Flag JS Literals and ignore assignments/accesses to invalid props, instead of adding an index

[weswigham]

Fixes #25987

As mentioned in-person with @mhegazy, since our goal with the index signature was to just squelch a class of errors in JS, this accomplishes the same thing without the unfortunate side effect of mangling our inferences. This has the fortunate side-effect that the meaningless index signatures are no longer present in quickinfo and the like.

These accesses are also now marked in noImplicitAny mode now - which I'd argue is a correct thing to do. That is the breaking change referenced in the PR labels, however in practice it's unlikely to break much (and when it does it's probably desirable!) as JS codebases using noImplicitAny are relatively rare.

[brendankenny]

As an aside, these accesses could totally be marked in noImplicitAny mode now - which I'd argue is a correct thing to do; we can put off deciding to do that, though.

One downside of this change (if I'm understanding it correctly) is that not having [x: string]: any; in the interface anymore means that it's going to be more difficult to diagnose what's going on when those accesses are not marked as an error. Right now it's kind of a process of hovering for quickinfo and then realizing, "oh yeah, tsc treats object literals loosely here" :)

Having keyof, etc work as expected is definitely an improvement for checked js, but it would be fantastic if we're also able to opt into errors for them with noImplicitAny.

[sandersn]

You may have told me in person, but: how do the user tests do with this change?

[mhegazy]

Since we are touching this area, i agree with @brendankenny that we should make --noImplicitAny work as expected. we have talked about this before but never done it.

[weswigham]

@mhegazy Alright, we'll no longer squeltch the errors under noImplicitAny. It's up now~

@sandersn I fixed a pair of small problems (like also disabling excess property checks on js literals outside of noImplicitAny mode), but it looks like now the changes are mostly pretty minimal. There's a lot of churn of index signatures getting removed and I believe errors from the recent lib update which we'll need to look over tomorrow; but this looks pretty good, IMO.

[sandersn]

Looks good, just some nits about comments.

[sandersn]

quibble: these properties are not "imaginary". They are "possibly existing" properties.

Seriously, though, I'd say "expando properties". I also don't think that referencing the old implementation helps, so I'd delete that line or at least change it to "This mirrors the behaviour of index signatures".

[sandersn]

I would move the comment block up here.

[weswigham]

I'm going to hold off on merging this until our user/rwc baselines are clean, so we can clearly see its impact on them. AFAIK we're not shipping this in a 3.0.2 (it's technically a break, after all), so this isn't in much of a rush.