If it is so easy, why most of them sucks?

Because it's only actually easy of you remove the irremovable thing that makes it hard. This is one of those tropes that honestly really hurts the credibility of ML in the eyes of non-practitioners. Practitioners get really excited about some 75% solution, and talk about how this thing which is bad everywhere is really easy. The average listener doesn't realize how hard that last 25% (why would they? They were just told it's easy!) and end up writing off the whole thing as snake oil.

To be honest, 90% of the code takes 90% of the time and the remaining 10% takes an extra 90% of the time.

That's the point... if you allow transtivity, you get discovery but your recommendations start sucking.

But my recommemdations sucks pretty much everywhere.

Spotify pretty reliably finds me one or two interesting tracks in my discover weekly. That still leaves dozens of tracks I don't particularly care for (although I generally don't actively dislike them) but I still think it's a decent result.

I guess the obvious problem is that for books the equation is different. If you get recommended 20 books and after reading them you only ended up really liking 3 of them, then you'll feel like you've wasted a lot of time reading books you don't care for. Meanwhile if you listen to 20 audio tracks in the worst case scenario you'll have wasted an hour, and you can do that in the background while you do something else.

I've found that Last.fm's recommendations are infinitely better than Spotify's. I feel like Spotify's recommendations have too much influence from their partners and sponsors.

2 interesting tracks out of a 100+, that's not a great result. I'm sure if you listen to 100 random tracks, you will like a couple.

If all recommendation engines can do is essentially try to find the right cohort for you (which is how they all work because they can just try to go "people who liked A also tended to like B"), then you'll never like them if you have narrow tastes. Recommendation engines are blunt instruments, after all. Just think of how fundamentally unreliable "people who liked A tended to like B, so you might like B too" actually is for such a personal topic.

You'd have to wait until recommendation engines can read your mind and run your simulation against books.