feat: batch queue for Anthropic Message Batches API (50% worker cost savings)

[BYK]

Summary

Gateway-only enhancement: accumulate non-urgent worker LLM calls and submit them via Anthropic's Message Batches API for 50% cost reduction on all background work.

• New BatchLLMClient wrapper (batch-queue.ts) intercepts LLMClient.prompt() calls. Non-urgent calls queue up; urgent calls bypass to the inner (synchronous) client immediately.
• urgent flag on LLMClient.prompt() opts — threaded through distillation.run(), distillSegment(), metaDistill(), and all gateway pipeline callers.
• Batch-safe (~80% of worker volume): incremental distillation, idle distillation, meta-distillation, curation, consolidation, worker model validation.
• NOT batch-safe (urgent=true): compaction (client blocking), overflow recovery (gradient needs result), query expansion (user waiting, 3s timeout).

How it works

prompt(system, user, {urgent: false})
  → enqueue with custom_id
  → flush timer (30s) or auto-flush at capacity (50 items)
  → POST /v1/messages/batches
  → poll timer (60s) checks status
  → stream JSONL results → resolve/reject promises

prompt(system, user, {urgent: true})
  → immediate delegation to inner LLMClient

Fallback safety

• Batch API error → fall back to synchronous for all queued items
• No API key → fall back to synchronous
• Batch exceeds max age (1h) → fall back to synchronous
• Shutdown → drain queue synchronously via inner client
• Disable entirely via LORE_BATCH_DISABLED=1

Cost impact

Worker call	Volume	Savings
Incremental distillation	~15-25% of turns	50% per call
Idle distillation	1x per idle	50%
Meta-distillation	1x per threshold	50%
Curation	1x per N turns	50%
Consolidation	Rare	50%
Worker model validation	Rare	50%
Estimated monthly savings		~$1,100

Files changed

• packages/core/src/types.ts — add urgent?: boolean to LLMClient.prompt() opts
• packages/core/src/distillation.ts — thread urgent through run(), distillSegment(), metaDistill()
• packages/core/src/search.ts — mark expandQuery() as urgent: true
• packages/gateway/src/batch-queue.ts — new BatchLLMClient wrapper
• packages/gateway/src/pipeline.ts — wire batch queue into getLLMClient(), mark urgent callers
• packages/gateway/src/index.ts — graceful shutdown via resetPipelineState()
• packages/gateway/test/batch-queue.test.ts — new 10 dedicated tests
• packages/gateway/test/helpers/harness.ts — async teardown for batch queue cleanup

754 tests pass, 0 failures.