This is a real effect, and this article gives an example with URL's: https://towardsdatascience.com/the-art-of-prompt-design-prompt-boundaries-and-token-healing-3b2448b0be38
":" and "://" are different tokens in this LLM, so prompting with a URL starting with "http:" gives bad results because it can't use the "://" token.
Although this can be improved with a technique called "token healing" that essentially steps backwards in the prompt and then allows any next token that starts with the same characters in the prompt (i.e. in the "http:" example, it steps backwards to "http" and allows any continuation that starts with ":" in its first token).
Note that this only applies at the level of tokens, so in your example it's true that the next token can't be ": T", but with standard tokenizers, you'll also get a token for every substring of your longer tokens, so it could be just "T". Whether this makes things better or worse depends on which usage was more common/better in the training data.
I was thinking about how prompt differs from training data in terms of tokenization. If i am to prompt with "solution:" as opposed to "solution: " it seems like it can influence the result, as in training data last token contain some information about next token. If there is token ": T" but my prompt ended in ": " it can be inferred than next token can't be "T[something]".
Is this real effect, or I just misunderstand how tokenization works?