feat: allow multiple recall tool calls per request (multi-turn recall)

[BYK]

Summary

Allows the model to call recall multiple times within a single turn, enabling drill-down from high-level search results into specific t:<id> source citations. Previously, the gateway stripped recall from the tools list after the first follow-up, preventing any further recall calls.

Problem

The gateway stripped the recall tool after the first recall follow-up (recall.ts:438-442). This prevented the LLM from:

1. Following up on t:<id> source citations in distillation results
2. Making multiple targeted recall queries in a single conversation turn
3. Drilling down from a general recall result to specific details

Approach

Treat recall like any other tool — stop artificially limiting it to one call per turn. The continuation path is now recall-aware, with a generous safety-net cap (MAX_RECALL_DEPTH = 10) to prevent pathological loops.

Changes

packages/gateway/src/recall.ts

• Add MAX_RECALL_DEPTH = 10 safety-net constant
• Remove tool stripping from buildRecallFollowUp() — recall stays in the tools list

packages/gateway/src/pipeline.ts

• Non-streaming path: Replace linear recall interception + defense-in-depth with a while loop. Each iteration executes recall, stores the result, builds a follow-up, and checks if the continuation also has recall. Usage accumulated across iterations.
• Streaming path: Replace inline continuation pipe + defense-in-depth suppression with a loop using a fresh RecallAwareAccumulator per continuation stream. Block indices tracked across iterations via cumulative blockOffset.

packages/gateway/src/stream/anthropic.ts

• Add blockOffset option to createRecallAwareAccumulator() — shifts all emitted block indices for continuation streams
• Add suppressMessageStart option — suppresses message_start events for continuation streams where the client already received one

Tests

• Updated buildRecallFollowUp test to verify recall is kept in tools
• Added blockOffset tests (correct index shifting, composition with suppression)
• Added suppressMessageStart tests
• Added continuation stream scenario test (simulates two chained recall follow-ups with correct cumulative block indexing)

Closes #399