fix: keep backticks on names containing $ in fmt

[prql-bot]

prqlc fmt drops the backticks around any name containing $, and the result no longer lexes as the same name — $ starts a $param token in PRQL, so the formatter turns one identifier into an identifier followed by a parameter.

$ printf 'from t\nselect {`a$b`}\n' | prqlc fmt
from t
select {a$b}

$ printf 'from t\nselect {`a$b`}\n' | prqlc fmt | prqlc compile
Error:
   ╭─[ :2:9 ]
   │
 2 │ select {a$b}
   │         ┬
   │         ╰── expected a function, but found `this.t.a`
───╯

This is the same round-trip break as #6200 and #6214, at the one remaining hole in the rule those two consolidated. valid_prql_ident() accepts $ in both leading and subsequent positions because its regex was copied from the SQL-side utils::valid_ident(), where $ genuinely is a Postgres identifier character. The PRQL lexer's ident_part() never accepts it. Dropping $ from the codegen regex is the whole fix; params are written separately as format!("${id}") and don't go through this path.

This change doesn't move the SQL: `a$b` compiles to SELECT a$b FROM t before and after it. Worth knowing when reading that claim, though — the default output path renders it as SELECT a $b FROM t, because sqlformat splits the $ the same way it does in #3188, so the SQL a user actually sees names a different identifier and isn't valid in the dialects that accept a$b. That's a pre-existing bug independent of fmt, and it currently has no open tracker: #3286 was closed as a duplicate of #3077, whose fix covered $param but not an identifier containing $.

Failure modes and verification

Three shapes, all now formatted with backticks intact:

Source	Before this PR	Re-parses as
select {`a$b`}	select {a$b}	error — a applied to $b
select {x = t.`$foo`}	select {x = t.$foo}	error — $foo where an ident part is expected
let `$foo` = 5	let $foo = 5	error — $foo where a declaration name is expected

The regression test is codegen::ast::test::test_quoted_dollar_in_name, covering all three. assert_is_formatted round-trips through the parser, so each case asserts the formatted output re-parses to the same source. Restoring $ to the character class while keeping the test fails it:

---- codegen::ast::test::test_quoted_dollar_in_name stdout ----
assertion failed: `(left == right)`
  left: `"from t\nselect {`a$b`}"`
 right: `"from t\nselect {a$b}"`

No snapshot moved — nothing in the repo's test corpus uses a $ in an identifier.

task prqlc:pull-request runs 763 tests; 762 pass. The one failure is queries::results::read_csv, which needs to download DuckDB's json extension and can't reach extensions.duckdb.org from this sandbox — unrelated to this change. cargo fmt --check and cargo clippy -p prqlc --all-targets --no-default-features --features=default,lsp -- -D warnings are both clean.

One related divergence I did not change: display_ident_part() in prqlc-parser has the same $-at-start hole, as the other surviving copy of the "needs backticks" rule. Since #6214 routed ExprKind::Ident through write_ident_part, that copy only backs Ident's Display impl — error messages and Debug output, which nothing re-parses — so there's no round-trip to break there. Worth folding into write_ident_part eventually, but that's a separate change with its own snapshot churn.

[prql-bot]

Two points, neither in the regex change itself.

No changelog entry. development.md asks for a CHANGELOG.md line on user-facing changes, and this is one — the sibling fmt fix #6311 has an entry in the same [unreleased] block. Pushing one to this branch.

The description's SQL claim only holds with --no-format. It says `a$b` "still compiles to SELECT a$b FROM t, which is correct for the dialects that allow it". The default output path splits the $:

$ printf 'from t\nfilter `a$b` == 1\n' | prqlc compile --hide-signature-comment
SELECT
  *
FROM
  t
WHERE
  a $b = 1

a $b is a different identifier from the one that was written, and it isn't valid in the dialects that accept a$b either. This is sqlformat splitting the token rather than anything in this diff — same class as #3188 — but the description asserts the opposite of what a user of the default path sees, so I'll correct it there.

Worth flagging separately, since it's adjacent and currently untracked: the parameter half of this was fixed ($sd renders intact now, via #3077), but an identifier containing $ still splits, and #3286 was closed as a duplicate of #3077 so nothing open covers it. Closing it on the PRQL side would mean dropping $ from utils::valid_ident() as well, so these names get dialect-quoted as "a$b" — that changes generated SQL for anyone using $ in identifiers on Postgres today, which is a maintainer call rather than a follow-on to this fix.