[reg] lade gemma3-1b-it (patch_sha=b042ef3127950a71) [transformers] `torch_dtype` is deprecated! Use `dtype` instead! Loading weights: 0%| | 0/340 [00:00 b042ef3127950a71), aber Verhalten identisch — Refactor ist verhaltenstreu. ✓ [AutoCalibrator] Loaded manifold for google/gemma-3-1b-it (T=0.60) [gemma3-px] Patched 26 attention modules with lossless mem-efficient attention. [gemma3-px] ACTIVE_MANIFOLD_LEAN (kausaler Kern). SR-59 for L26 (HS=1152). === REGRESSION: 18/18 Zellen OK === ALLE STRENGEN REGRESSIONEN BESTANDEN (Text/loops/path/zone exakt, phi/ent innerhalb Toleranz).