BDH fast-weight write-back lands open source

This is less a flashy launch than proof that BDH's inference-time plasticity can actually persist, and the selective consolidation result is the part worth watching. The catch is that it is still a synthetic-memory demo, so the natural-language question remains open.

• –The repo closes a real gap in the public BDH code: the Hebbian co-activation was computed before, but it was not written back into weights.
• –Sparse activation codes act as stable addresses, so the same token maps to the same fast-weight slot regardless of position.
• –The best confirmation run reaches 99.0% / 98.0% / 97.5% on n2 / n4 / n8, while the no-write-back baseline sits at 1.0% on n8.
• –Dense consolidation drops to 75.4% / 68.1% / 89.8%, but `rowtop10` preserves most of the control signal at 97.5% / 97.1% / 96.2%.
• –Independent H100 verification and counter-benchmarks in the 91-95% range make the result look reproducible rather than seed-chasing.
• –It is still a 25M-parameter benchmark model, so the real proof point is whether the same trick survives natural-language data.