AI Test Bug Fixed, Making Models More Reliable

On about six percent of random seeds, Muse Spark made test-llama-context fail on state_load, cp_h and cp_d, with an NMSE of 1.0e-05 to 1.4e-05 against a 1.0e-05 threshold. The sliding window is what exposed it: a restored state keeps the window alone, so an SWA layer's attention reduces over a different cell layout and the continuation isn't bit exact. Every SWA model carries that at 1e-08 — Muse Spark carried it at 1e-05.

The generator is what turned 1e-08 into 1e-05. A norm-preserving residual gate holds alpha^2 + beta^2 == 1 and the conversion resolves both weights from one raw weight, but the generator drew every tensor on its own — so the stream and branch scaled independently and the model became an amplifier. The weights the test ran on were not a Muse Spark.

The fix: give each per-tensor rule a filler, pick one from a table keyed on a substring of the tensor name, and keep a single write path behind it — replacing flags threaded through the write path. Mamba moves into the table and Muse Spark joins it, with a gate filler that draws the raw weight once per block and resolves alpha and beta the way the conversion does.

Over a thousand seeds on a CPU build, the worst case dropped from 1.36e-05 to 2.48e-06 and the median from 4.10e-06 to 5.48e-07, with no seed above the threshold. The other 131 models come out identical to the bit, since no other architecture carries these tensors, and ctest is green on CPU and CUDA builds. The fillers also stopped repeating the generator and distribution: the normal draw now takes its key and the other fillers transform its values, so the generator lives in one place and the generated