[abashev]

Eclipse marks a string literal that is deliberately not externalised with a trailing //$NON-NLS-n$ comment, where n is the position of the literal on its line. Eclipse recognises a marker only with that exact spelling and only on the line of its literal. The formatter broke this in three ways (#23); this fixes all three, in two commits.

Case 1: a space was inserted into a trailing marker

String a = "x"; //$NON-NLS-1$ came out as String a = "x"; // $NON-NLS-1$. A trailing line comment lands in the level of the statement's last break, and that level is written from its flat text when it fits on one line, so Comment.computeFlat() added the space itself, without the exemption that JavaCommentsHelper.rewrite() has for //noinspection and //$NON-NLS-n$. The missing-space rule is now one method, JavaCommentsHelper.addMissingSpace, applied by both paths, and the comment's width follows its flat text. A trailing /// comment that already had its space is no longer turned into // / either.

Cases 2 and 3: markers left behind when a statement is wrapped or reflowed

Before:

String message = Messages.format(
        "first argument that is long enough to matter",
        "second argument that is also rather long",
        "third one"); //$NON-NLS-1$ //$NON-NLS-2$ //$NON-NLS-3$
String reflowed =
        "a single string literal that is far too long to fit into one hundred and twenty columns, so the"
                + " formatter reflows it"; //$NON-NLS-1$

After:

String message = Messages.format(
        "first argument that is long enough to matter", //$NON-NLS-1$
        "second argument that is also rather long", //$NON-NLS-1$
        "third one"); //$NON-NLS-1$
String reflowed =
        "a single string literal that is far too long to fit into one hundred and twenty" //$NON-NLS-1$
                + " columns, so the formatter reflows it"; //$NON-NLS-1$

OpsBuilder.build pairs each marker of a markers-only trailing comment with the n-th string literal of the comment's input line (within the same partial-format region) and puts a zero-text ReservedWidth after that literal, so the layout already counts the marker on the literal's line and a second run changes nothing. JavaOutput.appendNonNlsMarkers writes each paired marker at the end of its literal's output line, numbered by the literal's position among the string literals written there; a marker whose number has no literal stays where it was. A comment of markers is never wrapped. StringWrapper recomputes the markers of a reflowed chain's line from the new line structure: one marker after every piece, the markers of literals before and after the chain renumbered for their lines, with their room reserved per line.

Left alone by design: a mixed comment such as // note //$NON-NLS-1$, a line on which two reflowed chains change (only reachable with a wrap limit below the formatter's), and a chain that ends in a text block.

Also in this PR

• The marker code hands its lists around as plain List (List.of, List.copyOf, Stream.toList), not Guava's ImmutableList.
• The line-comment wrapper no longer keeps // MOE: comments on one line: google/MOE has been dead for years, so such a comment wraps like any other.
• The conventions of the review, applied here and written down in CLAUDE.md: var for locals, no Error Prone @Immutable on the marker nodes, and the Apache header without a copyright line on the four new files (Audit the license headers of all source files #154 covers the rest of the tree).

Tests

Written first, and failing before each change: trailingLineCommentGetsItsMissingSpaceExceptIdeMarkers with the fixture ojf-issue-23-non-nls-trailing-marker (case 1); nonNlsMarkersStayOnTheLineOfTheirLiteral, aRunOfNonNlsMarkersIsNeverWrapped and the fixtures ojf-issue-23-wrapped-arguments, ojf-issue-23-reflowed-literal and ojf-issue-23-marked-literals (cases 2 and 3). The fixtures also run the idempotency checks. Unit tests cover the utility classes: NonNlsMarkersTest, ReservedWidthTest, and two JavaCommentsHelperTest cases for addMissingSpace.

Verification

• ./gradlew test over all modules is green on the rebased branch; ./gradlew formatDiff changes nothing.
• JDK 21 corpus, 15,747 files: case 1 changes 4 files, all a trailing /// ... comment that the flat path used to turn into // / ...; cases 2 and 3 change none, as the JDK carries no Eclipse markers.

Fixes #23. The same defect upstream: google/google-java-format#221, google/google-java-format#361.