Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Named op-assign like `x multiply= 10` is valid Andy (the lexer glues an
identifier to a single `=` not followed by `=`), but the grammar only
accepted symbolic compound operators, so it silently mis-parsed the
construct as two statements. Add an external scanner that recognises the
glued identifier-equals token with the one char of lookahead the lexer
uses, so `a==b` equality is unaffected. Reported by Codex on #165.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The locals query only treated a direct `identifier` name/pattern as a
definition, so identifiers inside `let [a, b] = …`, `let (a, b) = …`,
`let a, b = …` and nested for-patterns like `for (x, y), [a, b, c] in …`
fell through to `@local.reference`, losing definition highlighting and
go-to-definition. Descend through the list/tuple/pattern_sequence
containers (up to two levels) for let, for, and destructured parameters.
Reported by Codex on #165.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

A list comprehension may yield a bare comma tuple, e.g.
`[x, y for x in 1..10, y in 1..10]`. The grammar only accepted a single
expression before the comprehension clauses, so the comma forced the
list-literal path and the following `for` became an ERROR. Accept
`_expression_or_sequence` as the body, mirroring how the parser parses a
tuple before checking for `for`. Reported by Codex on #165.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Codex flagged a possible issue where the external scanner advancing past
`and`/`or`/`in` before returning false could drop the keyword. In
practice tree-sitter resets the lexer to the pre-scan position on a false
return, so `a and b not in c` parses correctly. Add a corpus test that
locks this in. No scanner change needed.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Raw strings use a `#` fence specifically to embed quotes, e.g.
`r#"with "" quotes"#` and `r###"... r#"x"# ..."###` from the manual. The
regex token stopped at the first inner `"`, breaking the rest of the
literal. Recognise raw strings in the external scanner, which counts the
opening `#` run and scans until a `"` followed by exactly that many `#`.
The leading `r` is disambiguated from identifiers (and named op-assign)
by the character after it. Reported by Codex on #165.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>