Full fine-tune (94% of params, vision encoder frozen) of Gemma 4 E4B-IT for Malaysian TTS-companion. CPT-free. Broke the warmth+grounding ceiling that LoRA couldn't: warmth 0.5 + A_fab 1 + 0 dead-ends (z.ai GLM-5.2 judge). Trained on v23 + ilmu-generated caring boosters + fragment-backchannel. Standalone — no adapter stacking. Serve with base_model = this repo.