Kilobug wrote "there is no room for an hypothesis like "the coin will fall heads 75% of times, tails 25% of time" There is room for both an hypothesis that predict the coin will always fall heads and this hypothesis has a probability of 75%. And an hypothesis that predict the coin will always fall tails and this hypothesis has a probability of 25%.

Say the coin falls HHTHTHH. Then the hypothesis that it will always fall heads has probability 0%, because they weren't all heads. Same for the hypothesis that it will always fall tails.

It seems like it would be pretty easy, though, to extend Solomonoff induction to have each hypothesis be an algorithm that outputs a probability distribution, and then update for each bit of evidence with Bayes's theorem. If we did that for this example, I think the hypothesis that generated the 25%:75% probability distribution would eventually win out.