Client-side ONNX inference (zero backend)

[ONDAAA]

Run embedding model directly in the browser via ONNX Runtime Web / WebAssembly.

Benefits:

• Zero server costs for inference
• Zero latency (no network roundtrip)
• Works offline
• No rate limiting concerns

Approach:

• Export MiniLM to ONNX format (~30MB)
• Bundle with extension or lazy-download on first use
• Use onnxruntime-web for inference
• Keep backend only for auth/billing

Trade-off: larger extension size (~30MB vs ~50KB)