Fix several Scala parse-print idempotency issues found on Apache Flink

[knutwannheden]

Motivation

Parsing the Apache Flink codebase (Scala 2) with rewrite-scala produced 44 parse failures ("is not print idempotent") across 42 files. Triaging them revealed five distinct parser bugs. All 1017 Scala files in apache/flink now round-trip with zero parse errors and zero print diffs.

Summary

• Argument prefix doubling in new Foo(...) calls — whitespace after a comma was captured into argPrefix without advancing the cursor past it, so the argument's own visit consumed it again into a nested prefix. Ordinary arguments masked this because withPrefix overwrites the prefix at the same level, but for J.TypeCast arguments (asInstanceOf) the duplicate sits on the inner qualifier, printing new MyPojo(a.asInstanceOf[Int], b) with a doubled space (or a doubled newline+indent for multi-line argument lists). The cursor now advances to the argument start, as the first-argument branch already did.
• Backticks dropped from package segments — package a.b.`trait` printed without backticks because the package expression was rebuilt from compiler name strings, which strip them. ScalaASTConverter now derives the package name from the pid source span, for both the file-header and braced-package forms.
• Trailing ; dropped — two defects in visitBlock: (a) dotty extends some statement spans to include the trailing ; (e.g. when a val's rhs sits on its own line), which the forward scan couldn't see; the absorbed ; is now recognized the same way consumeTrailingSemicolon does, including a cursor catch-up so the ; can't leak into whitespace; (b) a braced block's final expression had no semicolon handling at all — its ; is now consumed and preserved via the Semicolon marker. Braceless blocks (e.g. match-case bodies) are excluded: there the ; is the enclosing construct's separator and is preserved there.
• enum/given as identifiers produced error trees — valid Scala 2 names are hard keywords in Scala 3, so files like def qualifyEnum(enum: Enum[_]) printed with <error> fragments. ScalaCompilerBridge.parseOne now retries a file under -source 3.0-migration — where dotty's scanner demotes enum/given/erased to identifiers — whenever the default 3.6-migration parse has syntax errors, keeping the first attempt if the retry doesn't parse cleanly either. Each attempt parses under its own reporter so a failed attempt's errors don't surface on a file the retry parsed fine. A global source-version switch was not an option: 3.0-migration disables ≥3.1 parser features such as fewerBraces colon-arguments, which this parser supports.
• Scala 2 do { ... } while (cond) printed garbled — dotty desugars do-while into WhileDo(Block(body, cond), ()), which printed as whiledo {{ ... } while (...)}do. The desugared shape (span starting at the do keyword) is now detected and mapped to J.DoWhileLoop.

Test plan

• New round-trip tests for each bug: TypeCastTest#castAsFirstOfMultipleArguments, NewClassTest#newClassWithMultiLineChainedArgument, CompilationUnitTest#packageWithBacktickedSegment, BlockTest#trailingSemicolonAfterMultiLineInitializer, Scala2CompatTest#enumAsIdentifier / #givenAsIdentifier, ControlFlowTest#doWhileLoop
• ./gradlew :rewrite-scala:test green
• Round-tripped all 1017 Scala files in apache/flink: zero parse errors, zero print diffs

https://claude.ai/code/session_01R2LCngRXpEq7cuF6wu8XmE