I hardcoded a token buffer once and it silently wasted half the window on a big-context model, then overflowed on a small one. One setting can never fit all models, so budgets should scale with the window, not be a fixed number. Percentages beat absolute counts every time I'm juggling different models.