feat(gateway): streaming TTS adapter contract and consumer

[ctaylor86]

What does this PR do?

Adds gateway-level streaming TTS so voice platform adapters can consume the LLM token stream and synthesise audio clause-by-clause, instead of waiting for the full response before starting TTS. This closes the consumer gap referenced in #47896 and implements the feature requested in #60671.

Why this approach: The existing TTS path buffers the entire assistant response, then calls the TTS provider once. For voice interactions this adds latency proportional to response length. This PR introduces an opt-in streaming contract on BasePlatformAdapter that adapters can implement to receive PCM chunks incrementally as the LLM produces text. Non-streaming adapters are unaffected - the fallback to whole-file TTS is preserved.

Related Issue

Fixes #60671

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

• gateway/streaming_tts_consumer.py (new, 423 lines): StreamingTTSConsumer class that sits between the LLM token stream and the adapter. Iterates provider output, detects clause boundaries, and forwards PCM chunks to the adapter as they arrive. Handles timeouts, abort, fallback to whole-file TTS, and queue backpressure.
• gateway/platforms/base.py (+175 lines): Adds the supports_streaming_tts flag and the begin_streaming_tts / write_streaming_tts_chunk / finish_streaming_tts / abort_streaming_tts / get_streaming_tts_audio_format methods to BasePlatformAdapter. All default to no-ops/false so existing adapters are unaffected.
• gateway/run.py (+198/-52 lines): Wires the consumer into GatewayRunner._run_agent. When the adapter supports streaming TTS, the consumer is created on the gateway event loop and fed tokens from the LLM stream. Per-turn duplicate suppression prevents whole-file fallback from replaying after streaming has begun. Think blocks are suppressed (not synthesised).
• gateway/config.py (-2 lines): Minor cleanup.

How to Test

1. pytest tests/gateway/test_streaming_tts_consumer.py -v - 27 tests covering adapter contract defaults, consumer lifecycle, timeout/abort/fallback semantics, duplicate suppression, concurrent turn isolation, queue backpressure, and the finish-sentinel race.
2. pytest tests/gateway/test_streaming_tts_gateway_regression.py -v - 2 regression tests exercising the real _run_agent finalisation path (no NameError in voice or text turns).
3. Full suite: pytest tests/gateway/ -q

All 29 tests pass (verified on macOS 15, Python 3.11).

Checklist

Code

• I have read the Contributing Guide
• My commit messages follow Conventional Commits (feat(gateway): ...)
• I searched for existing PRs to make sure this is not a duplicate
• My PR contains only changes related to this feature
• I have run pytest tests/gateway/ -q and all tests pass
• I have added tests for my changes
• I have tested on my platform: macOS 15.2 (Tahoe), Python 3.11

Documentation & Housekeeping

• N/A - no new config keys, no architecture/workflow changes, no tool behaviour changes
• Cross-platform: uses only stdlib asyncio primitives (no platform-specific code)

Design Notes

• Opt-in by default: supports_streaming_tts defaults to False. No existing adapter changes required.
• Prompt-cache safe: No changes to system prompt, toolset, or message history mid-conversation.
• Fallback preserved: If streaming fails before any audio is audible, whole-file TTS runs as before. After audio starts, fallback is suppressed to prevent replay from the beginning.
• Per-turn isolation: Duplicate suppression markers are scoped per-turn and cleared before recursive turns.
• Event-loop safe: Provider iteration runs off the gateway event loop. PCM chunks are written immediately (not buffered per clause).

[teknium1]

Merged via PR #73862 (4de9e53) — cherry-picked clean with your authorship preserved in git history, and live-tested by Teknium working flawlessly. Your StreamingTTSConsumer (clause-boundary detection, backpressure, duplicate suppression, whole-file fallback) landed intact as the gateway half of the streaming subsystem, composed with the StreamingTTSProvider registry from #47588 for synthesis. Nice work writing it against the registry from the start — the two halves composed with zero seam changes. The plugin-stream()/opus half (#47896) is noted as the follow-up you scoped out. Thanks for the contribution!