Scrub remaining privacy markers from README + language.py docstring

Untrack CLAUDE.md (privacy: contains internal infra refs)

PyramidKV+SSD cache coexistence: 4 fixes restore cross-request reuse

PyramidKV: per-layer arithmetic-decay budget + trim() for MTP rollback

PyramidKV: post-prefill KV cache compression for long context

P4: MTP working — 75% accept, 15% net speedup (52.6 vs 45.6 tok/s)

README: document MTP experimental status + bench results

P3.7: Norm +1 shift on extract for Gemma-stored MTP weights

P3.5: Wire MTP forward into LanguageModel + outer Model

P3: enable MTP shard loading (config + sanitize + shard rewrite script)

P2: Add MTP head classes + sanitize routing (MLX port of vLLM Step3p5MTP)

Add root config.json as HF Hub download-tracking anchor

README: add HF metadata frontmatter + fail-fast on token-count mismatch

Initial: Step 3.7 Flash MLX adapter (text + vision)