feat(tts): orchestrator with per-conversation ElevenLabs fallback

[mikelrh-dev]

Summary

• PR 2/3 of the tts-elevenlabs-fallback change: TTSService is now a provider orchestrator with per-conversation fallback from ElevenLabs to Microsoft Edge TTS.
• Added TTS provider configuration (TTS_PRIMARY_PROVIDER, ELEVENLABS_API_KEY/VOICE_ID, TTS_ELEVENLABS_TIMEOUT) with safe defaults (microsoft, empty creds, 15s timeout).
• Strict TDD: 13 new tests (2 config + 11 orchestrator) written RED before implementation; full suite 104 passed.

Changes

File	Change
backend/config.py	Added TTS_PRIMARY_PROVIDER, ELEVENLABS_API_KEY, ELEVENLABS_VOICE_ID, TTS_ELEVENLABS_TIMEOUT env vars with defaults
.env.example	Documented the 4 new TTS vars
backend/services/tts.py	Refactored to orchestrator: TTSClient Protocol, _conversation_providers pinning dict, asyncio.wait_for EL timeout, fallback, forget_conversation, tuple returns (Path, provider)
tests/test_config.py	+test_tts_provider_defaults, +test_tts_provider_env_override
tests/test_tts.py	+TestTTSServiceOrchestrator (11 tests): pinning, isolation, timeout fallback, both-fail RuntimeError, missing creds, microsoft-only, no-pin, tuple returns, forget idempotent

Test Plan

• python -m pytest tests/ -v → 104 passed, 3 warnings (baseline 91)
• Strict TDD cycle: RED written first (15 failing), GREEN after implementation

Notes

• Stacked PR: targets feature/tts-elevenlabs-clients (PR 1) since PR 1 is still open; retargets to main automatically on merge.
• API wiring (conversation_id threading, JSON/SSE provider flags) lands in PR 3.