docs: eight bits a component, on the roadmap

[tanvincible]

Puts eight-bit vectors on the roadmap. The lab's follow-up run validated them with the plan, predictions and decision rule committed before any results.

• Scheme: LVQ-8. Each vector is stored relative to a center, the mean of the first vectors added, in its own 256-step range.
• Result: within 0.01 recall@10 of half precision on all seven text configurations, at ef 64 and 128, with nothing re-ranked. That holds with the center from the first 1,000 vectors (worst 0.005) and from 1,000 on one topic (worst 0.006).
• Checked by skeptics:
  • Recomputed independently, the verdict is the same.
  • No pairing of builds pushes a loss past 0.0063.
  • 40 topic seeds on openai3-384 and on bge-384: worst 0.0082 with every vector ranked, the bge worst.
  • One vector is too few to center on (within 0.002 of the bar); ten were enough.
• Not measured: the kernel and its speed, the file format, and the vectors added before there is a center.

The plan, the results and the skeptic pass are in bench/lab/quant-plan.md on lab/x86, from runs 38090885820 and 38131118757.