feat(ai): add manual generation and reconcile endpoints

[poboll]

背景

在 Core v10/v13 的 AI 模块里，后台已经有摘要、精读、翻译、翻译词表等能力，但这些能力主要依赖正常的内容创建/更新事件触发。

如果内容是通过数据库迁移、脚本导入、直接 SQL/ORM 写入等方式进入数据库，正常 domain event 会被绕过，导致：

• 搜索索引没有对应文档，文章在搜索或后台检索中不可见；
• AI 摘要、精读、翻译不会自动入队；
• 现有按 AI 记录分组的列表接口只能看到“已经生成过 AI 记录”的文章，无法列出完全没有 AI 记录的孤儿文章；
• 翻译词表原本只从最近的手记采集 mood/weather，历史手记里的词条可能无法被补齐。

这会让迁移或数据库直写后的恢复流程比较困难。

建议方案

我这边已经在 fork 上实现了一个比较小的后端补充方案，主要思路是：

1. 不把“孤儿文章”塞进原有 grouped/list 接口，而是为 AI 模块增加 candidates 接口。这样后台前端可以列出所有可处理文章/字段，包括还没有任何 AI 记录的对象。
2. 复用已有搜索重建能力，不在 AI 模块重复实现搜索孤儿修复逻辑。
3. 新增一个 POST /ai/reconcile 作为数据库直写/迁移后的恢复入口，默认只重建搜索索引；只有显式传入 features 时才会触发昂贵的 AI 补偿任务。
4. 为摘要、精读、翻译、翻译词表统一支持手动全量/批量补做，必要时支持 force 重做。

已实现分支

Compare:
master...poboll:codex/ai-manual-actions

Fork branch:
https://github.com/poboll/mxspace-c/tree/codex/ai-manual-actions

包含两个提交：

• feat(ai): add manual summary and insights tasks
• feat(ai): add reconcile and translation candidates

主要接口变化

• Summary

  • POST /ai/summaries/task/batch
  • POST /ai/summaries/task/all
  • GET /ai/summaries/candidates

• Insights

  • POST /ai/insights/task/batch
  • POST /ai/insights/task/all
  • GET /ai/insights/candidates

• Translation

  • POST /ai/translations/task/batch
  • POST /ai/translations/task/all
  • GET /ai/translations/candidates
  • force support for manual regeneration

• Translation entries

  • GET /ai/translations/entries/candidates
  • POST /ai/translations/entries/generate supports force
  • note mood/weather collection now scans all visible notes instead of only recent notes

• Reconcile

  • POST /ai/reconcile
  • default: rebuild search documents only
  • optional features: summary, insights, translation, translation-entries
  • optional force, targetLanguages, translationEntryKeyPaths

验证

已在 fork 分支本地通过：

pnpm -C apps/core exec vitest run test/src/modules/ai/ai.controller.faux.e2e.spec.ts test/src/modules/ai/ai-translation.service.spec.ts test/src/modules/ai/translation-entry.service.spec.ts test/src/modules/ai/ai-summary.service.spec.ts test/src/modules/ai/ai-insights.service.spec.ts test/src/modules/ai/ai-public-sse-controllers.spec.ts
pnpm -C apps/core run typecheck

备注

我尝试创建 Pull Request 时 GitHub 返回权限错误：

GraphQL: poboll does not have the correct permissions to execute `CreatePullRequest` (createPullRequest)

所以先把实现分支和 compare 链接放在 issue 里，方便维护者确认方案或直接拉取分支。

[linear]

MXD-230

[Innei]

Thanks for the writeup and the fork @poboll — the root-cause analysis was correct. After reviewing the branch I went with a smaller fix instead of cherry-picking, sticking to YAGNI for the parts the codebase didn't yet need.

Shipped in e51621afd

The grouped admin lists (/ai/summaries/grouped, /ai/insights/grouped, /ai/translations/grouped) now surface orphan articles — visible posts/notes (and pages for translation) that have zero AI records — alongside the records-having groups. No new query parameter; the new behavior is the default, since the prior "records-only" behavior was the underlying bug.

Shape per row is unchanged: { article, summaries | insights | translations: [] }. Orphan rows carry an empty array. Records-having groups are ordered by latest record createdAt desc (existing behavior); orphans follow, sorted by article snowflake id desc; pagination spans both. Search filter propagates through both sides.

Implementation lives in a single service-agnostic helper apps/core/src/modules/ai/grouped-with-orphans.util.ts (the three services each lost ~50 lines of duplicated grouping code in the process). New repo helpers: findDistinctRefIds on summary/insights/translation, and DatabaseService.findAllArticlesForAIText() (posts + notes, no pages — summary/insights skip pages by design). 26 service + util specs added/updated cover records-only / straddle / orphan-only pages and the search filter.

Deliberately NOT done

• POST /ai/reconcile — composite endpoint that fans out to search + AI. The existing POST /search/rebuild plus per-feature task endpoints already cover the workflow; bundling them adds API surface without removing a real call site.
• /candidates endpoints — once /grouped includes orphans, a separate candidates listing is duplicate surface.
• SummaryAll / InsightsAll task types — a real gap (parallel to existing TranslationAll), but separable from the orphan-visibility bug. Will be its own issue/PR.
• SummaryBatch / InsightsBatch task types — admin doesn't currently submit batches; defer until the UI asks for it.
• force flag on summary/insights/translation tasks — separable feature ("regenerate existing"), not part of "make orphans visible". And the implementation in the fork has a real correctness gap: force was added to the task payload but not to computeAITaskDedupKey, so a force request silently dedupes onto a non-force pending task. Worth its own issue + design.
• translation-entries: scan all visible notes — also a real gap (currently findRecent(100)), but orthogonal to the grouped-list bug. Separable.

I'd suggest splitting the remaining items into focused issues — happy to discuss the shape of each before any further patches.