fix: $ dropped from attribute names

[theRizwan]

Fixes #211. Fixes #306.

Both reports are the same defect seen through different syntax.

The bug

parser().astSync("[#{$attr}]").toString();   // "[#{attr}]"
parser().astSync("[$attr]").toString();      // "[attr]"
parser().astSync("[#{$a}=#{$b}]").toString() // "[#{a}=#{$b}]"

The $ is dropped when it belongs to the attribute name. The value position already worked, which is why [lang=#{$locale}] round-trips in the existing tests while [#{$attr}] does not.

Cause

attribute() has a branch for $ in the value position, and the tokens.caret case immediately below turns $ into the suffix match operator when the next token is =. Nothing handled a $ that is part of the attribute name, so it fell through to tokens.caret, failed the = check, and was discarded without being recorded anywhere.

The new branch mirrors the value-position handling, and is guarded on the next token not being =.

Why the guard matters

Without it, $= stops parsing as an operator. With it, a name and an operator can coexist:

parser().astSync('[$attr$="x"]')
// attribute "$attr", operator "$=", value "x"

[a$=x] and [data-weird-attr$="Something=weird"] are unaffected, and both remain covered by the existing tests.

Verification

Compared against main across 4,800 generated attribute selectors, varying the name, namespace, operator, value, quoting, insensitivity flag and padding:

newly round-tripping	991
regressions (round-tripped before, broken now)	0
parsed before, throws now	0
raw TypeError	45 → 30

The TypeError reduction is a side effect rather than the aim: a trailing $ such as [a$] now returns before reaching the unguarded next[TOKEN.TYPE] lookahead in tokens.caret. That lookahead is the subject of #334 and is not otherwise touched here.

npm test passes, including oxlint, the type check and the coverage thresholds: 94.99% lines, 95.41% branches, 97.71% functions against gates of 94/94/96. Test count 789 to 799.

Tests

Added to nonstandard.mjs, next to the existing sass escapes cases, since these arrive through preprocessor interpolation: interpolated attribute name, bare $ name, interpolation on both sides, a $ name combined with $=, and a namespaced $ name.

Not addressed

[#{$attr} i] still drops the insensitivity flag, both before and after this change. The flag survives only when an operator is present, which is a separate defect and left alone.