They don't. Too many attack vectors. I illustrate here:

https://news.ycombinator.com/item?id=10468624

My solution was to print the TCB on a process node that was verifiable by eye. Then, verify a random sample of each batch. Possibly speed it up with image processing algorithms if producing the same component or components.

Note: Deviations from intended circuitry in deep sub-micron can have measurable differences at analog or RF level. DARPA is funding research to do such things. A monitor at visible node could then be combined with CPU's on cutting-edge node. Common practice in commercial sector is obfuscations, though.

Basically splitting the trusted circuit and testing the parts separately. This requires a trusted master circuit, but it can be arbitrarily small.

See https://perso.uclouvain.be/fstandae/PUBLIS/177.pdf

But what if the malicious code is time activated? (just an example)

This is actually addressed in the paper. Basically you can use testing to detect the timebomb, up to a negligible probability.

This paper is approachable, it's understandable without too much background if you're interested in the topic.