Scalafmt on compile races with scaladoc

[nigredo-tori]

This one is tricky to reproduce. I think I've seen it in one of my CI builds (with scalafmt on compile) - the doc task has failed, citing missing types from the same project - but it was gone after I've restarted the build. I'm fairly certain that the reason is as follows.

At the moment, Scalafmt is wired into the compile / compileInputs task, which relies on client / Compile / sources:

sbt:root> inspect tree client/compile/compileInputs
[info] client / Compile / compile / compileInputs = Task[xsbti.compile.Inputs]
...
[info]   +-client / Compile / compile / compileOptions = Task[xsbti.compile.C..
...
[info]   | +-client / Compile / sources = Task[scala.collection.Seq[java.io.F..
...
[info]   +-client / Compile / scalafmtDoFormatOnCompile = Task[Unit]
[info]     +-ThisBuild / scalafmtOnCompile = true

However, the Scaladoc build (the Compile / doc task) directly uses Compile / sources instead:

sbt:root> inspect tree client/doc
[info] client / Compile / doc = Task[java.io.File]
...
[info]   +-client / Compile / sources = Task[scala.collection.Seq[java.io.Fil..

This means that we can have client / Compile / doc and scalafmtDoFormatOnCompile running in parallel. If formatting changes the sources (in my case that was caused by Windows line endings), client / Compile / doc can observe partially regenerated sources, which is problematic.

I suggest we wire formatting directly into Compile / sources instead.

[poslegm]

Published v2.3.3