Progress reporting on long-running scans and index builds #957

Problem

A cold scan over a large corpus gives the user nothing to look at. Core has no progress reporting; the only bar anywhere is the Hugging Face model download bar in the embeddings plugin (embedding/progress.py), TTY-gated.

This becomes acute with #956: a first vector-index build runs embed() once per ingested row via a worker round trip (41k rows on the reference corpus — minutes of wall time, silent). Trigger-fired scalar calls cannot batch, so the cost is irreducible; the answer is visibility, not policy — show the user what they are paying rather than deciding for them whether it is acceptable. The embeddings disk cache makes re-runs cheap; this is about first builds and full rebuilds.

Proposed

Progress on the scan path in core, and on the embedding path in the plugin. TTY-gated, consistent with the existing download bar. No new dependencies in core.