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

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