Summary

• Closes Bug: busy_input_mode: queue silently drops messages — single-slot overwrite instead of FIFO #28503. With busy_input_mode: queue, rapid follow-up text messages were silently overwriting a single pending slot (merge_pending_message_event(..., merge_text=False)), so the agent only saw the most recent message after the current turn finished. Now _queue_or_replace_pending_event routes through the FIFO infrastructure already used by /queue (_enqueue_fifo + the existing _promote_queued_event drain hook), preserving each follow-up as its own turn in arrival order.
• Photo bursts still merge into the head slot via merge_pending_message_event (album semantics intact); only distinct-turn events fall through to the overflow tail.
• Added a per-session hard cap of 32 pending follow-ups (_BUSY_QUEUE_MAX_PENDING) so a stuck agent + rapid-fire user cannot grow the overflow unboundedly. There is no obvious precedent in the codebase for a configurable bound here; 32 is far beyond realistic conversational backlog while still trivially small in memory. Items beyond the cap are dropped with a logger.warning (matches the existing pattern used elsewhere for queue caps).
• Affects every code path that hands off to _queue_or_replace_pending_event: busy_input_mode: queue, the steer-fallback-to-queue path, the subagent demotion path ([Bug]: Sending a message while delegate_task is running kills the subagent — interrupt propagates unconditionally to children #30170), and the draining/restart paths.

Test plan

• pytest tests/gateway/test_queue_consumption.py — 14 pass (11 existing + 3 new: test_rapid_text_followups_are_queued_in_fifo_order, test_queue_respects_bounded_cap, test_photo_burst_still_merges_in_head_slot).
• pytest tests/gateway/test_busy_session_ack.py tests/gateway/test_busy_session_auth_bypass.py tests/gateway/test_interrupt_key_match.py tests/gateway/test_text_batching.py tests/gateway/test_telegram_text_batching.py tests/tools/test_interrupt.py tests/cli/test_quick_commands.py — 79 pass (no regressions in adjacent suites).
• Manual smoke on a real platform adapter (Telegram/WeChat): set display.busy_input_mode: queue, send a slow-running prompt, then fire 5 rapid follow-ups — confirm all 5 are processed sequentially after the current turn ends.