[LPX-E5BD8]

Routing batches the messages that wait on one routing lane, for example a burst in a Feishu group or several messages sent quickly.

• One call per burst. A single fast-tier call (gateway::route_many) returns a JSON array with one route per message. Each message is decided on its own and handled in order, through the same pre-routed path as Thread: route messages while Steve is still answering #106.
• What stays out of a batch. Plain questions, which need no router, and direct messages. With fewer than 2 messages left, or when the call or its reply is unusable, every message is routed alone as before.
• While Steve is still answering. The batch is told what Steve is answering, and each message still goes through decide::now_while_answering.
• Edits. A message edited after its route was made is routed again.
• Single messages are unaffected. For one message the prompt is byte-identical to before (route_prompt_for with one text), so the prompt cache does not change.

Eval: batch_eval_real_model, with recorded outputs replayed in CI. 20 synthetic cases in batches of 5, on real Haiku:

• 20/20 hits, and no question became an item;
• 1.7 s per message instead of 4.7 s;
• 4 calls instead of 20.

Tests:

• steve-core: 650 passed, including new gateway and burst tests.
• steve-hub --no-fail-fast: everything passed except the known /tmp → /private/tmp failure in outputs_api, which is unrelated and passes in CI.
• clippy: clean.
• lint and the strict boundary check: passed.

🤖 Generated with Claude Code