• Looks like you can rip the idea of structural tags out too? (conjecture.data)
• To auto-close issues by mentioning them, you need a keyword immediately before each number. Comma-separated lists only close the first issue listed 😢
• Impressively fast, and wow there are several issues this will close!

To auto-close issues by mentioning them, you need a keyword immediately before each number. Comma-separated lists only close the first issue listed 😢

😭

Impressively fast

Deleting code is easy. 😁

and wow there are several issues this will close!

I did say. 😉 I wasn't just suggesting this out of petty deletionism.

Right. All tests passing, should be ready for review.

Do you anticipate reusing TargetSelector?

Maybe. I think there's scope for making its logic smarter in useful ways (and have some ideas that I may or may not pursue), and it's a reasonable thing to have encapsulated either way, so I thought I'd leave it as is.

I'm a little surprised to see the structural-tag machinery disappear entirely.

Did you decide that the diversity it adds to mutation isn't worth the complexity of retaining the full target selector, now that coverage no longer exists?

(I'm not familiar with how useful that diversity was in practice, so I don't have strong opinions about this. Just curious to see it go.)

Did you decide that the diversity it adds to mutation isn't worth the complexity of retaining the full target selector, now that coverage no longer exists?

Decided is maybe a bit of a strong word. Truth be told I started deleting code and just got a bit overenthusiastic.

I think this is the right thing to do though. The TargetSelector code as was was fairly complex and moderately slow (slow enough that until we stopped hitting an attrs performance bug it ended up as the bottleneck), and if it wasn't doing much with the more powerful coverage information, it's unlikely to be doing much now.

More generally, I don't think we have a good sense of how much of the current generation code is actually useful, or a solid way of acquiring that sense, so I think we probably need a stronger design pressure towards "simple and fast" for the main loop than we currently have, especially if the goal is to have an entirely separate fuzzer.