nx_blockfloat_stack_gate.nx

nx_blockfloat_stack_gate.nx -- BLOCK-FLOAT weights through a STACKED (multi-block) transformer -- the step toward a full model. Chains the existing gradcheck-verified block forward (blk_fwd) N_LAYERS times (each block's output feeds the next), with the weights block-float quantized. Proves block-float COMPOSES through depth, and the key depth property: per-tensor INT8 degradation COMPOUNDS with depth, while block-float stays stable. 1 the stacked transformer RUNS with block-float weights (non-trivial output) 2 EXCEED: block-float stack error vs full-precision < per-tensor over N_LAYERS blocks 3 DETERMINISTIC: the block-float stack run twice == bit-identical 4 DEPTH: per-tensor error at depth N > per-tensor error at depth 1 (degradation compounds; block-float stable) expect_exit: 0 license_tier: ORIGINAL