Gemma 4, Qwen 3.5 fail surgical edits

Vibe-coding hits a consistency wall as small models solve logic puzzles but ignore git-diff hygiene.

• –Quantization noise at Q4/Q5 levels likely degrades attention weights enough to trigger "hallucinated" changes to inactive code regions like comments.
• –Surgical editing requires a model to value token identity over logic generation, a skill that scales with parameter density rather than raw reasoning benchmarks.
• –TurboQuant KV compression provides massive context wins but introduces potential noise that breaks the precise structural preservation required for script edits.
• –Reliable first-attempt generation for 150-line scripts typically requires 30B+ dense models or high-rank MoEs like Gemma 4 26B A4B.