Do we need to seed a CSPRNG with a truly random number? Examples and documentation use truly random numbers, but no justification is given (that I can find).
If we were to seed one with a pseudorandom number, I don't see what the difference would be compared to with a truly random seed. If someone finds either of the seeds, then the encryption fails anyway.
You are correct, but the idea is that it's easier to find a successive pseudorandom seed than it is to find a truly random seed. This is especially true if a lot of numbers are generated in succession really quickly (and if the seed ever changes, which it usually does).
Essentially, determining the seed is sufficient to determine the entire output of a pseudorandom generator.
As a result, you want a seed that isn't predictable or determinable.
Pseudorandom output is (under some circumstances as described two paragraphs ago) determinable or predictable.
Beyond that, it is a trade-off. You've already decided to use pseudorandom numbers instead of real randomness, so it is probably an acceptable trade-off in your mind.