pmt fmt: a line comment between stacked labels and the command needs two passes to settle

[mellonis]

pmt fmt needs two passes to settle on one input shape: a line comment written between a statement's stacked labels and its command. The second pass is stable, so the output is correct — but pmt fmt --check reports a change on a file that has already been formatted, and format-on-save leaves the buffer dirty for one extra round.

Reproduction

main() {
 1:
 // mid
 2: left;
}

pass 1: "main() {\n  1: 2: // mid\n        left;\n}\n"
pass 2: "main() {\n 1: 2:\n    // mid\n    left;\n}\n"
pass 3: identical to pass 2

Pass 1 restacks the labels and pulls the comment up as a trailing comment on the label line, which pushes the command to column 8. Pass 2 sees the restacked labels, decides the comment is an own-line one after all, and settles.

Scope — narrower than it looks

Only the // form, and only between stacked labels and the command. These three neighbours are all idempotent on the first pass:

main() {\n 1:\n /* mid */\n 2: left;\n}\n     block comment      -> stable
main() {\n 1:\n // mid\n left;\n}\n          single label       -> stable
main() {\n // mid\n 1: left;\n}\n            comment above      -> stable

Not a regression

This is inherited, not introduced by the green-tree formatter. The pre-green printer produces byte-identical pass-1 and pass-2 output on the same input, verified against a build of it. The behavior is pinned by a two-pass fixture so it cannot drift silently, and the idempotence claim in the repo's own notes carries the exception.

Why it is worth fixing anyway

fmt --check is meant to answer "is this file formatted?" and on this shape it answers no about a file that is. Anything wiring --check into CI or a pre-commit hook sees a failure with no user-applicable fix other than running fmt twice.

Where to look

The interaction is between label restacking and trailing-comment attribution: on pass 1 the labels are still on separate lines when the comment's role is decided, so it qualifies as same-line trailing; after restacking it no longer does. A fix likely means deciding the comment's role against the layout the printer is about to produce rather than the one it read — or accepting the pass-1 shape as canonical and making pass 2 agree with it instead.

Whichever direction, the property suite in crates/post-machine/tests/fmt_property.rs currently EXCLUDES this shape from generation with the reason recorded; that exclusion comes out as part of the fix, and its removal is the test that the fix worked.