Dev fixes silent weight drift in Qwen3.5 35B

This proves that seemingly broken open-weight models often just need targeted weight surgery rather than complete retraining.

• –The bug exposes a dangerous edge case with AdamW optimizers in hybrid architectures, where rare experts suffer from massive weight drift.
• –Modifying just two tensors in a 64GB model achieved an 88.6% error reduction, highlighting the fragility of late-layer weights.
• –Developers training MoE and recurrent hybrids should manually inspect final block scales to prevent similar silent degradation.