llmfit's Fallback Estimate Can Be Fourteen Times Too Slow
Decode is memory bound and prefill is compute bound, so Q4 halves one and leaves the other exactly where it was. Both numbers come from formulas the project publishes, which is what makes them checkable.
