Return voices for single-model servers when model param is omitted

[kevin-ho]

Problem

When the server has a single model configured, GET /v1/audio/voices without a ?model= query parameter returns {"voices":[]}. Clients that call this endpoint to populate a voice picker — without prior knowledge of model IDs — receive an empty list and cannot discover available voices.

Fix

When no ?model= param is provided and exactly one model is configured, fall back to that model's voices. When multiple models are configured, the existing behavior is unchanged (empty response, since voices cannot be attributed to a specific model without the param).

size_t model_idx = SIZE_MAX;
if (!model_id.empty()) {
    const auto it = model_index_.find(model_id);
    if (it != model_index_.end()) {
        model_idx = it->second;
    }
} else if (models_.size() == 1) {
    model_idx = 0;
}

Rationale

The change is purely additive — it only affects the models_.size() == 1 path that previously returned empty. The multi-model path and the explicit ?model= path are unchanged.

Voice presets are static config metadata parsed at startup, so this works regardless of lazy_load setting. No runtime state or model session is involved — handle_voices only reads the in-memory voice_presets map and the embeddings/ directory listing.

Tested

• GET /v1/audio/voices (no param) → voices returned (was [])
• GET /v1/audio/voices?model=<id> → unchanged
• Backend: CUDA

[0xShug0]

@kevin-ho Thank you! PR merged.