Engineering doesn't typically work this way. There needs to be a tolerance because you can't account for all variables. The question really is, at what probability of likelihood of a crash do you want to attempt to avert it?
Averting 99.5%+ of crashes probably eliminates flexibility beyond what is appropriate for a fighter pilot.
> “Auto-GCAS continuously compares a prediction of the aircraft’s trajectory against a terrain profile generated from onboard terrain elevation data. If the predicted trajectory touches the terrain profile, the automatic recovery is executed by the Auto GCAS autopilot. The automatic recovery maneuver consists of an abrupt roll-to-upright and a nominal 5-G pull until terrain clearance is assured.”
Any nose dive will look like it will hit the ground, won't it? The question is at which attitude does it need to pull up. Obviously, there are many factors involved, but there may be other factors the system is not aware of, that's why it probably makes sense that's overridable. I bet there is a way to do it.
It is mostly likely timing based. It may well calculate how much time is left until collision, or perhaps until a 5G pull will become insufficient to recover. When that time remaining drops below N seconds (which from looks to be perhaps 1-3 seconds if it is the until insufficient time metric, or perhaps an extra few seconds for time until collision) it will activate.
Obviously there are other criteria to prevent a landing approach from being seen as an imminent crash etc. But an override system that cannot be accidentally held by a disoriented or unconscious pilot seems plausible.
Averting 99.5%+ of crashes probably eliminates flexibility beyond what is appropriate for a fighter pilot.