voxora-registry

Central model registry for voxora (ASR-specific).

Owns the answer to "given a model id, give me the engine descriptor

• the exact on-disk file". Before this crate, that answer was distributed across voxora-hf (cache key), voxora-whisper (lex-sort over *.bin), and voxora-cli (heuristic). Now it lives here.

Structural fix for voxora#79 (ggerganov/whisper.cpp/ggml-large-v3.bin incorrectly resolving to ggml-base.bin): ModelDir::entry, populated by voxora-hf 0.1.2 and surfaced through Registry::resolve, names the exact file. Consumers no longer need to lex-sort.

Phase 0 surface

• [ModelId::parse] — strict parser for org/repo, org/repo/file, /local/path, ./local-path.
• [Registry::resolve] → (EngineDescriptor, ModelDir).
• [builtin_whisper_descriptor] / [builtin_qwen3asr_descriptor] default descriptors.
• [CacheManifest] — .voxora-manifest.json sidecar.
• [hf::hf_registry] convenience builder (behind the hf feature).

ASR-specific: descriptors model only ASR engines (Whisper, Qwen3-Asr). Adding parakeet, voxtral, or granite-speech is a 0.2.x change.

Hardware backends

Model resolution is engine-agnostic; the hardware backend the resolved model runs on is selected downstream by the engine adapter. The cross-engine hardware / GPU support matrix (CUDA / Metal / Vulkan / CPU per engine, compute capability requirements, voxora-bridge forwarding rules) is documented in docs/GPU_SUPPORT.md.

Examples

• registry_resolve — build a [Registry] with both built-in descriptors and resolve a model id, printing the descriptor and on-disk [voxora_traits::ModelDir] the resolver chose. Run with cargo run --example registry_resolve -p voxora-registry -- <hf-model-id>.