Fix engram n-gram lookups during cached decode: add engram_context_ids so single-token steps match full-sequence computation

Upload README.md with huggingface_hub

Upload Escarda-LLMs.png with huggingface_hub

Update to v3: adds employment-law coverage (paycheck/wages/fired/non-compete now route to Legal Advisor); no regression on other roles

Escarda-Rewrite 86M prompt optimizer (v2 final) + loader + card

initial commit