources, build files, and bare paths such as a nested submodule pointer - needs a
-          # human. Note the trailing anchors: 'utfcpp' matches none of these patterns and so lands
-          # in review, which is the behaviour we want.
-          suspicious=$(echo "$files" | grep -vE '(^|/)[^/]+\.md$|^\.github/|^vcpkg\.json$|^vcpkg-configuration\.json$|^LICENSE$|^\.gitignore$' || true)
-
-          if [ -z "$suspicious" ]; then
-            echo "verdict=inert" >> "$GITHUB_OUTPUT"
-          else
-            echo "verdict=review" >> "$GITHUB_OUTPUT"
-            # A random delimiter: the file names come from another repository, so a path that
-            # happened to be the literal delimiter could otherwise close the block early.
-            delimiter="EOF_$(openssl rand -hex 16)"
-            {
-              echo "reason<<$delimiter"
-              echo "$suspicious"
-              echo "$delimiter"
-            } >> "$GITHUB_OUTPUT"
-          fi
-
-      # The product lock records the engine commit for provenance and for the release gate that
-      # refuses a commit nobody merged. A bump that moves the submodule without the lock leaves the
-      # two disagreeing, and verify-contracts fails the whole product-inputs job rather than this
-      # pull request, so say plainly which file has to move.
-      - name: Check the lock followed the submodule
-        id: lock
-        run: |
-          set -euo pipefail
-          if python3 scripts/product_lock.py verify-contracts .; then
-            echo "locked=yes" >> "$GITHUB_OUTPUT"
-          else
-            echo "locked=no" >> "$GITHUB_OUTPUT"
-          fi
-
-      - name: Enable auto-merge
-        # The lock guard belongs here. It used to be appended to the HEAD environment value instead,
-        # where it was not a condition at all: auto-merge ran whether or not the contract lock
-        # verified, and the comment below printed the commit followed by the literal text
-        # "&& steps.lock.outputs.locked == 'yes'".
-        if: steps.classify.outputs.verdict == 'inert' && steps.lock.outputs.locked == 'yes'
-        env:
-          BASE: ${{ steps.range.outputs.base }}
-          HEAD: ${{ steps.range.outputs.head }}
-        run: |
-          set -euo pipefail
-          gh pr merge --auto --squash "$PR_URL"
-          body="$RUNNER_TEMP/engine-bump-comment.md"
-          printf '引擎 `%s` → `%s` 的改动全部落在文档、CI 配置和依赖清单里，没有源码、构建文件或嵌套 submodule 指针。已开启自动合并，CI 通过后自行合入。\n' "$BASE" "$HEAD" > "$body"
-          gh pr comment "$PR_URL" --body-file "$body"
-
-      # Every value below reaches the shell through the environment, never through ${{ }}
-      # interpolation. REASON is a list of file names from another repository: spliced into the
-      # command text, a path containing a single quote would have closed the quoting and run as a
-      # command, on a runner holding a contents:write token for this repository.
-      - name: Hold for review
-        if: steps.classify.outputs.verdict == 'review'
-        env:
-          BASE: ${{ steps.range.outputs.base }}
-          HEAD: ${{ steps.range.outputs.head }}
-          REASON: ${{ steps.classify.outputs.reason }}
-        run: |
-          set -euo pipefail
-          gh pr edit "$PR_URL" --add-label "needs-review" || true
-          body="$RUNNER_TEMP/engine-bump-comment.md"
-          printf '引擎 `%s` → `%s` 的改动里有需要人看一眼的部分，未开启自动合并：\n\n```\n%s\n```\n\n源码、`CMakeLists.txt` 或嵌套 submodule 指针的变化都可能改变行为，而前端 CI 未必覆盖得到——`helpcode_enabled()` 与 `handle_character()` 的语义变更就是这样漏过去的。\n' "$BASE" "$HEAD" "$REASON" > "$body"
-          gh pr comment "$PR_URL" --body-file "$body"
-
-      - name: Ask for the lock to follow
-        if: steps.lock.outputs.locked == 'no'
-        run: |
-          set -euo pipefail
-          gh pr comment "$PR_URL" --body "引擎 submodule 动了但 \`product-lock.json\` 里的 \`engine.commit\` 没跟上，产物清单会记录错误的引擎来源，发布门禁也拦不住未合入的引擎提交。跑 \`python3 scripts/product_lock.py refresh --dictionary-tag <当前 tag>\` 或直接改这一个字段，再推上来。"

.github/workflows/release.yml

@@ -146,7 +146,6 @@ jobs:
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
-          python3 scripts/product_lock.py verify-contracts .
           python3 scripts/product_lock.py verify-published
           python3 scripts/product_lock.py outputs --github-output "$GITHUB_OUTPUT"
 
@@ -347,7 +346,6 @@ jobs:
         with:
           ref: ${{ needs.prepare.outputs.target_sha }}
           persist-credentials: false
-          submodules: true
 
       - name: Verify the shipped pages carry the engine's contract
         run: python3 ui-html/scripts/sync-contracts.py --check
@@ -407,13 +405,12 @@ jobs:
       # outside, so it lands beside the components in the one checkout.
       STAGING_ROOT: .
     steps:
-      # The helpcodes are collected out of the engine submodule now, so this job needs its contents
-      # and not just the gitlink.
+      # The helpcodes are collected out of engine/helpcode, which is a directory of this repository,
+      # so a plain checkout of the release commit already carries them.
       - name: Check out the release commit
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           ref: ${{ needs.prepare.outputs.target_sha }}
-          submodules: true
           persist-credentials: false
 
       # Some runner instances leave _work owned by SYSTEM while the job itself
@@ -489,7 +486,7 @@ jobs:
           pwsh -File ./Prepare-PackageFiles.ps1 -TargetVersion $env:TARGET_VERSION
           -RepoRoot (Resolve-Path '..').Path
           -TsfDirectory windows -ServerDirectory server -UiHtmlDirectory ui-html -NoticesDirectory .
-          -HelpCodeDirectory vendor/MetasequoiaImeEngine/helpcode
+          -HelpCodeDirectory engine/helpcode
 
       - name: Embed the exact product dependency manifest
         env: