fix(lint): apply suppression comments to relocated footnote definitions

[VXNCXNX]

Fixes #1078.

What's broken

Suppression comments do not suppress alerts inside footnote definitions.

$ cat mwe.md
Here is some text that has a footnote[^x].

<!-- vale Vale.Spelling = NO -->

[^x]: I want to talk about thingamajigg in a footnote.

<!-- vale Vale.Spelling = YES -->

Some more text here about nothing.

$ vale --output=line mwe.md
mwe.md:5:40:Vale.Spelling:Did you really mean 'thingamajigg'?

The same document with the footnote replaced by a plain paragraph reports nothing.

The cause

Not what the issue reporter guessed. goldmark's Footnote extension relocates every definition into a trailing div.footnotes, which I confirmed by dumping the rendered HTML:

<!-- vale Vale.Spelling = NO -->
<!-- vale Vale.Spelling = YES -->
<p>Some more text here about nothing.</p>
<div class="footnotes" ...>
<li id="fn:1"><p>I want to talk about thingamajigg ...</p></li>

The toggle state is a linear walk over the token stream, so by the time the relocated text is reached, the = YES has already re-enabled the check. The telling detail is the converse: an unterminated = NO does suppress the footnote, which is only possible if position in the stream, not the footnote itself, is the variable.

The fix

Inside that div, rebuild the state from the definition's own source line by replaying the comment controls that precede it, then restore the ambient state on the closing tag.

One deviation worth calling out: I key on the definition's source position, not the reference's. The issue's own example has the reference on line 1, outside the suppression block, with only the definition inside it. Keying on the reference leaves that example still reporting the alert. I built it that way first and confirmed it.

MyST and qmd

The div.footnotes container is a goldmark promise, not a general one, so the branch is gated on .md, .mdx, .myst and .qmd. Those are exactly the four whose converters register extension.Footnote, so rST, AsciiDoc and HTML never enter it. MyST and qmd are included because they relocate footnotes too and had the same bug; their existing tests pass unchanged.

Verification

go build ./... clean and go test ./internal/... passes in full.

The MWE above now reports nothing and exits 0. Deleting the suppression comments brings the alert back.

The interesting case is this one, which broke my first attempt:

An intro mentioning thingamajigg early on, plus a footnote[^x].

<!-- vale Vale.Spelling = NO -->

[^x]: I want to talk about thingamajigg in a footnote.

Resolving the definition by searching the whole document for its longest word finds line 1, so the replay stopped before the = NO and the definition was not suppressed. It now walks the source's definition starts and picks the one whose body contains the word. Only the line-1 alert is reported, which is correct.

Also checked: two footnotes where one is suppressed and one is not, and a definition whose longest word recurs in a later paragraph. Both behave.

One honest weakness. f.Content is prepMarkdown-processed, which stars out the label, so [^x]: becomes [**]: and the regex has to match definitions by shape rather than by \[\^. That means a link reference definition could in principle be selected first if it contained the same longest word. I could not construct a case where this produced a wrong suppression, since link definition labels are starred out and their URLs are not linted, but the ambiguity is there and I would rather you knew about it than found it.

This touches the same walker branch as #1001, #1052 and #1111. The change is purely additive, three touch points in the loop and a self-contained type below lintHTMLTokens, so it should not conflict badly.

[jdkato]

Thanks — the diagnosis is right, but I don't want to take this mechanism. It adds a second text-to-source heuristic beside the one the walker already uses for alert positions, mutates f.Comments mid-walk, and hardcodes goldmark's div.footnotes into the format-agnostic walker.