feat(vlm): Vision/VLM screen + full modality model catalog (0.20.9 parity)

[sanchitmonga22]

Summary

Brings the React Native starter to full SDK-feature + model-catalog parity with the monorepo reference example (examples/react-native/RunAnywhereAI), using only published @runanywhere/*@0.20.9 public APIs.

The headline change is the new Vision / VLM screen (the one capability the starter was missing) plus a curated model for every modality the starter exposes. This PR is stacked on the 0.20.9 integration baseline, so it also carries those preceding commits.

What's new (this work)

👁 Vision / VLM screen (src/screens/VisionScreen.tsx)

• On-device image understanding via RunAnywhere.processImageStream — streaming, token-by-token descriptions.
• Image input adapted to the starter's dependency set (no new native modules): tap a sample image or paste a custom image URL / local path. Remote sources are fetched to the app cache with react-native-fs and the on-disk path is fed to the VLM (VLM_IMAGE_FORMAT_FILE_PATH), mirroring the reference example's file-path flow.
• Editable prompt, live preview, and cancel mid-generation.
• Hermes-safe: manual iterator.next() loop over the NitroModules async iterable (no for await...of).

src/services/VLMService.ts

Thin wrapper over RunAnywhere.processImageStream / cancelVLMGeneration / currentModel(MULTIMODAL), mirroring the reference example's VLMService.

Model catalog — one curated model per modality (src/services/ModelService.tsx)

Registered with structured proto enums (INFERENCE_FRAMEWORK_*, MODEL_CATEGORY_*, MODEL_ARTIFACT_TYPE_*) — ids/urls/framework/category/artifactType copied from the reference ModelCatalogBootstrap:

Modality	Model	Framework / Category
LLM	lfm2-350m-q8_0 (+ smollm2-360m-q8_0 alt)	LLAMA_CPP / LANGUAGE
VLM	smolvlm-500m-instruct-q8_0	LLAMA_CPP / MULTIMODAL (tar.gz)
STT	sherpa-onnx-whisper-tiny.en	SHERPA / SPEECH_RECOGNITION
TTS	vits-piper-en_US-lessac-medium	SHERPA / SPEECH_SYNTHESIS
VAD	silero-vad	ONNX / VOICE_ACTIVITY_DETECTION

Also added the VLM (MULTIMODAL) download/load lifecycle to ModelService and its unload path.

Wiring

Vision added to the navigator (App.tsx), RootStackParamList, HomeScreen grid + model-info panel, screens/index.ts, and the FeatureCard / ModelLoaderWidget icon maps. README refreshed with the Vision feature and the full model table.

Not included (deferred, per scope)

• Diffusion / image generation — intentionally out of scope.

Verification

• npx tsc --noEmit — clean (0 errors).
• eslint on new/changed source — no new errors.
• No native run / pod install attempted (tsc is the gate).

🤖 Generated with Claude Code