build(installer): 打包神经整句模型并在发布流程中下载

Prepare-PackageFiles 断言并拷贝两个模型与授权声明；release workflow 在收集产物前
运行 fetch-neural-model.ps1。

.github/workflows/release.yml

@@ -478,6 +478,13 @@ jobs:
         shell: pwsh
         run: ./scripts/build-language-model.ps1
 
+      # The neural sentence models are downloaded from the chinese-ime-lm release pinned in
+      # neural-model/lock.json and checked against the SHA256 recorded there. Prepare-PackageFiles.ps1
+      # asserts they exist, so this has to run before the payload is collected.
+      - name: Download the neural sentence models
+        shell: pwsh
+        run: ./scripts/fetch-neural-model.ps1
+
       - name: Decide the signing mode
         id: signing
         shell: pwsh

installer/Prepare-PackageFiles.ps1

@@ -16,6 +16,9 @@ param(
     # 词格打分用的语言模型不随词库发布下发，而是由 scripts/build-language-model.ps1
     # 从钉住的上游语料现场转换到这个目录。
     [string]$LanguageModelDirectory = 'language-model',
+    # 神经整句模型（chinese-ime-lm 的 safetensors）从这个目录下发，与 sc.lm 同为资源根下的
+    # 只读文件。缺文件时引擎静默不出神经整句，但打包仍然必须带上，故这里断言存在。
+    [string]$NeuralModelDirectory = 'neural-model',
     # THIRD_PARTY_NOTICES.txt used to sit next to the tip's sources. In the consolidated repository
     # the notice covers the whole product and lives at the root, one level above windows/, so where
     # to read it is no longer answered by where the tip is.
@@ -76,6 +79,9 @@ $englishDb = Join-Path $RepoRoot (Join-Path $DictionaryDirectory 'out\english.db
 $othersDb = Join-Path $RepoRoot (Join-Path $DictionaryDirectory 'out\others.db')
 $languageModel = Join-Path $RepoRoot (Join-Path $LanguageModelDirectory 'sc.lm')
 $languageModelNotice = Join-Path $RepoRoot (Join-Path $LanguageModelDirectory 'NOTICE.md')
+$neuralModelDesktop = Join-Path $RepoRoot (Join-Path $NeuralModelDirectory 'sentence-model-desktop.safetensors')
+$neuralModelKeyboard = Join-Path $RepoRoot (Join-Path $NeuralModelDirectory 'sentence-model.safetensors')
+$neuralModelNotice = Join-Path $RepoRoot (Join-Path $NeuralModelDirectory 'NOTICE.md')
 
 Assert-PathExists -LiteralPath $RepoRoot -Description '源码仓库根目录'
 Assert-PathExists -LiteralPath $serverRelease -Description 'Server Release 输出目录'
@@ -134,6 +140,9 @@ if 'weight' not in names or pk != ['word', 'display']:
     Assert-PathExists -LiteralPath $languageModel `
         -Description '词格整句语言模型 sc.lm（用 scripts\build-language-model.ps1 生成）'
     Assert-PathExists -LiteralPath $languageModelNotice -Description 'libime 语言模型授权声明 NOTICE.md'
+    Assert-PathExists -LiteralPath $neuralModelDesktop -Description '神经整句模型 sentence-model-desktop.safetensors'
+    Assert-PathExists -LiteralPath $neuralModelKeyboard -Description '神经整句模型 sentence-model.safetensors'
+    Assert-PathExists -LiteralPath $neuralModelNotice -Description 'chinese-ime-lm 授权声明 NOTICE.md'
 }
 
 $targetAppData = Join-Path $PSScriptRoot 'app_data'
@@ -169,6 +178,12 @@ else {
     # 模型数据是 LGPL-2.1-or-later 的第三方作品，声明必须跟着二进制到用户磁盘上，
     # 理由同下面 rime-ice 那段。
     Copy-Item -LiteralPath $languageModelNotice -Destination (Join-Path $targetAppData 'libime-lm-NOTICE.md') -Force
+    # 神经整句模型与 sc.lm 一样落在资源目录根下，文件名由 engine/contracts/assets 清单声明。
+    # 缺文件时引擎静默不出神经整句候选（shared_sentence_model 返回 nullptr），不影响其它候选。
+    Copy-Item -LiteralPath $neuralModelDesktop -Destination (Join-Path $targetAppData 'sentence-model-desktop.safetensors') -Force
+    Copy-Item -LiteralPath $neuralModelKeyboard -Destination (Join-Path $targetAppData 'sentence-model.safetensors') -Force
+    # 权重是 Apache-2.0 的第三方作品，授权声明必须随二进制到用户磁盘上。
+    Copy-Item -LiteralPath $neuralModelNotice -Destination (Join-Path $targetAppData 'chinese-ime-lm-NOTICE.md') -Force
 
     $defaultConfigPath = Join-Path $targetAppData 'config.default.toml'
     # 出厂配置来自本仓库的 default_config，不依赖本机是否已安装输入法。