Java 25: strip javac's EOI sentinel from /// markdown doc comment text

[timtebeek]

A /// markdown doc comment whose content ends in a backslash round-trips with a stray 0x1A appended, so the file fails the print idempotency check:

java.lang.IllegalStateException: <file> is not print idempotent.
	at org.openrewrite.Parser.requirePrintEqualsInput(Parser.java:52)

• Found while triaging ParseFailures rows from a large ingest. Split out of Java: fix two comment-scanning defects that drop modifiers from the LST #8564, which fixes two unrelated comment-scanning defects in the same area.

Root cause

This one is upstream in javac, not in our code — we copy the string it hands us. com.sun.tools.javac.tree.DCTree.DCRawText#getContent() can return the content of a JEP 467 markdown doc comment with javac's LayoutCharacters.EOI (0x1A) sentinel appended; its escape lookahead runs past the end of the comment buffer. Visible with a plain DocTrees probe, no OpenRewrite involved:

///x\   ->  DCRawText.getContent() = [x\<0x1A>]
///x    ->  DCRawText.getContent() = [x]

Confirmed on both JDK 25.0.3 and JDK 26.0.1, so it isn't something we can wait out upstream. I'll report it to OpenJDK separately.

Reproducer

class Test {
    ///x\
    void m() {}
}

prints back as ///x\<0x1A>.

The practical trigger is a doc comment ending in backslashes. Banner comments hit this, because four-or-more leading slashes still start with /// and javac therefore classifies them as markdown doc comments rather than ordinary line comments:

class Test {
	////Section\\\\
	void first() {}
}

The exact javac predicate is erratic — trailing runs of 1, 2, 4, 5, 7 and 8 backslashes leak but 3 and 6 do not; ///a\b leaks while ///a\bc and ///a\*b do not — so this strips defensively rather than trying to model the condition.

Fix

Strip a trailing sentinel in visitText(String), the single point every raw content string passes through. Leaving it in place corrupts the text and also desynchronizes the cursor, since the sentinel isn't present in source.

Scope

• Regression in 8.76.0, introduced by Add support for Java 23+ triple-slash markdown Javadoc (///) #7087, which routed /// comments through ReloadableJava25JavadocVisitor. Last good release is 8.75.11 — before that these lines stayed plain TextComments and round-tripped fine.

Java 25 parser only, since /// doc comments are JDK 23+.

Testing

Three new tests in JavadocTest.MarkdownDocComment; all three fail on main and pass with the fix. rewrite-java-25:compatibilityTest green.

[timtebeek]

Split out of #8564 — unrelated root cause, so it can land independently.

[timtebeek]

Couldn't really find a good way to shorten this oddity to two lines or less

[timtebeek]

Shortened to two lines in 670e4ef — kept that it's javac's doing, which API, the trailing-backslash trigger, and the cursor consequence:

// javac can append its end-of-input sentinel to `DCRawText#getContent()` for `///` doc comments,
// e.g. when the content ends in a backslash. It is absent from `source`, so it also skews the cursor.

Rebased onto your eb860a2 rather than force-pushing over it.