feat(speech): integrate TTS/ASR cloud vendors into speech registry (task 21)

[Nixxx19]

Summary

Wire OpenAI, ElevenLabs, and Deepgram cloud speech adapters into the MoFA framework so that callers can configure and register them through a single, config-driven API surface rather than constructing each adapter manually.

Related Issues

Closes #1256 -- Integrate TTS/ASR from 2-3 Cloud Vendors

Context

The vendor adapters (OpenAI TTS/ASR, ElevenLabs TTS, Deepgram ASR) already existed in mofa-integrations and already implemented the TtsAdapter / AsrAdapter kernel traits. The gap was:

1. No SDK re-exports -- callers had to depend directly on mofa-integrations
2. No config-driven factory -- callers had to construct every adapter manually
3. No single function to populate a SpeechAdapterRegistry from a config block

Changes

crates/mofa-sdk/Cargo.toml

• Added mofa-integrations as an optional dependency
• Added three forwarded feature flags: openai-speech, elevenlabs, deepgram

crates/mofa-sdk/src/lib.rs -- new pub mod speech

• Re-exports kernel speech traits (TtsAdapter, AsrAdapter, AudioFormat, TtsConfig, AsrConfig, TtsOutput, AsrResult, VoiceInfo)
• Re-exports foundation types (SpeechAdapterRegistry, VoicePipeline, VoicePipelineConfig, VoicePipelineResult)
• Re-exports config types (SpeechConfig, SpeechProviderConfig) under all three feature flags
• Per-feature adapter re-exports (OpenAiTtsAdapter, OpenAiAsrAdapter, ElevenLabsTtsAdapter, DeepgramAsrAdapter)
• register_speech_adapters(registry, config) -- feature-gated factory that constructs and registers the correct adapter per provider entry

crates/mofa-integrations/src/speech/registry_builder.rs (new file)

• SpeechProviderConfig -- per-vendor config: provider, api_key, default_tts, default_asr
• SpeechConfig -- top-level config holding a list of provider entries
• Builder methods: new(), with_provider(), as_default_tts(), as_default_asr()
• Full serde support with #[serde(default)] on booleans for forward-compatible deserialization
• Internal unit tests (4 tests)

crates/mofa-integrations/src/speech/mod.rs

• Unconditionally exports the new registry_builder module

crates/mofa-integrations/tests/speech_example_tests.rs

• Added mod registry_builder_tests (4 tests): empty config, builder chain, JSON roundtrip, missing default flags

Testing

All tests pass without network access:

# Build with all speech features
cargo build -p mofa-sdk --features openai-speech,elevenlabs,deepgram

# Run registry builder tests (no network required)
cargo test -p mofa-integrations \
    --features openai-speech,elevenlabs,deepgram \
    -- speech

# Results: 12 passed, 3 ignored (live-API tests skipped in CI)

Live-API tests are marked #[ignore] and require vendor API keys via environment variables:

OPENAI_API_KEY=sk-... cargo test -p mofa-integrations \
    --features openai-speech -- --ignored

Breaking Changes

None. All new API surface is additive. Existing code that does not enable the new feature flags is unaffected.

Checklist

• All new public enums are #[non_exhaustive]
• Feature flags are correctly gated with optional = true
• No circular dependencies introduced (factory in sdk, config types in integrations)
• Serde roundtrip tests added
• Live-API tests remain #[ignore]d for CI