>I'm not convinced by reason #1 because words are distinct even in speech. No matter how fast or continuous your speech when you say, "The dog jumped", everyone will agree that "dog" is a distinct thing, even if you're illiterate.
Segmenting words is actually a quite complex cognitive task that takes a while to acquire, usually considered in the framework of statistical learning[0].
Another aspect here is that the concept of a word is somewhat fuzzy, with compound words[1] emerging quite naturally, particularly in some languages. For relatively recent examples of compounds in English, consider "website" and "cellphone" (both of which are now, as a separate step, gradually becoming the default meanings of "site" and "phone").
So I don't find it too surprising that once upon a time word segmentation was even more flexible.
[0] https://en.wikipedia.org/wiki/Statistical_learning_in_langua...