Volume-based weights, undot-before-restore, startup prefetch, UI polish

chore(frontend): upgrade to vite 8 / vitest 4 / plugin-react 6 (0 vulnerabilities)

fix(frontend): model-search highlight, compare-all model guard, drop dead i18n key; fix README CPU note

fix(frontend): fetch methods once on mount; show model name in picker

docs: sync plan to LlmPanel import deviation

fix(frontend): build uses tsc --noEmit; sync plan to deviations

feat(backend): weight-download script for Docker build

chore(backend): scaffold uv project (Python 3.10, CPU torch via tnqeet + PyTorch CPU index)

Add implementation plan for tnqeet demo app

Pin Python 3.10 (KenLM fails to build on 3.11+)

Pin Python 3.11; use CPU-only torch wheels

Add health endpoint, input cap, Python pin, and Railway ops notes

Expand honeybee Wikipedia example to ~68 words

Replace Damascus example with neutral honeybee topic

Add Wikipedia-style and two public-domain poetry examples

Revise seed examples: drop two short, add long (15-50 word) examples

Bake largest model variants; link package note to PyPI

Add tnqeet demo app design spec