Oh definitely, the array is also a monotonically increasing array of numbers. But I often find that when your data are in L1 cache, linear scan is pretty fast.

Then you can store 102/4=25 floats in these cache-lines. So you could cut the search space in 26 with very little - to no overhead.
I've only implemented this in C++ and Rust, I don't know how to do this in C# 😕

You can stick a fixed sized array of size N in the rest of the space, split the original array into N+1 chunks, and you can use a linear scan to find which chunk you need to search.
So if you assume you have 128 bytes to play with, 24 needed for the array, and you're using 32 bit floats

So, it's still O(n), but what I did in the past for the same problem is build a skiplist inside the data structure's cache-line, the CPU prefetcher *usually* also gets the next line for you, so you have 128 bytes you can query for effectively free, while the array only really needs 16 or 24 bytes.

Me literally every day:
docker-compose up
Shit
docker compose up