In many languages it's a requirement. For example, in Romanian, there are 8 words that collide with „fata“ if you remove the diacritics (fata, fată, fața, față, făta, făță, fâța, fâță).
Given that we have to use diacritics, spaces don't seem like a big deal.
There is one big difference: CLI utilities don't usually care about diacritics (though encoding issues can throw a wrench in that), but they care a lot about spaces. So putting spaces in filenames requires properly quoting or escaping parameters, whereas diacritics does not. That makes one-off shell snippets and scripts a lot more annoying (though TBH I tend to shy away from those anyway, these days).
That has lead to phantom diacritics: reading letters in unfamiliar words/names based on what you assume they are. For example some pronounce Chirica as Chirică because they assume someone forgot to type the breve in ă.
The resulting text was understandable by the "computer people" but not the general population who did not use the networks back then, perhaps somewhat comparable to when some time ago USA parents encountered the "SMS slang" used by their teenagers.
* fată - girl
* fața - the face
* față - face
* făta - was giving birth
* făță - a small fish, or a child who won't sit still
* fâța - was fussing
* fâță - variant of făță
As you might infer from the first 4, Romanian uses postfix "the" and for singular feminine words you can't tell the difference if you use only ASCII.
Use a proper dictionary for that. The very nature of statistical models makes proper translation without context impossible for these systems, especially when uncommon words and diacritics are involved.
That is what context is for.