Slow first run — 90MB download + minutes of CPU embedding

[krishpinto]

Problem

First run downloads ~90MB and embeds the whole history on CPU (slow). One-time, but it's the first impression.

Acceptance criteria

• Progress UX for download + backfill.
• Consider quantized/smaller model or better batching.
• Document expected first-run time.

[krishpinto]

Partially resolved in v0.1.0: the 90 MB model download is gone — the model is now bundled into the release zip, so first run is fully offline with no download. The remaining half (CPU embedding time on a large initial import) is still real; keeping this open to track that specifically.