OrionAI stays up while the data index works

Changes since 2.22.1.

#Fixed

• OrionAI outage (3–4 November): drawing thumbnails for the newly spatialized statistics tables attached a full-detail boundary outline to every row of a table, which pushed Orion Core past its memory limit about every 10–30 minutes and took OrionAI down with it. Previews and thumbnails of a table now draw one simplified polygon per area: the same picture at about 1% of the size. Loading or exporting a table still gives every row with its full-detail polygon.
• The hourly spatial_plan backfill failed on every run since 2.22.0 with an import error; it now runs.

#Improved

• OrionAI runs apart from background work: Orion Core now runs as two containers from one image. The API (OrionAI, search, loading) no longer runs harvests, thumbnails, enrichment or embeddings; a new index worker does, and also draws the preview samples, each under its own memory limit. A heavy job can no longer slow down or stop OrionAI.
• Thumbnail backlog: batches are now 500 records that queue the next batch themselves, so the backlog drains continuously and a restart loses one small batch, not hours of work. The best candidates go first: never-tried records, then those closest to publishing, with spatialized tables ahead. Offline sources wait for their link check, and a source with nothing Orion can draw is retried only after the record changes.
• No crash loops: a record being drawn when the worker stopped is skipped for a day, and renders pause while the worker is above 75% of its memory limit. Both are staff settings (thumbnail_batch_size, thumbnail_memory_pause), as is thumbnail_concurrency (default now 2).