A controlled six-way benchmark compared Qwen3.8-Flash-Next and Opus 5.5 on implementing a real LlamaStash feature: persisting and restoring llama-server's KV cache across unloads. Twelve pull requests were generated (six model/harness combinations, each run twice) using Claude Code, with Flash-Next executed locally on a Strix Halo laptop via Halogen (~Q4 quantization). The test aimed to fairly evaluate model performance on a practical Rust task after prior methodology criticisms.