Add end-of-turn batching layer for same-channel messages (#91)

[strix-tkellogg]

Summary

• End-of-turn drain: after _process_event runs for a discord_message, pop every queued same-channel discord_message and process them all in a single follow-up turn. A burst of N messages now costs 2 turns instead of N.
• Newest drained event is the synthetic trigger; _render_prompt already pulls full recent-channel history, so every batched message lands in context.
• Queue accounting preserved — the drain helper mutates the backing deque directly and leaves _unfinished_tasks untouched; caller balances with task_done per drained event.
• Other-channel events and non-discord triggers are never absorbed (scheduler ticks, web messages, etc. stay in queue).

Telemetry

• New batched_turn_start event: channel_id, trigger_event_type, batch_size.
• turn_timing now carries a batched boolean — distinguishes the triggering turn from the follow-up batch leg so you can jq the events log for turns where batching actually saved work.

Closes #91. Part 1 (turn-time instrumentation) shipped in #94 so this can be evaluated against real numbers.

Test plan

• tests/test_batched_processing.py::test_same_channel_messages_drain_into_one_batched_turn — 3 queued (2 same-channel + 1 other) → exactly 2 agent invocations, batched_turn_start{batch_size: 2}, one turn_timing{batched: true}, other-channel event survives.
• tests/test_batched_processing.py::test_no_batched_turn_when_queue_has_no_same_channel_messages — unrelated queued event does not trigger batching.
• tests/test_batched_processing.py::test_non_discord_trigger_does_not_batch — scheduler tick does not absorb a matching discord_message.
• tests/test_batched_processing.py::test_drain_helper_preserves_order_of_kept_events — order-preserving filter across mixed-channel queue.
• Full suite green via pre-commit: 246 passed, 1 skipped (was 242, added 4).

🤖 Generated with Claude Code