fix(core): robust embedding model download (#740)

[ajianaz]

Summary

Fixes #740 — embedding model download fails on first run.

Problem

had several issues:

1. No timeout — with infinite timeout. Slow HF connections → hang
2. Full RAM buffer — loaded entire 187MB model_q4.onnx_data into memory at once
3. No retry — any transient network error → permanent failure
4. No progress — user sees nothing, assumes crash

Changes

• Timeout: 30s connect, 300s read (generous for 187MB over slow connections)
• Retry: Up to 3 attempts with cleanup of partial downloads between retries
• Streaming: 64KB chunked write to disk — avoids buffering 187MB in RAM
• Progress: Shows file name, total size, percentage milestones (every 10%)
• Integrity: Verifies downloaded bytes match Content-Length when available
• Cleanup: Removes partial .tmp files on retry/failure

Testing

• cargo fmt ✅
• cora review ✅
• Full clippy/test will run in CI (Woodpecker)

Closes #740

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-cli · BYOK · MIT