v0.1.0 — Initial research PoC release

Initial public research PoC release.

Highlights:

• Metadata-aware low-bit KV-cache compression benchmark.
• Main point: 5.11x modeled KV payload compression vs FP16.
• In a local GPT-2 CPU/Python fake-quant benchmark: +11.0% attention-output cosine vs local RotorQuant-3b baseline, at +4.35% modeled memory.
• Includes conservative scope notes: not a production runtime, CUDA, real-VRAM, or throughput claim.

Feedback wanted on:

• effective-bit accounting,
• metadata overhead modeling,
• benchmark scope and baselines,
• next experiments worth running.