Revert #1284 Include mutual dependency in initial invalidation

[lihaoyi]

Ref #1284
Fixes #1461
Fixes #1420
Also fixes com-lihaoyi/mill#3748 downstream

Not sure if there's a better way to fix this? Just opening this to spark discussion

The original bugs are fine, but the solution seems incorrect, and is both overly conservative (invalidating everything based on whitespace?) and not conservative enough (doesn't handle cycles with length > 2?).

I don't know if there's a better answer here, but "invalidate everything in a cycle every time you touch a single file" is problematic, so much so that occasional crashes due to undercompilation seems preferable to universal slowness due to over-compilation. After all, the cost of a compiler crash is just the time spent on a full-clean-recompilation, so the frequency/pain of that needs to be weighed against the time spent re-compiling stuff unnecessarily the other 99% of the time. Maybe we can make it opt-in/out so people can choose the tradeoff that works for themselves?

Now I hate dependency cycles as much as anybody (hence https://github.com/com-lihaoyi/acyclic), but many real-world codebases I have to deal with are not nearly as strict, so I'd like Zinc to remain fast when working with them

[Friendseeker]

[error] 	source-dependencies/abstract-class-to-trait
[error] 	source-dependencies/false-error
[error] 	source-dependencies/nested-type-params
[error] 	source-dependencies/no-type-annotation

@lihaoyi Would you like to rename

/zinc/src/sbt-test/source-dependencies/abstract-class-to-trait/test ->/zinc/src/sbt-test/source-dependencies/abstract-class-to-trait/pending

(and vice versa for the other 3 tests)?

Since's Mill's next release is imminent and Mill build codes heavily feature mutual reference, this is a blocking issue for Mill and I am thinking if we should make a new Zinc release ASAP.

While Mill downgrading Zinc is certainly an option, Zinc fixed a significant performance regression that affected mixed Scala / Java projects so downgrading it would cause a whole different set of slowness problem.

[eed3si9n]

LGTM. We can always try again on the fix now that we know about the general problem with circular deps.

[lihaoyi]

Thanks @eed3si9n ! Could you help me cut a release once this is merged? So I can include it in Mill and close out that issue there

[eed3si9n]

Yea. I've been holding off on making a release since we've added a new Java interface and it's good to batch changes. As far as I know we should be able to get something out soon.

[lihaoyi]

Thanks @eed3si9n, if you can get it out in the next few days let me know and I'll include it in Mill 0.12.0

[eed3si9n]

Yea. I've started the gradual release towards a 1.10.x patch release - sbt/sbt#7776. Hopefully we'll get a new Zinc out today or tomorrow.

[eed3si9n]

https://github.com/sbt/zinc/releases/tag/v1.10.3