fix: honour default_voice_code in _split_multi_voice

[Christian-Sidak]

Summary

• _split_multi_voice() in api/src/services/tts_service.py computed segment_lang from the first character of the voice name when no explicit lang_code was provided, skipping settings.default_voice_code entirely.
• Voices whose filename does not start with a recognised Kokoro language code (e.g. ursa, eve, iris) were therefore routed to the wrong phonemiser pipeline, returning HTTP 200 with an empty body.
• The fix mirrors the precedence already used in kokoro_v1.py: lang_code > settings.default_voice_code > voice[:1].

Fixes #514

Changes

api/src/services/tts_service.py (line ~295)

Before:

segment_lang = lang_code if lang_code else segment_voice[:1].lower()

After:

segment_lang = (
    lang_code
    or settings.default_voice_code
    or segment_voice[:1].lower()
)

api/tests/test_tts_service.py

Two new unit tests added alongside the existing _split_multi_voice tests:

• test_split_multi_voice_default_voice_code_used_when_no_lang_code -- verifies that a voice named ursa (no language prefix) uses settings.default_voice_code when no lang_code is passed.
• test_split_multi_voice_explicit_lang_code_beats_default -- verifies that an explicit lang_code still wins over settings.default_voice_code.

Test plan

• Reviewed that settings is already imported in tts_service.py (line 15)
• Verified the fix matches the precedence in kokoro_v1.py
• Added unit tests following the existing mock/AsyncMock pattern in test_tts_service.py
• Confirmed no other callers of _split_multi_voice are affected (the precedence is strictly additive)

[remsky]

Good catch, and thanks. I was able to repro, and looks like generate_from_phonemes has the same omission on the /dev path, I'll follow up with a commit on it