Showcase MagicQuant with real results, reusable recipes, and a stronger README

[magiccodingman]

What changes

The README now explains what makes MagicQuant useful: learning real tensor recipes, discovering and measuring mixed-precision hybrids, sharing reproducible manifests, cloning configurations onto compatible variants, and scheduling storage/GPU work. It leads readers into real published models and a practical NuGet/CLI starting point.

• Showcase the Qwen3.6 35B MoE survivor set, its ten-recipe uncensored clone, and the specialized Qwen3.8 MXFP4 adaptation with measured sizes and KLD.
• Add a research showcase with complete tables and eight pinned Hugging Face evidence links. Distinguish source discovery from clone measurements, and specialized MXFP4 work from ordinary CLI capabilities.
• Restore the original 4B worked example's useful tradeoff in the main narrative. Preserve the original research overview unchanged, including its tables, tensor recipes, and diagrams.
• Remove the README's development/history section; contributor and migration guidance remains in its existing documentation. Retain funding, upstream attribution, scratch-disk advice, platform status, and license information.

Validation

• Eight Python documentation/release tests passed.
• Ten configuration contract tests passed, including the README YAML example.
• Release package built and clean-installed successfully; package smoke verified the exact root README, icon, native assets, CLI help, config generation, and path preflight.
• All eight pinned source links returned HTTP 200; numerical tables checked against the published manifests.
• git diff --check passed.

Documentation only; no new model benchmarks or runtime changes.