Mamba 1, 2 Transplant Weights Into 3

This is an ambitious, very niche open-source experiment: clever if it works, but the writeup reads more like a proof-of-concept than a validated migration recipe.

• –The core idea is checkpoint surgery, not a new model, so the value is in compatibility engineering and memory discipline
• –The phase-based freeze/unfreeze plan is the practical hook; that is what makes the 12GB claim plausible
• –The mathematical claims around gate inversion, pooling, and inverse-softplus reparameterization are specific enough to be interesting, but they still need benchmarks and reproducibility data
• –For local-model users, this is relevant as a “reuse what you have” path; for everyone else, it is probably too bespoke to generalize