Improve embedding visualization color spread using UMAP

[monneyboi]

Problem

Article visualization colors cluster into only 2-3 color variants (purple/red and green/orange) instead of spreading across the full color spectrum.

Current Implementation

The compute_content_hue() function in embeddings/embeddings.py derives hue by:

1. Computing the mean embedding vector for the article
2. Hashing it with SHA-256
3. Using hash_int % 360 to get a hue

This produces deterministic but poorly distributed colors when articles are semantically similar.

Proposed Solution

Use UMAP to reduce the mean embedding to 1D, then map that value to hue (0-360). This approach:

• Spreads articles along their primary semantic axis
• Similar articles get nearby (but distinct) colors
• Very different articles get contrasting colors
• Scales well as more articles are added

Implementation

Modify compute_content_hue() to:

1. Use UMAP with n_components=1 on the mean embedding
2. Normalize the resulting value to 0-360 range
3. Return as hue

This may require processing all articles together for optimal spread, or using a pre-fitted UMAP model.