perf: use Anthropic Message Batches API for eval extraction (50% cost reduction)

[hdviettt]

Problem

LongMemEval-M extraction costs ~$1K+ (238K Haiku calls). Each call waits for a response synchronously, creating rate-limit bottlenecks.

Solution

Use Anthropic's Message Batches API for extraction calls:

• 50% cost discount on all batch-processed messages
• No rate limits — batch is processed asynchronously
• Results retrieved when ready (poll or webhook)

Implementation

1. Collect all extraction prompts (session text → extraction request)
2. Submit as a single batch via client.batches.create()
3. Poll for completion
4. Parse results and feed into the normal pipeline

Trade-offs

• Latency: batch processing takes longer per-call (up to 24h) but total throughput is higher
• Complexity: need to handle batch lifecycle (create → poll → retrieve)
• Could combine with --types filter for even more savings

Expected savings

• $1K → ~$500 for full run
• $330 → ~$165 for targeted run (temporal + preference)