Unfinished knowledge indexing automatically resumes after restart

What changed

Knowledge base indexing now persists every in-flight task to disk and resumes it automatically after a process restart or crash. Previously, indexing that was interrupted by quit/crash stayed stuck in "processing" forever and required the user to manually re-trigger reindex; that no longer happens.

Why this matters to the user

• Indexing that was running when the app exited (graceful quit, crash, OS shutdown) is picked up again ~1 minute after the next launch — no user action needed.
• A few side effects are visible:
  • Per-base concurrency moves from a shared 5-slot global pool to 5 slots per base with a 50-slot global cap, so importing into several bases in parallel finishes faster.
  • Transient embedding failures are now retried up to 3 times with exponential backoff before the item shows as "failed".
  • Each prepare-root (directory/sitemap scan) job has a 10-minute wall-clock cap, and each index-leaf (single file/url/note embed) job has a 5-minute wall-clock cap. Very large single files served by a slow embedding endpoint may now hit the leaf timeout where before they would run indefinitely.
• The processing status may briefly persist after force-quit; it self-heals once the recovery pass finishes after the next launch.

What the user should do

Nothing — automatic. If a previously failed indexing item is still stuck after the next launch + ~1 minute, run "Reindex" from the item's context menu as before.

Notes for release manager

• Recovery is gated by JobManager.onAllReady, which fires 60 s after the lifecycle reaches WhenReady. On a cold start this means indexing resumes ~60 s into the session, not at process start.
• The 5-minute per-leaf timeout is a deliberate change from "no timeout". If telemetry shows large-document users hitting it, raise defaultTimeoutMs on indexLeafJobHandler rather than rolling this back.
• v1 → v2 migrated knowledge bases that had in-flight indexing at upgrade time are mapped to idle / failed by KnowledgeMigrator (unchanged from prior v2 behavior). Users must manually reindex those — auto-recovery covers v2 → v2 restarts only.