Bulk reindex DISPATCHER - queue re-processing of every document

One POST queues the full extraction pipeline (fetch → parse → chunk → embed) on EVERY known document of the knowledge base - uploaded files, connector/remote docs AND existing web pages (including auto-discovered ones) - re-processing them with the current store settings plus each document’s persisted overrides. Use after a store-level parser or chunking change to roll it out across the whole KB. No request body: per-document overrides stay on POST /documents/{document_id}/reindex.

Dispatch-only, by design. The request does only auth, a count, one atomic lease acquire and one internal batch kick, then returns 202. All per-document work runs ASYNCHRONOUSLY afterwards (batches of 100, keyset-paginated by id, chained through the platform events API so each link restarts at call depth 0). The response reports what was DISPATCHED, never real parse/embed completion - status is dispatching (or complete for an empty KB), and the terminal storage.kb.reindex.dispatch_completed event likewise reports counts of what was queued (visited, dispatched, skipped_missing_file, dispatch_failed). There is intentionally NO completion or progress tracking; a document that later fails to parse/embed surfaces through its own status, not here.

Documented characteristics of the dispatch-only design:

• Best-effort live set, not a snapshot. Documents are walked by keyset pagination during the run; rows inserted or deleted mid-run may or may not be visited.
• Burst, not throttled backpressure. Batches are spaced by a small delay only; a large KB enqueues its whole document set into the pipeline over a short window (milder than, and replacing, the legacy synchronous fan-out).
• Concurrency guard is dispatcher-only. An atomic lease prevents two overlapping reindex DISPATCHES; it does NOT guard a reindex overlapping ordinary ingestion, does not interact with the crawl pause state, and does not block on a running crawl.
• No new-page discovery. Existing/known web pages are re-fetched (depth-0); discovery of NEW pages runs on the crawl periodicity schedule, not here (there is no manual recrawl - removed 2026-07-20 as it never discovered pages).

Requires editor+.