Support flexible TypedDict creation/update (#15425)
Fixes #9408
Fixes #4122
Fixes #6462
Supersedes #13353
This PR enables two similar technically unsafe behaviors for TypedDicts,
as @JukkaL explained in
https://github.com/python/mypy/issues/6462#issuecomment-466464229
allowing an "incomplete" TypedDict as an argument to `.update()` is
technically unsafe (and a similar argument applies to `**` syntax in
TypedDict literals). These are however very common patterns (judging
from number of duplicates to above issues), so I think we should support
them. Here is what I propose:
* Always support cases that are safe (like passing the type itself to
`update`)
* Allow popular but technically unsafe cases _by default_
* Have a new flag (as part of `--strict`) to fall back to current
behavior
Note that unfortunately we can't use just a custom new error code, since
we need to conditionally tweak some types in a plugin. Btw there are
couple TODOs I add here:
* First is for unsafe behavior for repeated TypedDict keys. This is not
new, I just noticed it when working on this
* Second is for tricky corner case involving multiple `**` items where
we may have false-negatives in strict mode.
Note that I don't test all the possible combinations here (since the
phase space is huge), but I think I am testing all main ingredients (and
I will be glad to add more if needed):
* All syntax variants for TypedDicts creation are handled
* Various shadowing/overrides scenarios
* Required vs non-required keys handling
* Union types (both as item and target types)
* Inference for generic TypedDicts
* New strictness flag
More than half of the tests I took from the original PR #13353
