Refactor PEP 661 sentinel handling to use a synthetic nominal class

[edgarrmondragon]

Related

• Fixes Invalid return type with sentinel and dict.get #21866
• Hard left turn from initial implementation in Add support for PEP 661 (sentinels) #21647
• Motivation: Fix sentinel identity loss through generic substitution and inference #21888 (comment)

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[ilevkivskyi]

I guess you want to cc @JelleZijlstra here.

[JelleZijlstra]

If this fits better in mypy's model, I'm OK with landing it, but it feels worse to me. A sentinel isn't its own type, it's a single object that is an instance of the type sentinel. Treating it more like a literal makes more sense than treating it like a NewType.

[A5rocks]

Maybe we can do the best of both worlds (less fragile, but fits @JelleZijlstra's mental model) by adding a test in copy_modified that if last_known_value is being removed, the type isn't a sentinel? IDK if there's a way to compile a check only into testing mypys instead of released mypys though!

[edgarrmondragon]

adding a test in copy_modified that if last_known_value is being removed, the type isn't a sentinel?

I'd be happy with that compromise, but yeah I don't know if that's feasible.