dense_h216: add seed replicates, correct the published number to a 3-seed mean

card: repeat-seed confirmation of the dense-vs-Monarch gaps

card: param-matched dense controls and the controlled latency comparison

add the param-matched dense controls (dense_h52 .. dense_h216)

add monarch_40; refresh the card with the full block-count sweep and the int8 scoping

add the block-diagonal block-count sweep: blockdiag_5 / 10 / 20 / 40

add the Monarch block-count sweep: monarch_5 / monarch_10 / monarch_20

rename monarch_*->blockdiag_* (they were single block-diagonal factors), publish genuine 2-factor Monarch runs, re-quantize int8 with Einsum weight fix

Update README.md

200-epoch retrain: 9 variants with FP32 + int8 ONNX + int8 PESQ findings

Initial upload: 9 best-PESQ checkpoints from compression sweep

initial commit