The implementation really depends on the hardware and driver. AMD hardware is capable of synthesizing sampler descriptors using shader code, so it's very possible that they do this for static samplers. With Nvidia hardware nobody knows how it really works, so it's possible that they turn the static samplers into registers or store descriptors located in memory. You'll notice in the docs that there's a limit of 2032 total unique static samplers across all live root signatures, which suggests that there are implementations that may need to convert static samplers into something heap-based.
With that in mind, I wouldn't make decisions based off of implementation details since they're not guaranteed. None of the IHV's have said anything positive or negative about static sampler performance or limitations, so I would just make your decision based on what works best for your engine. Personally I like static samplers since they spare me from having to create and bind descriptor tables, which I find pretty convenient.