bug(embedder): search --json leaks model loading messages to stdout

[carlos-alm]

Found during dogfooding v2.4.0

Severity: Low
Command: codegraph search "query" --json

Reproduction

codegraph search "build graph" --json 2>/dev/null | head -3

Expected behavior

Clean JSON on stdout with all status messages on stderr, allowing pipe usage:

{
  "results": [

Actual behavior

Loading embedding model: Xenova/all-MiniLM-L6-v2 (384d)...
Model loaded.
{

The "Loading embedding model..." and "Model loaded." lines appear on stdout, breaking JSON pipe consumers.

Root cause

These messages come from @huggingface/transformers via console.log. The codegraph logger routes through stderr, but the HF library's own console output goes to stdout.

Suggested fix

Redirect console.log to stderr during model loading, or suppress HF library output when --json is active.