Faster pattern clustering

Pattern clustering now uses the Anthropic Batch API for non-real-time work. Nightly runs that previously took 8–12 minutes per workspace now complete in under 3.

What changed: clustering moved from sequential real-time API calls to batched requests. The batch result poller picks up completions every two hours. No change to the output — same pattern structure, same deduplication logic, faster delivery.

Why this matters: the old approach hit rate limits for workspaces with large item volumes. The Batch API has higher throughput ceilings and costs 50% less per token, which we pass through in the LLM cost accounting.

All pattern operations (merge, track, feature status) work the same as before.