Trigger Backfill

Queue semantic-search embedding generation for records (admin only).

Pass ?force=true to regenerate every record and replace its stored vectors. Without it, the run embeds only records that lack a current-model embedding.

The run covers the calling tenant's records. In a multi-tenant deployment the embedding model and width are shared by every tenant, so a change leaves each tenant to regenerate. Pass ?all_tenants=true to also queue a run for every other registered tenant; there it needs manage_tenants instead of manage_users, and other_tenants reports each run. When the calling tenant already has a run in flight, the status is already_running, job_id is that run (null when the caller may not read it) and the other tenants are still queued. This request did not queue that run and it may predate the change, so run the backfill again once it ends. A single-tenant deployment ignores the flag.

The run happens on the job queue because a full regeneration can exceed request timeouts. This endpoint returns the job id; poll GET /jobs/{job_id} for the outcome.