docs: upstream-first KV-cache guidance (q8_0/q4_0, mainline Hadamard rotation); fork demoted to experimental

docs: remove AWQ variants from family table (AWQ lane paused, repos private)

fix: family model links majentik/gemma4-* -> majentik/gemma-4-* (reported by @shramee in gemma-4-31B-TurboQuant-MLX-4bit#1)

docs: Tier 2 polish — variant matrix + quant trade-off

chore(card): enrich YAML frontmatter (pipeline_tag, language, library_name, inference)

Add MLX quantized model with KV cache compression

initial commit