A point overlooked—which may or may not be a factor here—is how an ‘error’ is triggered, regardless of whether checking is ‘on’.
Different (parts of) apps trigger on differing initial string length. Most often this is 3 or 4 characters at string start. In some cases, I’ve found spelling mistakes in the first 3 characters aren’t even detected. More benignly errors in the first 3–4 characters may be detected and have the red ‘underlining’ but no corrections are offered. A lack of completions for short stems reflects the large number of possible completions.
For a dyspraxic typist like myself where errors are less unknown spelling that left/right input queuing errors, spell remains an unreliable assistant. I would note, as a general AI skeptic, that Apple’s Writing tools → Proofread, is actually remarkably good even if it punishes idiomatic writing.