fix(gateway): preserve depth-capped interrupt follow-ups

[embwl0x]

Summary

Fixes #61008.

When an interrupt-drain chain reaches _MAX_INTERRUPT_DEPTH, the gateway currently tries to park a remaining string interrupt_message via adapter.queue_message(...). Normal gateway adapters use _pending_messages and do not implement that legacy method, so a transcribed voice follow-up can be dropped exactly when the recursion guard fires.

This patch adds a shared GatewayRunner._requeue_interrupt_depth_pending() helper that:

• requeues a full pending MessageEvent through the existing FIFO pending-event path,
• converts a plain interrupt transcript string into a real MessageEvent when no full event remains,
• preserves FIFO ordering when another pending slot is already occupied,
• keeps the legacy queue_message() fallback only for non-standard adapters without _pending_messages.

Validation

• uv run --extra dev pytest tests/gateway/test_queue_consumption.py -q — 16 passed
• uv run --extra dev pytest tests/gateway/test_stt_config.py tests/gateway/test_telegram_audio_vs_voice.py -q — 12 passed
• uv run --extra dev pytest tests/gateway/test_busy_session_ack.py -q — 26 passed
• uv run --extra dev pytest tests/gateway/test_run_progress_interrupt.py tests/gateway/test_pending_drain_no_recursion.py tests/gateway/test_internal_event_never_interrupts_busy_session.py -q — 10 passed
• uv run --extra dev pytest tests/gateway/test_queue_command.py tests/gateway/test_busy_session_auth_bypass.py tests/gateway/test_subagent_protection_30170.py -q — 26 passed
• uv run --extra dev pytest tests/gateway -q -k "queue or interrupt or voice" — 442 passed, 7 skipped, 8515 deselected
• uv run --extra dev ruff check gateway/run.py tests/gateway/test_queue_consumption.py — passed
• python -m py_compile gateway/run.py tests/gateway/test_queue_consumption.py — passed
• git diff --check origin/main..HEAD — passed
• gitleaks git --log-opts='origin/main..HEAD' --redact . — no leaks found

Duplicate / overlap check

• Exact open PR search for 61008 in:title,body returned no results immediately before push.
• preflight_issue.py 61008 --keyword interrupt_depth --keyword _MAX_INTERRUPT_DEPTH --keyword queue_message --keyword voice-interrupt warned on broad queue/interrupt work only.
• Manually checked the closest overlaps:
  • fix(gateway): complete agent:start/end hook symmetry on interrupt/drain turns #39126 is hook start/end symmetry on interrupt drain turns.
  • fix(gateway): label goal continuations and avoid silent queued turns #58039 is goal-continuation labeling and queue notices.
  • Bug: Voice transcript causes infinite re-queue loop in TUI #54917 is a TUI voice transcript requeue loop.
  • Telegram: out-of-band messages (text and voice) interrupt current task instead of being queued #58780 is a broader Telegram feature request for auto-queueing out-of-band messages.

Those do not preserve a depth-capped gateway interrupt transcript as a pending turn.

[teknium1]

Thanks for isolating the plain-text fallback separately from the retained voice-event path.

• Current main still assigns an interrupt_message when no pending_event remains (gateway/run.py:19571-19580), then calls adapter.queue_message(...) at the depth cap (gateway/run.py:19681-19692). A repository-wide search finds no adapter implementation of that legacy method.
• The proposed helper routes the fallback into the existing FIFO event mechanism (gateway/run.py:4371-4385), and BasePlatformAdapter drains that pending event into a fresh task (gateway/platforms/base.py:5164-5201). This preserves the current non-recursive drain invariant.
• The issue's original voice-event report is already covered by the retained-event drain path, as noted in the issue discussion; this PR addresses the distinct interrupt_message-only fallback still present on current main.

Automated hermes-sweeper review.