Every problem is easy if you oversimplify it.
I send people who come to me struggling with their tests away with permission to fail at them but not permission to give up on getting better. You're gonna write broken tests. Just don't keep writing the same broken tests.
If anyone looking for a PhD or book idea is reading along with this, my suspicion is that it's so difficult because we are either 1) fundamentally doing it wrong (in which case there's room for at least 2 more revolutions in testing method) 2) someone will prove mathematically that it's an intractable problem, Gödel-style, and then someone will apply SAT solvers or similar to the problem and call it a day. Property based testing already pretty much does Monte Carlo simulation...
For backend tests, the penalty at each level of the testing pyramid is about 8x cost for a single test (and IME, moving a test down one level takes 5x as many tests for equivalent coverage, so moving a test down 2 layers reduces the CPUtime by half but also allows parallel execution).
For frontend I think that cost is closer to 10x. So you want to push hard as you can to do component testing and extract a Functional Core for unit tests even harder than you do for backend code. Karma is not awesome but is loads better than Selenium, particularly once you have to debug a failing test. I've been on long projects for a minute so I can't really opine on Puppeteer, but Selenium is hot flaky garbage. I wouldn't tolerate more than 100 E2E tests in it, even on a half million line project. Basically a smoke test situation if you use it like that, but you won't have constant red builds on the same eight tests on shuffle.
I want to say we had 47 E2E tests on a project I thought was going swimmingly from a SDLC perspective. But it might have been 65.