Calibrate token reserves for each destination model

When a tool loop uses different models for investigation and review, keep byte-to-token estimates separate by destination model. A ratio calibrated on investigation requests can underestimate an unseen review request and spend the tokens needed to finish. Initialize the review estimate conservatively, reserve its input and output before another retrieval step, and calibrate it only with that model's provider usage. Estimates remain heuristics: they do not guarantee an exact spend ceiling or adequate source coverage. Applies when: A bounded tool workflow routes its final synthesis or review to another model. Keywords: token budget review reserve mixed models calibration tool loop