Adds `openrouter` to the STT provider list, alongside local/groq/openai/
mistral/xai. OpenRouter exposes /v1/audio/transcriptions with a
JSON+base64 protocol (not OpenAI-multipart), so it needs its own client
path rather than re-using the openai-compatible code.

Why this is useful
- One key (OPENROUTER_API_KEY) covers chat + STT, which simplifies
  configuration for users already on OpenRouter for LLMs.
- OpenRouter routes requests across providers (Groq Whisper, OpenAI
  Whisper, etc.), giving fallback for free.

Provider selection
- Explicit: stt.provider: openrouter (or STT_PROVIDER=openrouter).
- Auto-detect: chosen when local + groq are unavailable but
  OPENROUTER_API_KEY is set; ranks below local/groq, above openai.

Implementation notes
- _transcribe_openrouter() in tools/transcription_tools.py:
  - JSON body with base64-encoded input_audio.data + format string.
  - Format whitelist (wav/mp3/flac/m4a/ogg/webm/aac); rejects unsupported
    extensions early, with .oga/.opus → ogg and .mpeg → mp3 aliases.
  - 25 MiB size cap matching the documented limit.
  - Typed exception branches for Timeout / ConnectionError so transient
    failures surface cleanly instead of as tracebacks.
- voice_mode.py: voice_doctor now reports the openrouter/mistral/xai
  branches (previously only listed local/groq/openai).
- cli-config.yaml.example: documents the new provider + model + optional
  base_url override.

Tests
- 19 new tests in TestTranscribeOpenRouter / TestGetProviderOpenRouter /
  TestTranscribeAudioOpenRouterDispatch covering: missing key, success,
  whitespace/empty handling, HTTP errors, permission errors, format
  whitelist + aliases, JSON-not-multipart wire format, base_url override,
  header shape, connection errors, and provider auto-detect ranking.
- pytest tests/tools/test_transcription_tools.py: 106 passed, 7 skipped.

Adds `openrouter` to the TTS provider list, alongside edge/elevenlabs/
openai/minimax/mistral/gemini/xai/neutts/kittentts/piper. Mirrors the
STT provider added in the prior commit on this branch — same
OPENROUTER_API_KEY covers chat + STT + TTS.

OpenRouter exposes /v1/audio/speech with the OpenAI request shape
({model, voice, input, response_format, speed}), so this provider
reuses the OpenAI client with a swapped base_url + key. The model slug
selects the underlying provider (openai/google/mistral/...).

Implementation
- _generate_openrouter_tts() in tools/tts_tool.py: OpenAI-client based,
  default model openai/gpt-4o-mini-tts-2025-12-15 (the slug OR returns
  for the OpenAI route as of the May 2026 audio API launch).
- response_format is hard-coded to "mp3" — OR's endpoint only accepts
  mp3 / pcm. Telegram-bound .ogg outputs go through the existing ffmpeg
  conversion path (same path Edge TTS takes).
- openrouter is in BUILTIN_TTS_PROVIDERS so a user's
  tts.providers.openrouter command block can never shadow it.
- PROVIDER_MAX_TEXT_LENGTH["openrouter"] = 4096 (follows the underlying
  OpenAI cap).
- Configurable base_url via tts.openrouter.base_url or
  TTS_OPENROUTER_BASE_URL env (matches the xAI / OpenAI patterns).

Tests
- 13 new tests in test_tts_openrouter.py covering missing key, success,
  default + custom model/voice, default + config + env base_url, the
  always-mp3 invariant (via .ogg output_path), speed clamp + omitted
  default, dispatcher routing, BUILTIN_TTS_PROVIDERS membership, and
  PROVIDER_MAX_TEXT_LENGTH.
- pytest tests/tools/test_tts_openrouter.py: 13 passed.
- Full TTS + STT suite: 265 passed, 7 skipped (one collection error in
  test_tts_kittentts is a local-only missing-numpy issue, unrelated).

Live verification
- Direct curl against POST https://openrouter.ai/api/v1/audio/speech
  with model openai/gpt-4o-mini-tts-2025-12-15 — returns audio/mpeg.
- _generate_openrouter_tts() against the live endpoint — produces a
  valid 75 KiB MP3.
- Full dispatcher via text_to_speech_tool with HERMES_SESSION_PLATFORM=
  telegram — produces a valid Opus OGG (24kHz mono) via ffmpeg.

Docs
- website/docs/user-guide/features/tts.md: added the provider row,
  config block, and per-request input cap entry.