[phdoerfler]

Keep coverage data in the class directory on sbt 2

Fixes #696.

On sbt 2, compile is cached. A cache hit restores only the declared outputs, i.e. the class directory. The scoverage-data directory under crossTarget (scoverage.coverage plus the directory the instrumented classes write measurements into) is a side effect of compilation, so it is not restored. After clean, a fresh checkout, or on a CI agent with a warm cache:

• forked tests fail with FileNotFoundException on scoverage.measurements.*, or
• coverageReport logs No coverage data, skipping reports and skips the minimum check.

This changes the sbt 2 default of coverageDataDir to Compile / classDirectory, the workaround from #696. The sbt 1 default stays crossTarget. The switch lives in ScoverageSbtPluginCompat, next to the existing Def.uncached.

Trade-offs:

• On sbt 2, reports now go to <classes>/scoverage-report and <classes>/coverage-report.
• A jar packaged in the same coverage-enabled invocation contains scoverage-data. A normal compile rewrites the class directory, so regular packaging is unaffected.
• Users who set coverageDataDir themselves keep their value. The README notes it should stay inside the class directory on sbt 2.

An alternative would be to move only scoverage-data into the class directory and keep reports under crossTarget. That needs a separate path for data vs. reports, including in coverageAggregate. I decided against that since both variants have their downsides and the one presented in this PR is at least a very minimal change.

The largest change is the addition of a new scripted test. cached-compile runs clean; coverage; testOnly; coverageReport twice with Test / fork := true and a 100% minimum. On sbt 2 (sbt "++ 3.x" "scripted scoverage/cached-compile"), it fails without the change (FileNotFoundException in round 2) and passes with it. It also passes on sbt 1. Note that CI currently runs scripted on sbt 1 only.

fwiw: with Test / fork := false, round 2 reports 0% (noted in #696). That looks like a separate in-JVM Invoker issue (#522).

[tgodzik]

Thanks for the contribution! Looks like the tests are failing when Scala 3.8.x is used on JDK 8. Not sure if there is a way to run specific test only on sbt 2, so maybe we can just run it on 2.13 (we are not really testing Scala versions itself here)

[tgodzik]

Looks like we can't run it on JDK 8, but since we are actually testing sbt 2, we can just use Scala 2.13 here.

Suggested change

	scalaVersion := "3.8.4"
	scalaVersion := "2.13.18"

[tgodzik]

Alternatively, we can check the java version and switch between Scala versions.