Enable '--strictNullChecks'

[ghost]

Fixes #9432

The changes generally amount to:

• Adding | undefined as appropriate.
• Adding ! in many places
  • Particularly on optional members of interfaces like Symbol and Type.
  • Also, flags enums are often used as E | undefined, so we need to write e! & E.SomeFlag.
  • The no-unnecessary-type-assertion lint rule ensures that these are only added where necessary.
  • Suspicious uses of ! are marked with Tracking invalid type assertions #18217.
• Adding !! to make expressions be boolean instead of boolean | undefined.
  • Also sometimes need to add = false to initialize the boolean.
  • Also change condition && optional to condition ? optional : undefined to avoid getting a Optional | undefined | false.
• Change Debug.fail() to return Debug.fail() to tell the compiler that we definitely return from a function

[sandersn]

• Looks like the public API needs to be updated, which is why Travis fails.
• How does this change compile time of the compiler? I think this is higher value than the all-unions change, but I would still hope it's not too much slower than what we have now.

[weswigham]

@sandersn On my workstation jake clean tests goes from a wallclock time of ~1m7s to ~1m10s, so pretty much no difference. Seems good to me.

[weswigham]

I'm on board with all the changes except the non-optional non-undefined parent-pointers. While internally we use them with reckless abandon (the checker only operates on fully bound trees, after all), people using our transforms APIs won't have parent pointers bound (and shouldn't use them) - and making our public API correct is important. The, like, 1000 extra assertions we'd need internally are probably worth it.

Also flags! | <some flags> is a bit odd (I mean, the assertion is never going to go away)? Maybe we should probably be allowing undefined in binary arithmetic without an error?

[ghost]

API tests pass when I run them locally on either windows or linux -- don't know why travis is failing.

[weswigham]

The travis failure repros locally for me (looks like tests\baselines\reference\api\typescript.d.ts has more undefined's than it needs). Do you, perhaps, have an different, cached copy of the input or baselined typescript.d.ts knocking around locally that just won't flush?

[ghost]

Just needed to get Jake to run with --strictNullChecks too, thanks for the help.

[weswigham]

TIL jenkins will refuse to build when there are merge conflicts.