If yes, doesn't this mean that if the plaintext is not easily and
quickly recognizable as such (as for example ASCII-encoded English
prose might be), brute-force is in practice MUCH harder than the
so-much-time-on-such-hardware description suggests?
If the plaintext is not recognizable at all (random numbers (from a good generator)) and if the attacker has no other way to test if it is the real data, you are right, just brute-forcing solves nothing.
But, other than good random numbers, what kind of data is not recognizable for someone knowing a bit about computers? Text no, Images no, any file format from known programs no, any file format from unknown programs no (eg. low entropy and many more things), sometimes there are checksums as part of the data, and so on...
And if yes, doesn't this mean that a trivial (but unknown)
"pre-encryption", e.g. ROT13 or base-64 encoding makes a brute-force
attack hugely more difficult?
No, ROT13, data compression and other preprocessing steps (other than good encryption with a different key) do not add security.
Because, even if the attacker doesn't know about the preprocessing step (which should not be assumed, listen to Kerckhoff), such things can be detected easily. Eg. ASCII text with ROT13 is still ASCII text (ie. A-Z), just not as readable. Getting an output with mostly english letters when trying AES keys is unusual => time look at it closer.
The practical scenario I have in mind is the encryption of one's own
files on disk, using a standard algorithm, e.g. AES.
Lets take a disk with 1TB. The chance that, with 2^128 keys, multiple keys lead to a sane output (each one of size 2^(8*1024*1024*1024*1024)), is pretty small. And content of a HDD with OS and music and whatever on it is easily recognizable.