Make spec RandomTree recognise spec MRCAPrior constraints

[alexeid]

Problem

beast.base.spec.evolution.tree.coalescent.RandomTree still imported the legacy beast.base.evolution.tree.MRCAPrior. The spec MRCAPrior extends Distribution directly rather than subclassing the legacy class, so the instanceof MRCAPrior scan over the tree's outputs never matched and monophyletic constraints were silently dropped when initialising the starting tree.

Any calibrated analysis then starts from a random tree that violates its own monophyly constraints:

Start likelihood: -Infinity after 10 initialisation attempts
    P(HumanGorilla.prior) = -Infinity
    P(MonkeyApe.prior) = NaN  **
    ...
java.lang.RuntimeException: Could not find a proper state to initialise. Perhaps try another seed.

Reproduced on seeds 1, 7, 42 and 12345 — it is not seed-dependent.

There was no XML-level workaround: calibrationsInput was typed to the legacy class too, so an explicit <constraint idref="..."/> would fail type-checking, and legacy MRCAPrior.distInput requires a ParametricDistribution while spec priors are ScalarDistribution — incompatible hierarchies.

Impact

This is not limited to one package. beast-fx/src/main/resources/beast.fx/fxtemplates/Standard.xml pairs spec RandomTree with spec MRCAPrior, so every calibrated analysis produced by BEAUti under BEAST 3 is affected.

Fix

Switch the import, calibrationsInput and the distributions list to the spec MRCAPrior / ScalarDistribution types. ScalarDistribution.inverseCumulativeProbability returns a boxed value, so it is unboxed explicitly when computing the calibration bounds.

The legacy beast.base.evolution.tree.coalescent.RandomTree is untouched, so legacy XMLs keep working through the legacy class. No other caller passes MRCAPriors into the spec class — RandomTreeTest constructs one without constraints and RandomGeneTree does not touch the constraint machinery.

Verification

• Full beast3 suite: BUILD SUCCESS, 527 tests (47 pkgmgmt + 458 base + 22 fx), 0 failures.
• Found via bModelTest's examples/beast2-paper.xml (7 monophyletic clade calibrations), which failed to initialise on every seed before this change and now starts at a finite likelihood and samples normally:

  Start likelihood: -41052.324347071335
                0    -41052.3243              N    -39681.3021     -1371.0222
            10000    -26682.0022         2.0       -26499.3254      -182.6768
  

• The other two bModelTest examples were re-run as regression checks and are unaffected.

---

Second commit: LoggerTest isolation

CI on this branch went red on LoggerTest.testFileLog, which is unrelated to the change above — it fails or passes purely on where surefire schedules it relative to the integration tests. Three JVM-wide globals are set by those tests and never restored: file.name.prefix (XMLPathUtil.setUpOutputDir), Logger.FILE_MODE, and Logger.sampleOffset (ResumeTest).

Run ordering is not stable between CI runs, which is why master has been green:

Run	prefix-setting tests	LoggerTest	Result
master 29714663831	69, 70, 71	57 (before)	green
this branch 30044812322	26, 28, 30	58 (after)	red

When LoggerTest runs second, openLogFile() prepends the stale prefix and writes to <prefix>beast.log while the test resolves beast.log against the working directory, so the existence check fails; and the stale sampleOffset is added to every logged sample, so the round-tripped sample number is off by the leaked offset.

Fixed by resetting all three in setUp(). Verified by forcing the ordering — with -Dsurefire.runOrder=reversealphabetical over ExampleXmlParsingTest, ResumeTest and LoggerTest, testFileLog fails on a clean tree before the change and passes after it, as it does under alphabetical and filesystem ordering. Full beast-base suite: 458 tests, 0 failures.