card: GSM8K 100q across all three runtimes (98/97/95) + the two-pass budget note

card: record the stop-token defect and its fix

fix: declare <|eot|> as a stop token — the bundle never terminated

card: correct the machine — Mac Studio M4 Max, not MacBook Pro

card: add the raw-MLX arm — Core AI ties MLX, both beat the ExecuTorch build by 14%

Card: lossless n-gram spec-decode on this bundle (1.34x chat / 1.96x code / 1.84x tool calling), no drafter, no re-export

card: same-machine head-to-head vs the ExecuTorch Metal build

Muse-Glimmer-30B text decoder -> Core AI: decode int4hu (--head-sym), token-exact vs fp16 oracle, 26.69 tok/s on M4 Max

initial commit