Logic pass: CPU attention mirrors get the same int32 bound as the GPU kernels (peers refuse identically); gpuMlpChain returns pooled buffers when the row-max guard throws
Four local commits in one upload (ec432c8, c8db1cf, 47724cd, ae9cfa2): bounded elementwise units and LUT build (407 MB -> 2 MB, 277 MB -> 1 MB at startup); requant16 retrained from floor (-0.4981 LSB bias) to round-half-up, N/N 65,536/65,536, browser requant LUT re-exported; qat.py and every browser quantizer now refuse NaN/Inf instead of silently quantizing the whole layer to 0 (new tests fail on the old code); README September section. python test_verified_units.py 25/25, npm test 12/12 suites green.
