sqlglot_rust/parser/

sql_parser.rs

use crate::ast::*;
use crate::errors::{Result, SqlglotError};
use crate::tokens::{Token, TokenType, Tokenizer};

/// Convert a token's `quote_char` into a `QuoteStyle`.
fn quote_style_from_char(c: char) -> QuoteStyle {
    match c {
        '"' => QuoteStyle::DoubleQuote,
        '`' => QuoteStyle::Backtick,
        '[' => QuoteStyle::Bracket,
        _ => QuoteStyle::None,
    }
}

/// A recursive-descent SQL parser.
///
/// Supports CTEs (WITH), subqueries, UNION/INTERSECT/EXCEPT, CAST,
/// window functions (OVER), EXISTS, EXTRACT, INTERVAL, and more.
pub struct Parser {
    tokens: Vec<Token>,
    pos: usize,
    /// Whether to preserve comments during parsing.
    #[allow(dead_code)]
    preserve_comments: bool,
    /// Accumulated comments pending attachment to the next AST node.
    pending_comments: Vec<String>,
}

impl Parser {
    /// Create a new parser from a SQL string.
    pub fn new(sql: &str) -> Result<Self> {
        let mut tokenizer = Tokenizer::new(sql);
        let tokens = tokenizer.tokenize()?;
        Ok(Self {
            tokens,
            pos: 0,
            preserve_comments: false,
            pending_comments: Vec::new(),
        })
    }

    /// Create a new parser that preserves SQL comments in the AST.
    pub fn new_with_comments(sql: &str) -> Result<Self> {
        let mut tokenizer = Tokenizer::with_comments(sql);
        let tokens = tokenizer.tokenize()?;
        Ok(Self {
            tokens,
            pos: 0,
            preserve_comments: true,
            pending_comments: Vec::new(),
        })
    }

    // ── Comment helpers ────────────────────────────────────────────

    /// Consume any comment tokens at the current position, accumulating
    /// their text into `pending_comments`.
    fn collect_comments(&mut self) {
        while self.pos < self.tokens.len() {
            match self.tokens[self.pos].token_type {
                TokenType::LineComment | TokenType::BlockComment => {
                    let token = &self.tokens[self.pos];
                    self.pending_comments.push(token.value.clone());
                    self.pos += 1;
                }
                _ => break,
            }
        }
    }

    /// Take all pending comments, leaving the buffer empty.
    fn take_comments(&mut self) -> Vec<String> {
        std::mem::take(&mut self.pending_comments)
    }

    // ── Token helpers ──────────────────────────────────────────────

    fn peek(&self) -> &Token {
        &self.tokens[self.pos.min(self.tokens.len() - 1)]
    }

    fn peek_type(&self) -> &TokenType {
        &self.peek().token_type
    }

    fn advance(&mut self) -> &Token {
        let token = &self.tokens[self.pos.min(self.tokens.len() - 1)];
        if self.pos < self.tokens.len() {
            self.pos += 1;
        }
        token
    }

    fn expect(&mut self, expected: TokenType) -> Result<Token> {
        let token = self.peek().clone();
        if token.token_type == expected {
            self.advance();
            Ok(token)
        } else {
            Err(SqlglotError::ParserError {
                message: format!(
                    "Expected {expected:?}, got {:?} ('{}') at line {} col {}",
                    token.token_type, token.value, token.line, token.col
                ),
            })
        }
    }

    fn match_token(&mut self, expected: TokenType) -> bool {
        if self.peek().token_type == expected {
            self.advance();
            true
        } else {
            false
        }
    }

    /// Check if the current token's uppercased value matches a keyword string.
    fn check_keyword(&self, keyword: &str) -> bool {
        self.peek().value.to_uppercase() == keyword
    }

    /// Check if the token at `current + offset` matches a keyword string.
    fn check_keyword_offset(&self, keyword: &str, offset: usize) -> bool {
        let idx = self.pos + offset;
        if idx < self.tokens.len() {
            self.tokens[idx].value.to_uppercase() == keyword
        } else {
            false
        }
    }

    /// Match a keyword by string value (for multi-word context-sensitive keywords).
    fn match_keyword(&mut self, keyword: &str) -> bool {
        if self.check_keyword(keyword) {
            self.advance();
            true
        } else {
            false
        }
    }

    /// Expect a keyword by string value, returning an error if not found.
    fn expect_keyword(&mut self, keyword: &str) -> Result<()> {
        if self.check_keyword(keyword) {
            self.advance();
            Ok(())
        } else {
            let token = self.peek().clone();
            Err(SqlglotError::ParserError {
                message: format!(
                    "Expected keyword '{keyword}', got '{value}' at line {line} col {col}",
                    value = token.value,
                    line = token.line,
                    col = token.col
                ),
            })
        }
    }

    /// Reconstruct a single token's surface representation for raw command
    /// preservation. String literals are wrapped in their original quotes;
    /// identifiers may carry a quote_char from the tokenizer.
    fn token_text(token: &Token) -> String {
        match token.token_type {
            TokenType::String => format!("'{}'", token.value.replace('\'', "''")),
            TokenType::Identifier if token.quote_char != '\0' => {
                let (l, r) = match token.quote_char {
                    '[' => ('[', ']'),
                    c => (c, c),
                };
                format!("{l}{}{r}", token.value)
            }
            _ => token.value.clone(),
        }
    }

    /// Join a slice of tokens with whitespace tuned for SQL — no space
    /// before `,` `)` `;` `.`, no space after `(` or `.`.
    fn join_tokens_for_raw(tokens: &[Token]) -> String {
        let mut out = String::new();
        let mut prev_no_space_after = true; // suppress leading space
        for t in tokens {
            let no_space_before = matches!(
                t.token_type,
                TokenType::Comma
                    | TokenType::RParen
                    | TokenType::Semicolon
                    | TokenType::Dot
                    | TokenType::RBracket
            );
            if !out.is_empty() && !prev_no_space_after && !no_space_before {
                out.push(' ');
            }
            out.push_str(&Self::token_text(t));
            prev_no_space_after = matches!(
                t.token_type,
                TokenType::LParen | TokenType::Dot | TokenType::LBracket
            );
        }
        out
    }

    /// Consume tokens up to (but not including) the next top-level `;` or EOF,
    /// returning the raw text of the consumed tokens with whitespace
    /// reconstructed by [`join_tokens_for_raw`]. Honors parenthesis depth so
    /// embedded `;` inside `(...)` does not terminate the statement.
    fn consume_raw_to_statement_end(&mut self) -> String {
        let start = self.pos;
        let mut depth: i32 = 0;
        while self.pos < self.tokens.len() {
            let tt = &self.tokens[self.pos].token_type;
            match tt {
                TokenType::Eof => break,
                TokenType::Semicolon if depth == 0 => break,
                TokenType::LParen | TokenType::LBracket => {
                    depth += 1;
                    self.pos += 1;
                }
                TokenType::RParen | TokenType::RBracket => {
                    // A closing paren at depth 0 belongs to an enclosing
                    // context (e.g. CTE body, subquery) — stop without
                    // consuming it.
                    if depth == 0 {
                        break;
                    }
                    depth -= 1;
                    self.pos += 1;
                }
                _ => self.pos += 1,
            }
        }
        Self::join_tokens_for_raw(&self.tokens[start..self.pos])
    }

    /// Parse a comma-separated list of raw items inside an already-opened
    /// parenthesized context. Stops at the matching `)` and returns each item
    /// reconstructed from tokens.
    fn parse_parenthesized_raw_items(&mut self) -> Result<Vec<String>> {
        let mut items = Vec::new();

        // Allow empty parens for tolerance.
        if self.match_token(TokenType::RParen) {
            return Ok(items);
        }

        loop {
            let start = self.pos;
            let mut paren_depth: i32 = 0;
            let mut bracket_depth: i32 = 0;

            while self.pos < self.tokens.len() {
                match self.peek_type() {
                    TokenType::Eof => break,
                    TokenType::LParen => {
                        paren_depth += 1;
                        self.pos += 1;
                    }
                    TokenType::RParen => {
                        if paren_depth == 0 && bracket_depth == 0 {
                            break;
                        }
                        if paren_depth > 0 {
                            paren_depth -= 1;
                        }
                        self.pos += 1;
                    }
                    TokenType::LBracket => {
                        bracket_depth += 1;
                        self.pos += 1;
                    }
                    TokenType::RBracket => {
                        if bracket_depth > 0 {
                            bracket_depth -= 1;
                        }
                        self.pos += 1;
                    }
                    TokenType::Comma if paren_depth == 0 && bracket_depth == 0 => break,
                    _ => self.pos += 1,
                }
            }

            if start == self.pos {
                let token = self.peek().clone();
                return Err(SqlglotError::ParserError {
                    message: format!(
                        "Expected expression inside parenthesized list, got '{}' at line {} col {}",
                        token.value, token.line, token.col
                    ),
                });
            }

            items.push(Self::join_tokens_for_raw(&self.tokens[start..self.pos]));

            if self.match_token(TokenType::Comma) {
                continue;
            }

            self.expect(TokenType::RParen)?;
            break;
        }

        Ok(items)
    }

    /// Helper for the dispatcher: consume one verb token (already known) and
    /// then capture the entire tail as a [`CommandStatement`].
    fn parse_command_kind(&mut self, kind: &str) -> Result<Statement> {
        self.advance(); // consume the verb token
        let body = self.consume_raw_to_statement_end();
        Ok(Statement::Command(CommandStatement {
            comments: vec![],
            kind: kind.to_string(),
            body,
        }))
    }

    /// `COMMENT ON {TABLE|COLUMN|...} <name> IS '...'` — preserved as raw.
    /// `COMMENT` can also appear inside `CREATE TABLE` column definitions and
    /// in other positions; only the standalone DDL form lands here because
    /// the dispatcher peeks at the *first* token.
    fn parse_comment_on_command(&mut self) -> Result<Statement> {
        // Look ahead for "COMMENT ON" — if not "ON", fall back to parser error
        // (the COMMENT token would otherwise have been consumed inside an
        // expression / column-def parser, not at statement boundary).
        if self.peek_offset(1).map(|t| t.value.to_uppercase()) != Some("ON".to_string()) {
            return Err(SqlglotError::UnexpectedToken {
                token: self.peek().clone(),
            });
        }
        self.advance(); // COMMENT
        let body = self.consume_raw_to_statement_end();
        Ok(Statement::Command(CommandStatement {
            comments: vec![],
            kind: "COMMENT".to_string(),
            body,
        }))
    }

    /// Returns `true` when the current Identifier token is a known
    /// statement-starting verb that we preserve verbatim.
    fn match_command_keyword(&self) -> bool {
        let v = self.peek().value.to_uppercase();
        matches!(
            v.as_str(),
            "GO" | "DECLARE"
                | "LOAD"
                | "REM"
                | "REMARK"
                | "RESET"
                | "PRAGMA"
                | "VACUUM"
                | "REINDEX"
                | "CALL"
                | "LOCK"
                | "UNLOCK"
                | "CLUSTER"
                | "REFRESH"
                | "CHECKPOINT"
                | "LISTEN"
                | "NOTIFY"
                | "PREPARE"
                | "EXECUTE"
                | "DEALLOCATE"
                | "DISCARD"
                | "COPY"
                | "ATTACH"
                | "DETACH"
                | "COMMENT"
                | "DESCRIBE"
                | "DESC"
                | "OPTIMIZE"
                | "SYSTEM"
                | "KILL"
                | "FLUSH"
                | "RESTORE"
                | "BACKUP"
                | "EXCHANGE"
                | "RENAME"
                | "WATCH"
                | "MSCK"
                | "UNLOAD"
                | "ASSERT"
                | "REPAIR"
                | "PURGE"
                | "ABORT"
                | "VALIDATE"
                | "MOVE"
                | "CLOSE"
                | "FETCH"
                | "REPLICATE"
                | "START"
                | "RAISE"
                | "UNDROP"
                | "EXCEPTION"
                | "CONNECT"
                | "DISCONNECT"
                | "SEND"
                | "ENABLE"
                | "DISABLE"
                | "REPLAY"
                | "SYNCHRONIZE"
                | "CHECK"
                | "REPORT"
                | "BIND"
                | "UNBIND"
                | "INCLUDE"
                | "EXPORT"
                | "IMPORT"
                | "ADMIN"
                | "SPLIT"
                | "TRACE"
                | "RESUME"
                | "SUSPEND"
                | "ROUTE"
                | "EMIT"
                | "FOR"
                | "WHILE"
                | "LOOP"
                | "RETURN"
                | "REPEAT"
                | "EXIT"
                | "LEAVE"
                | "ITERATE"
                | "CONTINUE"
                | "GOTO"
                | "RAISERROR"
                | "PRINT"
                | "WAITFOR"
                | "TRUNCATE"
                | "DO"
                | "CONNECTION"
                | "ELSEIF"
                | "ELSIF"
                | "UNTIL"
                | "CONNECT_BY_ROOT"
                | "APPLY"
                | "EXEC"
                | "OPEN"
                | "REVERT"
                | "DEALLOC"
                | "GRANT"
                | "REVOKE"
                | "DENY"
                | "UNSET"
                | "USE"
                | "PRELOAD"
                | "RECOMPRESS"
                | "COMPUTE"
                | "INVALIDATE"
                | "ANALYSE"
                | "BOOTSTRAP"
                | "LATCH"
                | "UNLATCH"
                | "SETOF"
                | "CHECKSUM"
                | "DELIMITER"
                | "GET"
                | "HELP"
                | "BINLOG"
                | "RELOAD"
                | "PARSE"
                | "BUFFER"
                | "BUILDS"
                | "COMPACT"
                | "FREEZE"
                | "UNFREEZE"
                | "BORROW"
                | "UNLISTEN"
                | "REPACK"
                | "RESIGNAL"
                | "SIGNAL"
                | "THROW"
                | "DBCC"
                | "SUMMARIZE"
                | "BATCH"
        )
    }

    /// Variant of [`parse_command_kind`] for verbs that arrive as an
    /// Identifier token (no dedicated TokenType).
    fn parse_command_from_identifier(&mut self) -> Result<Statement> {
        let verb = self.peek().value.to_uppercase();
        self.advance();
        let body = self.consume_raw_to_statement_end();
        Ok(Statement::Command(CommandStatement {
            comments: vec![],
            kind: verb,
            body,
        }))
    }

    /// Look at the token `offset` positions ahead of the current one,
    /// returning `None` if past EOF.
    fn peek_offset(&self, offset: usize) -> Option<&Token> {
        self.tokens.get(self.pos + offset)
    }

    /// Helper to check if current token is an identifier or keyword that can serve as a name.
    fn is_name_token(&self) -> bool {
        matches!(
            self.peek_type(),
            TokenType::Identifier
                | TokenType::All
                | TokenType::Year
                | TokenType::Month
                | TokenType::Day
                | TokenType::Hour
                | TokenType::Minute
                | TokenType::Second
                | TokenType::Interval
                | TokenType::Key
                | TokenType::Filter
                | TokenType::First
                | TokenType::Next
                | TokenType::Only
                | TokenType::Respect
                | TokenType::Epoch
                | TokenType::Schema
                | TokenType::Database
                | TokenType::View
                | TokenType::Collate
                | TokenType::Comment
                | TokenType::Left
                | TokenType::Right
                | TokenType::Replace
                | TokenType::Cube
                | TokenType::Rollup
                | TokenType::Grouping
                | TokenType::Pivot
                | TokenType::Unpivot
                | TokenType::Sets
                | TokenType::Range
                | TokenType::Conflict
                | TokenType::Unnest
                | TokenType::Text
                | TokenType::Show
                | TokenType::Describe
                | TokenType::Analyze
                | TokenType::Index
                | TokenType::Cast
                | TokenType::Group
                | TokenType::Order
                | TokenType::Explain
                | TokenType::Table
                | TokenType::Offset
                | TokenType::Merge
                | TokenType::Nulls
                | TokenType::Temp
                | TokenType::Temporary
                | TokenType::Rows
                | TokenType::Partition
                | TokenType::Any
                | TokenType::Escape
        )
    }

    /// Consume a name token (identifier or unreserved keyword used as identifier).
    fn expect_name(&mut self) -> Result<String> {
        let (name, _) = self.expect_name_with_quote()?;
        Ok(name)
    }

    /// If the current token is `@` / `:` / `Parameter` immediately followed by
    /// a name token (no whitespace tracking — they are adjacent in the token
    /// stream), consume both and return them as a combined alias name.
    /// Used to accept auto-generated aliases like `AS @rpm` or `AS :minutes`
    /// without changing parameter-marker handling elsewhere.
    fn try_parse_prefixed_alias(&mut self) -> Result<Option<(String, QuoteStyle)>> {
        let prefix = match self.peek_type() {
            TokenType::AtSign => '@',
            TokenType::Colon => ':',
            // Standalone Parameter token (`$` not absorbed into an identifier).
            TokenType::Parameter if self.peek().value == "$" => '$',
            _ => return Ok(None),
        };
        let next = match self.peek_offset(1) {
            Some(t) => t,
            None => return Ok(None),
        };
        let is_name_like = matches!(
            next.token_type,
            TokenType::Identifier
                | TokenType::Year
                | TokenType::Month
                | TokenType::Day
                | TokenType::Hour
                | TokenType::Minute
                | TokenType::Second
                | TokenType::Key
                | TokenType::Filter
                | TokenType::First
                | TokenType::Next
                | TokenType::Only
                | TokenType::Schema
                | TokenType::Database
                | TokenType::View
                | TokenType::Collate
                | TokenType::Comment
                | TokenType::Replace
                | TokenType::Text
                | TokenType::Show
                | TokenType::Describe
                | TokenType::Analyze
                | TokenType::Index
                | TokenType::Cast
                | TokenType::Group
                | TokenType::Order
                | TokenType::Range
        );
        if !is_name_like {
            return Ok(None);
        }
        self.advance(); // consume prefix
        let name_tok = self.advance().clone();
        let mut combined = String::with_capacity(name_tok.value.len() + 1);
        combined.push(prefix);
        combined.push_str(&name_tok.value);
        Ok(Some((combined, quote_style_from_char(name_tok.quote_char))))
    }

    /// Like `expect_name` but also returns the quote style of the token.
    fn expect_name_with_quote(&mut self) -> Result<(String, QuoteStyle)> {
        if self.is_name_token() {
            let token = self.advance().clone();
            let qs = quote_style_from_char(token.quote_char);
            let mut name = token.value.clone();
            // Append trailing `${...}` template variables so identifiers
            // like `t1_${type}` round-trip as a single name token.
            while matches!(self.peek_type(), TokenType::Parameter)
                && self.peek().value.starts_with("${")
            {
                name.push_str(&self.advance().value.clone());
            }
            return Ok((name, qs));
        }
        // Leading `${...}` template variable as a name (rare).
        if matches!(self.peek_type(), TokenType::Parameter) && self.peek().value.starts_with("${") {
            let mut name = self.advance().value.clone();
            // Only fuse plain identifiers or further `${...}` segments —
            // never reserved keywords (Order, By, etc.) even though those
            // tokenize as name-like, or the template would swallow the
            // surrounding clause.
            while matches!(self.peek_type(), TokenType::Identifier)
                || (matches!(self.peek_type(), TokenType::Parameter)
                    && self.peek().value.starts_with("${"))
            {
                name.push_str(&self.advance().value.clone());
            }
            return Ok((name, QuoteStyle::None));
        }
        // ClickHouse typed placeholder used as an identifier:
        // `{db:Identifier}`, `{tbl:Identifier}`. Accept anywhere a name is
        // expected so `FROM {db:Identifier}.t` and friends parse.
        if matches!(self.peek_type(), TokenType::Parameter) && self.peek().value.starts_with('{') {
            let name = self.advance().value.clone();
            return Ok((name, QuoteStyle::None));
        }
        // Also accept any keyword-like identifier
        let token = self.peek().clone();
        if matches!(
            token.token_type,
            TokenType::Identifier
                | TokenType::Int
                | TokenType::Integer
                | TokenType::BigInt
                | TokenType::SmallInt
                | TokenType::TinyInt
                | TokenType::Float
                | TokenType::Double
                | TokenType::Decimal
                | TokenType::Numeric
                | TokenType::Real
                | TokenType::Varchar
                | TokenType::Char
                | TokenType::Text
                | TokenType::Boolean
                | TokenType::Date
                | TokenType::Timestamp
                | TokenType::TimestampTz
                | TokenType::Time
                | TokenType::Interval
                | TokenType::Blob
                | TokenType::Bytea
                | TokenType::Json
                | TokenType::Jsonb
                | TokenType::Uuid
                | TokenType::Array
                | TokenType::Map
                | TokenType::Struct
                | TokenType::Offset
                | TokenType::Limit
                | TokenType::Default
                | TokenType::Begin
                | TokenType::Recursive
                | TokenType::Ignore
                | TokenType::Pivot
                | TokenType::Unpivot
                | TokenType::Rows
                | TokenType::Range
                | TokenType::Values
        ) {
            let t = self.advance().clone();
            let qs = quote_style_from_char(t.quote_char);
            Ok((t.value.clone(), qs))
        } else {
            Err(SqlglotError::ParserError {
                message: format!(
                    "Expected identifier, got {:?} ('{}') at line {} col {}",
                    token.token_type, token.value, token.line, token.col
                ),
            })
        }
    }

    // ── Top-level parsing ──────────────────────────────────────────

    /// Parse a single SQL statement.
    pub fn parse_statement(&mut self) -> Result<Statement> {
        self.collect_comments();
        let stmt = self.parse_statement_inner()?;
        // ClickHouse trailing `WITH TOTALS` / `WITH TIES` / `WITH ROLLUP` /
        // `WITH CUBE` postfix at the end of a SELECT — these are query-level
        // modifiers we don't model; swallow them so the statement closes.
        if matches!(self.peek_type(), TokenType::With) {
            let after = self.peek_offset(1);
            let is_postfix_modifier = after
                .map(|t| {
                    matches!(
                        t.token_type,
                        TokenType::Identifier | TokenType::Cube | TokenType::Rollup
                    ) && matches!(
                        t.value.to_uppercase().as_str(),
                        "TOTALS" | "TIES" | "FILL" | "ROLLUP" | "CUBE"
                    )
                })
                .unwrap_or(false);
            if is_postfix_modifier {
                self.advance();
                self.advance();
                // Swallow any chained option words up to `;`/EOF/FORMAT/SETTINGS.
                while !matches!(self.peek_type(), TokenType::Semicolon | TokenType::Eof) {
                    if self.is_name_token()
                        && matches!(
                            self.peek().value.to_uppercase().as_str(),
                            "SETTINGS" | "FORMAT"
                        )
                    {
                        break;
                    }
                    self.advance();
                }
            }
        }
        // ClickHouse trailing `SETTINGS k=v, k=v` clause / `FORMAT name`
        // (statement-level). Swallow up to the next `;` or EOF.
        if self.is_name_token()
            && matches!(
                self.peek().value.to_uppercase().as_str(),
                "SETTINGS" | "FORMAT"
            )
        {
            while !matches!(self.peek_type(), TokenType::Semicolon | TokenType::Eof) {
                self.advance();
            }
        }
        // BigQuery pipe-syntax: `<query> |> WHERE … |> AGGREGATE … |> …`.
        // The `|>` operator chains query stages. We don't model them; swallow
        // the entire chain to end of statement so the leading query stands.
        if self.peek_type() == &TokenType::BitwiseOr
            && self
                .peek_offset(1)
                .map(|t| matches!(t.token_type, TokenType::Gt))
                .unwrap_or(false)
        {
            while !matches!(self.peek_type(), TokenType::Semicolon | TokenType::Eof) {
                self.advance();
            }
        }
        // Consume trailing semicolons
        while self.match_token(TokenType::Semicolon) {}
        Ok(stmt)
    }

    fn parse_statement_inner(&mut self) -> Result<Statement> {
        self.collect_comments();
        let comments = self.take_comments();
        // MySQL / PSM labeled block: `mylabel: BEGIN … END mylabel`.
        // Swallow the leading `<name>:` so the block dispatches normally.
        if self.is_name_token()
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::Colon)
            )
        {
            let saved = self.pos;
            self.advance();
            self.advance();
            // Only treat as a label if a known block keyword follows;
            // otherwise rewind so we don't misinterpret `alias: type`.
            let is_block = matches!(
                self.peek_type(),
                TokenType::Begin | TokenType::If | TokenType::Case
            ) || self.check_keyword("WHILE")
                || self.check_keyword("LOOP")
                || self.check_keyword("FOR")
                || self.check_keyword("REPEAT");
            if !is_block {
                self.pos = saved;
            }
        }
        let mut stmt = match self.peek_type() {
            TokenType::With => self.parse_with_statement(),
            TokenType::Select => {
                let select = self.parse_select_body(vec![])?;
                self.maybe_parse_set_operation(Statement::Select(select))
            }
            TokenType::LParen => {
                // Could be a parenthesized SELECT / VALUES / TABLE form.
                let saved_pos = self.pos;
                self.advance(); // consume '('
                if matches!(
                    self.peek_type(),
                    TokenType::Select
                        | TokenType::With
                        | TokenType::From
                        | TokenType::Values
                        | TokenType::Table
                        | TokenType::LParen
                ) {
                    let inner = self.parse_statement_inner()?;
                    self.expect(TokenType::RParen)?;
                    self.maybe_parse_set_operation(inner)
                } else {
                    self.pos = saved_pos;
                    Err(SqlglotError::ParserError {
                        message: "Expected statement".into(),
                    })
                }
            }
            TokenType::Insert => self.parse_insert().map(Statement::Insert),
            TokenType::Replace => self.parse_insert().map(Statement::Insert),
            TokenType::Update => self.parse_update().map(Statement::Update),
            TokenType::Delete => self.parse_delete().map(Statement::Delete),
            TokenType::Merge => self.parse_merge().map(Statement::Merge),
            TokenType::Create => self.parse_create_or_command(),
            TokenType::Drop => self.parse_drop(),
            TokenType::Alter => self.parse_alter_or_command(),
            TokenType::Truncate => {
                let saved = self.pos;
                match self.parse_truncate() {
                    Ok(t) => {
                        // Tolerate Oracle-flavored trailing modifiers on
                        // TRUNCATE (PURGE, DROP STORAGE, REUSE STORAGE,
                        // KEEP …, CASCADE, etc.) by swallowing all trailing
                        // tokens up to the statement boundary.
                        while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon) {
                            self.advance();
                        }
                        Ok(Statement::Truncate(t))
                    }
                    Err(_) => {
                        self.pos = saved;
                        self.parse_command_kind("TRUNCATE")
                    }
                }
            }
            TokenType::Begin | TokenType::Commit | TokenType::Rollback | TokenType::Savepoint => {
                // PL/pgSQL / MySQL stored-procedure block: `BEGIN <stmt> …
                // END`. If `BEGIN` is followed by anything that isn't an
                // obvious transaction modifier, capture the whole block as
                // a command so the surrounding parse completes.
                if matches!(self.peek_type(), TokenType::Begin) {
                    let next = self.peek_offset(1).map(|t| &t.token_type);
                    let is_psm_block = matches!(
                        next,
                        Some(TokenType::Identifier)
                            | Some(TokenType::If)
                            | Some(TokenType::Case)
                            | Some(TokenType::Select)
                            | Some(TokenType::Insert)
                            | Some(TokenType::Update)
                            | Some(TokenType::Delete)
                    );
                    if is_psm_block {
                        return self.parse_command_kind("BEGIN");
                    }
                }
                self.parse_transaction().map(Statement::Transaction)
            }
            TokenType::Explain => self.parse_explain().map(Statement::Explain),
            TokenType::Use => self.parse_use().map(Statement::Use),
            // Raw-tail command statements: SET / SHOW / DESCRIBE / ANALYZE
            // (when standalone, not as part of EXPLAIN) / COMMENT ON ... .
            // We preserve the verb plus the entire remainder up to `;` or EOF
            // so the AST round-trips even though we don't model these in detail.
            TokenType::Set => self.parse_command_kind("SET"),
            TokenType::Show => self.parse_command_kind("SHOW"),
            TokenType::Describe => self.parse_command_kind("DESCRIBE"),
            // `DESC <name>` is a Hive/MySQL synonym for DESCRIBE. The lone
            // `Desc` token also appears mid-statement (ORDER BY x DESC), so
            // we only treat it as a statement when at the very start.
            TokenType::Desc => self.parse_command_kind("DESC"),
            // Hive multi-insert: `FROM tbl INSERT OVERWRITE TABLE x SELECT ...`
            // [INSERT OVERWRITE TABLE y SELECT ...]+. Capture the whole thing
            // as a raw command body so it round-trips.
            TokenType::From => {
                // Hive `FROM tbl INSERT OVERWRITE TABLE x …` / `FROM tbl
                // SELECT cols`. DuckDB implicit SELECT: `FROM tbl …`. Try
                // the structured DuckDB FROM-first parse only when there is
                // no INSERT/SELECT marker at the top paren level; otherwise
                // capture as a raw command so it round-trips. Fall back to
                // command capture on parse failure as well.
                let mut i = self.pos + 1;
                let mut depth = 0i32;
                let mut hive = false;
                while i < self.tokens.len() {
                    match &self.tokens[i].token_type {
                        TokenType::Eof | TokenType::Semicolon => break,
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            if depth == 0 {
                                break;
                            }
                            depth -= 1;
                        }
                        TokenType::Insert | TokenType::Select if depth == 0 => {
                            hive = true;
                            break;
                        }
                        _ => {}
                    }
                    i += 1;
                }
                if hive {
                    self.parse_command_kind("FROM")
                } else {
                    let saved_from = self.pos;
                    match self.parse_select_body(vec![]) {
                        Ok(select) => self.maybe_parse_set_operation(Statement::Select(select)),
                        Err(_) => {
                            self.pos = saved_from;
                            self.parse_command_kind("FROM")
                        }
                    }
                }
            }
            TokenType::Analyze => self.parse_command_kind("ANALYZE"),
            TokenType::Check => self.parse_command_kind("CHECK"),
            TokenType::Comment => self.parse_comment_on_command(),
            TokenType::Grant => self.parse_command_kind("GRANT"),
            TokenType::Revoke => self.parse_command_kind("REVOKE"),
            // Procedural / control-flow statements (Spark, MySQL stored
            // procs, PL/SQL, T-SQL): IF / FOR / WHILE / LOOP / CASE blocks
            // and the matching ELSE / END / WHEN tokens at statement start.
            // Capture verbatim so the AST round-trips.
            TokenType::If => self.parse_command_kind("IF"),
            TokenType::Else => self.parse_command_kind("ELSE"),
            TokenType::End => self.parse_command_kind("END"),
            TokenType::Case => self.parse_command_kind("CASE"),
            TokenType::When => self.parse_command_kind("WHEN"),
            TokenType::Then => self.parse_command_kind("THEN"),
            TokenType::Do => self.parse_command_kind("DO"),
            // Spark: `TABLE name` and `TABLE name |> …` are SELECT-equivalent
            // shorthand. Capture verbatim so the AST round-trips.
            TokenType::Table => self.parse_command_kind("TABLE"),
            TokenType::Values => self.parse_command_kind("VALUES"),
            // DuckDB SQL-shorthand: `PIVOT tbl ON col USING agg(...)` and
            // `UNPIVOT tbl ON col INTO ...`. Preserve verbatim.
            TokenType::Pivot => self.parse_command_kind("PIVOT"),
            TokenType::Unpivot => self.parse_command_kind("UNPIVOT"),
            // PG cursor verbs: FETCH, MOVE, CLOSE.
            TokenType::Fetch => self.parse_command_kind("FETCH"),
            // Vendor-specific verbs that tokenize as plain identifiers:
            //   GO (T-SQL batch separator), DECLARE (T-SQL/PL-pgSQL),
            //   LOAD (PG / MySQL extensions), REM / REMARK (SQL*Plus),
            //   RESET / PRAGMA / VACUUM / REINDEX (PG / SQLite), CALL (PSM).
            TokenType::Identifier if self.match_command_keyword() => {
                self.parse_command_from_identifier()
            }
            // PL/pgSQL / MySQL stored-procedure assignment `var := expr` or
            // `var = expr` at statement position. Preserve verbatim.
            TokenType::Identifier
                if matches!(
                    self.peek_offset(1).map(|t| &t.token_type),
                    Some(TokenType::Colon)
                ) && matches!(
                    self.peek_offset(2).map(|t| &t.token_type),
                    Some(TokenType::Eq)
                ) =>
            {
                self.parse_command_kind("ASSIGN")
            }
            // PL/SQL / PL/pgSQL variable declaration at top level:
            //   `name TYPE [:= default]`. Some corpora split DECLARE blocks
            //   into individual lines; treat these as opaque commands.
            //   Heuristic: <identifier> followed by either a data-type
            //   token, or an identifier that looks type-like (uppercase
            //   keyword such as NUMBER/VARCHAR2/BOOLEAN/PLS_INTEGER/etc.).
            TokenType::Identifier
                if self
                    .peek_offset(1)
                    .map(|t| {
                        self.is_data_type_token_kind(&t.token_type)
                            || (matches!(t.token_type, TokenType::Identifier)
                                && matches!(
                                    t.value.to_uppercase().as_str(),
                                    "NUMBER"
                                        | "VARCHAR2"
                                        | "NVARCHAR2"
                                        | "PLS_INTEGER"
                                        | "BINARY_INTEGER"
                                        | "ROWID"
                                        | "UROWID"
                                        | "CLOB"
                                        | "NCLOB"
                                        | "BFILE"
                                        | "LONG"
                                        | "RAW"
                                        | "XMLTYPE"
                                        | "RECORD"
                                ))
                            || matches!(t.token_type, TokenType::Percent | TokenType::Percent2)
                    })
                    .unwrap_or(false)
                    && self
                        .peek_offset(2)
                        .map(|t| {
                            // Confirm declaration shape: trailing `:=`,
                            // `%TYPE`/`%ROWTYPE`, semicolon, EOF, or
                            // `(precision)` parenthesised type modifier.
                            matches!(
                                t.token_type,
                                TokenType::Colon
                                    | TokenType::Semicolon
                                    | TokenType::Eof
                                    | TokenType::Percent
                                    | TokenType::Percent2
                                    | TokenType::LParen
                            ) || matches!(t.token_type, TokenType::Identifier)
                                && matches!(
                                    t.value.to_uppercase().as_str(),
                                    "NOT" | "DEFAULT" | "CONSTANT"
                                )
                        })
                        .unwrap_or(true) =>
            {
                self.parse_command_kind("PLSQL_DECL")
            }
            _ => Err(SqlglotError::UnexpectedToken {
                token: self.peek().clone(),
            }),
        }?;
        if !comments.is_empty() {
            attach_comments_to_statement(&mut stmt, comments);
        }
        Ok(stmt)
    }

    /// Parse multiple statements separated by semicolons.
    pub fn parse_statements(&mut self) -> Result<Vec<Statement>> {
        let mut stmts = Vec::new();
        while !matches!(self.peek_type(), TokenType::Eof) {
            while self.match_token(TokenType::Semicolon) {}
            if matches!(self.peek_type(), TokenType::Eof) {
                break;
            }
            stmts.push(self.parse_statement()?);
            // ClickHouse trailing `FORMAT <name>` after a statement is a
            // client-side output directive, not part of the AST. Swallow
            // it (and any whitespace-separated payload up to the next
            // semicolon / EOF) so the statement still parses.
            if self.peek().value.eq_ignore_ascii_case("FORMAT") {
                let saved = self.pos;
                self.advance();
                if self.is_name_token() {
                    self.advance();
                    while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon) {
                        self.advance();
                    }
                } else {
                    self.pos = saved;
                }
            }
        }
        Ok(stmts)
    }

    // ── WITH / CTE parsing ─────────────────────────────────────────

    fn parse_with_statement(&mut self) -> Result<Statement> {
        self.expect(TokenType::With)?;
        let recursive = self.match_token(TokenType::Recursive);

        // T-SQL `WITH XMLNAMESPACES ('uri' AS prefix [, ...]) <stmt>`. The
        // XML namespaces are not modeled in the AST; swallow the keyword
        // and its parenthesized binding list opaquely so the surrounding
        // SELECT / INSERT / UPDATE / DELETE / MERGE parses cleanly.
        if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("XMLNAMESPACES") {
            self.advance(); // XMLNAMESPACES
            if self.match_token(TokenType::LParen) {
                let mut depth = 1_i32;
                while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth -= 1,
                        _ => {}
                    }
                    self.advance();
                }
            }
            return self.parse_with_body(vec![]);
        }

        // ClickHouse scalar-binding form: `WITH (expr) AS name [, ...] SELECT …`
        // (and the symmetric `WITH expr AS name`). Detect by peeking for a
        // `<expr> AS <name>` pattern rather than the canonical `<name> AS
        // (select …)`. We swallow these bindings — they aren't modeled as
        // CTEs — then fall through to the main query.
        if self.is_clickhouse_scalar_with() {
            loop {
                let _ = self.parse_expr()?;
                self.expect(TokenType::As)?;
                // The binding name may use a data-type keyword (`Uuid`,
                // `Text`, etc.) — accept any single token that isn't a
                // structural delimiter so the loop advances.
                if self.is_name_token() || self.is_data_type_token() {
                    self.advance();
                } else if !matches!(
                    self.peek_type(),
                    TokenType::Comma
                        | TokenType::Eof
                        | TokenType::Semicolon
                        | TokenType::Select
                        | TokenType::Insert
                        | TokenType::Update
                        | TokenType::Delete
                        | TokenType::Merge
                ) {
                    self.advance();
                }
                if !self.match_token(TokenType::Comma) {
                    break;
                }
                // The next binding might still be `name AS (select …)`; if so,
                // fall back to the canonical CTE parser for the remainder.
                if !self.is_clickhouse_scalar_with() {
                    let mut ctes = vec![self.parse_cte(recursive)?];
                    while self.match_token(TokenType::Comma) {
                        ctes.push(self.parse_cte(recursive)?);
                    }
                    return self.parse_with_body(ctes);
                }
            }
            return self.parse_with_body(vec![]);
        }

        let mut ctes = vec![self.parse_cte(recursive)?];
        while self.match_token(TokenType::Comma) {
            ctes.push(self.parse_cte(recursive)?);
        }
        // PostgreSQL recursive-query SEARCH / CYCLE clauses appear between
        // the last CTE and the main query body. Swallow them opaquely.
        // Forms:
        //   SEARCH { DEPTH | BREADTH } FIRST BY <col_list> SET <col>
        //   CYCLE <col_list> SET <col> [TO <val> DEFAULT <val>] USING <col>
        loop {
            let saved = self.pos;
            if self.match_keyword("SEARCH") {
                let _ = self.match_keyword("DEPTH") || self.match_keyword("BREADTH");
                let _ = self.match_keyword("FIRST");
                let _ = self.match_token(TokenType::By);
                // Swallow tokens until SET or end-of-search clause.
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    && !self.check_keyword("SET")
                {
                    self.advance();
                }
                if self.match_keyword("SET") {
                    let _ = self.is_name_token() && {
                        self.advance();
                        true
                    };
                }
                continue;
            }
            if self.check_keyword("CYCLE") {
                self.advance();
                while !matches!(
                    self.peek_type(),
                    TokenType::Select
                        | TokenType::Insert
                        | TokenType::Update
                        | TokenType::Delete
                        | TokenType::Merge
                        | TokenType::With
                        | TokenType::Eof
                        | TokenType::Semicolon
                ) {
                    self.advance();
                }
                continue;
            }
            self.pos = saved;
            break;
        }
        self.parse_with_body(ctes)
    }

    /// Returns true if the current token sequence looks like a ClickHouse
    /// scalar `WITH expr AS name` rather than a canonical `name AS (select …)`
    /// CTE binding. Used by [`parse_with_statement`] to switch parsing modes.
    fn is_clickhouse_scalar_with(&self) -> bool {
        // Canonical CTE binding starts with `<name>` then either `(` (column
        // list) or `AS`. Anything else — a parenthesized expression, a number,
        // a string, a function call, an operator — must be the scalar form.
        match self.peek_type() {
            TokenType::LParen => true,
            TokenType::LBracket => true,
            TokenType::Number | TokenType::String | TokenType::HexString => true,
            t if matches!(t, TokenType::Minus | TokenType::Plus) => true,
            _ => {
                // Plain identifier followed by anything other than `(` or `AS`
                // also indicates the scalar form (e.g. `WITH x + 1 AS y`).
                if self.is_name_token() {
                    let next = self.peek_offset(1).map(|t| &t.token_type);
                    match next {
                        Some(TokenType::LParen) => {
                            // `name(...)` is canonical column-list form only
                            // if the body is a `name [, name]*` followed by
                            // `) AS`. Otherwise (function call like
                            // `arrayJoin([...])`) it's the scalar form.
                            !self.parens_are_name_list_then_as(1)
                        }
                        Some(TokenType::As) => false,
                        _ => true,
                    }
                } else {
                    false
                }
            }
        }
    }

    /// Starting at `tokens[self.pos + offset]` (which must be `(`), check
    /// whether the body is a comma-separated identifier list followed by
    /// `)` and then `AS` — the shape of a CTE column-list binding.
    fn parens_are_name_list_then_as(&self, offset: usize) -> bool {
        let mut i = self.pos + offset;
        if self.tokens.get(i).map(|t| &t.token_type) != Some(&TokenType::LParen) {
            return false;
        }
        i += 1;
        loop {
            // Accept any name-like token in the column list, not just plain
            // identifiers — DuckDB CTEs frequently use unreserved keywords
            // like `key`, `value`, `order`, `range` as column names.
            let is_name_like = matches!(
                self.tokens.get(i).map(|t| &t.token_type),
                Some(TokenType::Identifier)
                    | Some(TokenType::Key)
                    | Some(TokenType::Year)
                    | Some(TokenType::Month)
                    | Some(TokenType::Day)
                    | Some(TokenType::Hour)
                    | Some(TokenType::Minute)
                    | Some(TokenType::Second)
                    | Some(TokenType::Filter)
                    | Some(TokenType::First)
                    | Some(TokenType::Next)
                    | Some(TokenType::Only)
                    | Some(TokenType::Schema)
                    | Some(TokenType::Database)
                    | Some(TokenType::View)
                    | Some(TokenType::Collate)
                    | Some(TokenType::Comment)
                    | Some(TokenType::Replace)
                    | Some(TokenType::Text)
                    | Some(TokenType::Show)
                    | Some(TokenType::Describe)
                    | Some(TokenType::Analyze)
                    | Some(TokenType::Index)
                    | Some(TokenType::Cast)
                    | Some(TokenType::Group)
                    | Some(TokenType::Order)
                    | Some(TokenType::Range)
                    | Some(TokenType::Partition)
                    | Some(TokenType::Rows)
                    | Some(TokenType::Table)
                    | Some(TokenType::Offset)
                    | Some(TokenType::Temp)
                    | Some(TokenType::Temporary)
                    | Some(TokenType::Nulls)
                    | Some(TokenType::Conflict)
                    | Some(TokenType::Unnest)
                    | Some(TokenType::Explain)
                    | Some(TokenType::Merge)
                    | Some(TokenType::Any)
                    | Some(TokenType::Escape)
            );
            if is_name_like {
                i += 1;
            } else {
                return false;
            }
            match self.tokens.get(i).map(|t| &t.token_type) {
                Some(TokenType::Comma) => i += 1,
                Some(TokenType::RParen) => {
                    i += 1;
                    // DuckDB recursive cycle clause: `(cols) USING KEY (...)
                    // AS (...)`. Treat the cycle keyword as a sign this is a
                    // canonical CTE binding, not a ClickHouse scalar.
                    if self.tokens.get(i).map(|t| t.value.to_uppercase())
                        == Some("USING".to_string())
                    {
                        return true;
                    }
                    if self.tokens.get(i).map(|t| &t.token_type) != Some(&TokenType::As) {
                        return false;
                    }
                    // Canonical form requires the body after `AS` to be
                    // a parenthesized SELECT (or `[NOT] MATERIALIZED (…)`
                    // for DuckDB / PostgreSQL). If it isn't, this is the
                    // ClickHouse scalar form.
                    i += 1;
                    let after_as = self.tokens.get(i).map(|t| &t.token_type);
                    if after_as == Some(&TokenType::LParen) {
                        return true;
                    }
                    let after_as_value = self.tokens.get(i).map(|t| t.value.as_str());
                    if matches!(
                        after_as_value,
                        Some(v) if v.eq_ignore_ascii_case("MATERIALIZED")
                            || v.eq_ignore_ascii_case("NOT")
                    ) {
                        return true;
                    }
                    return false;
                }
                _ => return false,
            }
        }
    }

    fn parse_with_body(&mut self, ctes: Vec<Cte>) -> Result<Statement> {
        match self.peek_type() {
            TokenType::Select => {
                let select = self.parse_select_body(ctes)?;
                self.maybe_parse_set_operation(Statement::Select(select))
            }
            // DuckDB `WITH x AS (...) FROM tbl SELECT cols` (FROM-first form).
            // We rely on parse_select_body's existing FROM-first tolerance.
            TokenType::From => {
                let select = self.parse_select_body(ctes)?;
                self.maybe_parse_set_operation(Statement::Select(select))
            }
            // PostgreSQL / DuckDB `WITH x AS (...) TABLE tbl` body — equivalent
            // to `SELECT * FROM tbl`. Swallow the table reference and trailing
            // clauses opaquely and emit a stub Select so the surrounding
            // statement parses cleanly.
            // DuckDB / PostgreSQL `TABLE tbl` as the body of a WITH query —
            // shorthand for `SELECT * FROM tbl`. Swallow the trailing tokens
            // opaquely and emit a stub Select so the surrounding parse runs.
            TokenType::Table => {
                self.advance();
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon) {
                    self.advance();
                }
                let select = SelectStatement {
                    comments: vec![],
                    ctes,
                    distinct: false,
                    top: None,
                    columns: vec![SelectItem::Wildcard],
                    from: None,
                    joins: vec![],
                    where_clause: None,
                    group_by: vec![],
                    having: None,
                    order_by: vec![],
                    limit: None,
                    offset: None,
                    fetch_first: None,
                    qualify: None,
                    window_definitions: vec![],
                };
                Ok(Statement::Select(select))
            }
            TokenType::Insert => {
                let ins = self.parse_insert()?;
                let _ = ctes;
                Ok(Statement::Insert(ins))
            }
            TokenType::Update => {
                let upd = self.parse_update()?;
                let _ = ctes;
                Ok(Statement::Update(upd))
            }
            TokenType::Delete => {
                let del = self.parse_delete()?;
                let _ = ctes;
                Ok(Statement::Delete(del))
            }
            TokenType::Merge => {
                let mrg = self.parse_merge()?;
                let _ = ctes;
                Ok(Statement::Merge(mrg))
            }
            _ => Err(SqlglotError::ParserError {
                message: "Expected SELECT or INSERT after WITH clause".into(),
            }),
        }
    }

    fn parse_cte(&mut self, recursive: bool) -> Result<Cte> {
        let (name, name_quote_style) = self.expect_name_with_quote()?;

        let columns = if self.match_token(TokenType::LParen) {
            let mut cols = vec![self.expect_name()?];
            while self.match_token(TokenType::Comma) {
                cols.push(self.expect_name()?);
            }
            self.expect(TokenType::RParen)?;
            cols
        } else {
            vec![]
        };

        // DuckDB recursive CTE cycle clause:
        //   `WITH RECURSIVE tbl(a, b) USING KEY (a, max(b)) AS (...)`.
        // Swallow `USING KEY (...)` opaquely so the surrounding parse runs.
        if self.check_keyword("USING") {
            let saved = self.pos;
            self.advance();
            if self.check_keyword("KEY") {
                self.advance();
                if self.match_token(TokenType::LParen) {
                    let mut depth = 1_i32;
                    while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                        match self.peek_type() {
                            TokenType::LParen => depth += 1,
                            TokenType::RParen => depth -= 1,
                            _ => {}
                        }
                        self.advance();
                    }
                }
            } else {
                self.pos = saved;
            }
        }

        self.expect(TokenType::As)?;
        let materialized = if self.match_keyword("MATERIALIZED") {
            Some(true)
        } else if self.check_keyword("NOT") {
            let saved = self.pos;
            self.advance();
            if self.match_keyword("MATERIALIZED") {
                Some(false)
            } else {
                self.pos = saved;
                None
            }
        } else {
            None
        };

        self.expect(TokenType::LParen)?;
        let query = self.parse_statement_inner()?;
        self.expect(TokenType::RParen)?;

        Ok(Cte {
            name,
            name_quote_style,
            columns,
            query: Box::new(query),
            materialized,
            recursive,
        })
    }

    // ── SELECT ──────────────────────────────────────────────────────

    fn parse_select_body(&mut self, ctes: Vec<Cte>) -> Result<SelectStatement> {
        // DuckDB allows starting a query with `FROM ...` and implies
        // `SELECT *`. Detect that and synthesise the wildcard projection.
        let from_first = !matches!(self.peek_type(), TokenType::Select)
            && matches!(self.peek_type(), TokenType::From);
        if !from_first {
            self.expect(TokenType::Select)?;
        }

        // MySQL `SELECT` modifiers (between SELECT and the column list):
        // DISTINCTROW (alias of DISTINCT), HIGH_PRIORITY, STRAIGHT_JOIN,
        // SQL_SMALL_RESULT, SQL_BIG_RESULT, SQL_BUFFER_RESULT, SQL_CACHE /
        // SQL_NO_CACHE, SQL_CALC_FOUND_ROWS. Swallow any number of these.
        let mut distinctrow = false;
        loop {
            if self.is_name_token() {
                let v = self.peek().value.to_uppercase();
                if matches!(
                    v.as_str(),
                    "DISTINCTROW"
                        | "HIGH_PRIORITY"
                        | "STRAIGHT_JOIN"
                        | "SQL_SMALL_RESULT"
                        | "SQL_BIG_RESULT"
                        | "SQL_BUFFER_RESULT"
                        | "SQL_CACHE"
                        | "SQL_NO_CACHE"
                        | "SQL_CALC_FOUND_ROWS"
                ) {
                    if v == "DISTINCTROW" {
                        distinctrow = true;
                    }
                    self.advance();
                    continue;
                }
            }
            break;
        }
        let distinct = distinctrow || self.match_token(TokenType::Distinct);
        // PostgreSQL / DuckDB `DISTINCT ON (expr, ...)` — swallow the column
        // list so the surrounding query parses. We don't model DISTINCT ON in
        // the AST; treat it as plain DISTINCT.
        if distinct && self.match_token(TokenType::On) {
            self.expect(TokenType::LParen)?;
            let mut depth = 1;
            while depth > 0 {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    TokenType::Eof => break,
                    _ => {}
                }
                self.advance();
            }
        }
        // SQL-standard `SELECT ALL` quantifier (§7.12). Equivalent to omitting
        // the quantifier; consume it so it does not get mis-parsed as a column.
        if !distinct {
            let _ = self.match_token(TokenType::All);
        }

        // BigQuery `SELECT [DISTINCT] AS STRUCT|VALUE …` — type-tag for the
        // implicit row constructor. We don't model it; swallow the prefix.
        if self.peek_type() == &TokenType::As {
            let v = self
                .peek_offset(1)
                .map(|t| t.value.to_uppercase())
                .unwrap_or_default();
            if matches!(v.as_str(), "STRUCT" | "VALUE") {
                self.advance(); // AS
                self.advance(); // STRUCT|VALUE
            }
        }

        // TOP N (SQL Server style)
        // Use parse_primary() instead of parse_expr() to prevent the parser
        // from consuming `*` (SELECT all columns) as a multiplication operator.
        // This correctly handles: TOP 5, TOP 100, TOP (expr), TOP (@var)
        let top = if self.match_token(TokenType::Top) {
            Some(Box::new(self.parse_primary()?))
        } else {
            None
        };

        let columns = if from_first {
            vec![SelectItem::Wildcard]
        } else {
            self.parse_select_items()?
        };

        let from = if self.match_token(TokenType::From) {
            Some(FromClause {
                source: self.parse_table_source()?,
            })
        } else {
            None
        };

        let joins = self.parse_joins()?;

        // ClickHouse `PREWHERE expr` hint clause (sits between FROM/joins and
        // WHERE). Parsed as a regular boolean expression and folded into the
        // WHERE clause via `AND` so the AST stays simple.
        let prewhere = if self.check_keyword("PREWHERE") {
            self.advance();
            Some(self.parse_expr()?)
        } else {
            None
        };

        let where_clause = if self.match_token(TokenType::Where) {
            let e = self.parse_expr()?;
            // ClickHouse: `WHERE (expr) AS alias` — alias-binds the
            // predicate. Swallow the AS-alias tail; we don't model it.
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            Some(e)
        } else {
            None
        };

        let where_clause = match (prewhere, where_clause) {
            (Some(pw), Some(w)) => Some(Expr::BinaryOp {
                left: Box::new(pw),
                op: BinaryOperator::And,
                right: Box::new(w),
            }),
            (Some(pw), None) => Some(pw),
            (None, w) => w,
        };

        // Teradata `PREFERRING <expr> [PARTITION BY <list>]` skyline clause.
        // Sits between WHERE and GROUP BY. Swallow opaquely up to a known
        // terminator so the surrounding query parses.
        if self.check_keyword("PREFERRING") {
            self.advance();
            loop {
                match self.peek_type() {
                    TokenType::Eof
                    | TokenType::Semicolon
                    | TokenType::Group
                    | TokenType::Order
                    | TokenType::Having
                    | TokenType::Qualify
                    | TokenType::Limit
                    | TokenType::Union
                    | TokenType::Intersect
                    | TokenType::Except
                    | TokenType::RParen => break,
                    _ => {}
                }
                self.advance();
            }
        }

        let group_by = if self.match_token(TokenType::Group) {
            self.expect(TokenType::By)?;
            let items = self.parse_group_by_list()?;
            // ClickHouse / MySQL `GROUP BY ... WITH ROLLUP|CUBE|TOTALS` —
            // swallow the modifier; we don't model it in the AST.
            if self.match_token(TokenType::With) {
                let _ = self.match_token(TokenType::Rollup)
                    || self.match_token(TokenType::Cube)
                    || self.match_keyword("TOTALS");
            }
            // Hive / Spark `GROUP BY k1, k2 GROUPING SETS ((k1), (k2))` —
            // swallow the trailing parenthesized list.
            if self.match_token(TokenType::Grouping) {
                if self.check_keyword("SETS") {
                    self.advance();
                }
                if self.match_token(TokenType::LParen) {
                    let mut depth = 1;
                    while depth > 0 {
                        match self.peek_type() {
                            TokenType::LParen => depth += 1,
                            TokenType::RParen => {
                                depth -= 1;
                                if depth == 0 {
                                    self.advance();
                                    break;
                                }
                            }
                            TokenType::Eof => break,
                            _ => {}
                        }
                        self.advance();
                    }
                }
            }
            items
        } else {
            vec![]
        };

        let having = if self.match_token(TokenType::Having) {
            let expr = self.parse_expr()?;
            // ClickHouse corpora occasionally include a trailing alias after
            // HAVING expression text (`HAVING cond AS x`). Swallow alias so it
            // doesn't leak as an unexpected token.
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            Some(expr)
        } else {
            None
        };

        let qualify = if self.match_token(TokenType::Qualify) {
            Some(self.parse_expr()?)
        } else {
            None
        };

        // Named WINDOW definitions
        let window_definitions = if self.match_token(TokenType::Window) {
            self.parse_window_definitions()?
        } else {
            vec![]
        };

        let order_by = if self.match_token(TokenType::Order) {
            self.expect(TokenType::By)?;
            self.parse_order_by_items()?
        } else {
            vec![]
        };

        // Hive / Spark non-standard ordering clauses; behave syntactically
        // like ORDER BY. We parse and discard them so the surrounding query
        // continues to parse.
        loop {
            let is_sort = self.check_keyword("SORT");
            let is_distribute = self.check_keyword("DISTRIBUTE");
            let is_cluster = self.check_keyword("CLUSTER");
            if !(is_sort || is_distribute || is_cluster) {
                break;
            }
            let saved = self.pos;
            self.advance();
            if self.peek_type() == &TokenType::By {
                self.advance();
                let _ = self.parse_order_by_items()?;
            } else {
                self.pos = saved;
                break;
            }
        }

        let (mut limit, mut offset) = if self.match_token(TokenType::Limit) {
            let first = self.parse_expr()?;
            // MySQL / ClickHouse `LIMIT offset, count` form — convert to
            // `LIMIT count OFFSET offset`.
            if self.match_token(TokenType::Comma) {
                let count = self.parse_expr()?;
                (Some(count), Some(first))
            } else {
                (Some(first), None)
            }
        } else {
            (None, None)
        };

        // ClickHouse `LIMIT N BY col[, ...]` / `LIMIT N BY col LIMIT M` —
        // consume the BY-list and an optional outer LIMIT so the trailing
        // SETTINGS / FORMAT clauses still parse.
        if limit.is_some() && self.match_token(TokenType::By) {
            let _ = self.parse_expr_list_allow_item_alias()?;
            if self.match_token(TokenType::Limit) {
                let _ = self.parse_expr()?;
            }
        }

        if offset.is_none() && self.match_token(TokenType::Offset) {
            let expr = self.parse_expr()?;
            // T-SQL / ANSI SQL:2008 form: OFFSET n ROWS [FETCH …].
            // Consume the optional ROWS/ROW keyword so FETCH can match next.
            let _ = self.match_token(TokenType::Rows) || self.match_keyword("ROW");
            offset = Some(expr);
        } else if offset.is_some() {
            // Already populated from `LIMIT a, b`; still consume an explicit
            // `OFFSET n` if it appears so it does not leak into the trailer.
            if self.match_token(TokenType::Offset) {
                let expr = self.parse_expr()?;
                let _ = self.match_token(TokenType::Rows) || self.match_keyword("ROW");
                offset = Some(expr);
            }
        }

        // Trino / Presto: `OFFSET n LIMIT m` (ordering opposite to MySQL).
        // We've parsed OFFSET; accept a trailing LIMIT n.
        if limit.is_none() && self.match_token(TokenType::Limit) {
            limit = Some(self.parse_expr()?);
        }

        // FETCH FIRST|NEXT n ROWS ONLY (Oracle / ANSI SQL:2008 / T-SQL)
        let fetch_first = if self.match_token(TokenType::Fetch) {
            // consume FIRST or NEXT
            let _ = self.match_token(TokenType::First) || self.match_token(TokenType::Next);
            let count = self.parse_expr()?;
            // consume ROWS or ROW
            let _ = self.match_keyword("ROWS") || self.match_keyword("ROW");
            // consume ONLY
            let _ = self.match_token(TokenType::Only);
            Some(count)
        } else {
            None
        };

        // ClickHouse trailing `WITH TOTALS` / `WITH TIES` / `WITH ROLLUP` /
        // `WITH CUBE` / `WITH FILL` modifiers in subquery position. These
        // are query-level modifiers we don't model; swallow so the
        // surrounding `)` is reached.
        if matches!(self.peek_type(), TokenType::With) {
            let after = self.peek_offset(1);
            let is_postfix_modifier = after
                .map(|t| {
                    matches!(
                        t.token_type,
                        TokenType::Identifier | TokenType::Cube | TokenType::Rollup
                    ) && matches!(
                        t.value.to_uppercase().as_str(),
                        "TOTALS" | "TIES" | "FILL" | "ROLLUP" | "CUBE"
                    )
                })
                .unwrap_or(false);
            if is_postfix_modifier {
                self.advance(); // WITH
                self.advance(); // modifier keyword
            }
        }

        // ClickHouse `SETTINGS k = v, ...` / `FORMAT <name>` and MySQL
        // `INTO OUTFILE 'file'` style trailing clauses. None of these have
        // a dedicated AST representation; consume to keep the surrounding
        // statement parseable.
        loop {
            if self.check_keyword("SETTINGS")
                || self.check_keyword("FORMAT")
                || self.check_keyword("INTO")
            {
                self.skip_trailing_options();
                break;
            }
            break;
        }

        Ok(SelectStatement {
            comments: vec![],
            ctes,
            distinct,
            top,
            columns,
            from,
            joins,
            where_clause,
            group_by,
            having,
            order_by,
            limit,
            offset,
            fetch_first,
            qualify,
            window_definitions,
        })
    }

    fn parse_window_definitions(&mut self) -> Result<Vec<WindowDefinition>> {
        let mut defs = Vec::new();
        loop {
            let name = self.expect_name()?;
            self.expect(TokenType::As)?;
            self.expect(TokenType::LParen)?;
            let spec = self.parse_window_spec()?;
            self.expect(TokenType::RParen)?;
            defs.push(WindowDefinition { name, spec });
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }
        Ok(defs)
    }

    /// Check if we should parse a set operation (UNION / INTERSECT / EXCEPT)
    fn maybe_parse_set_operation(&mut self, left: Statement) -> Result<Statement> {
        let op = match self.peek_type() {
            TokenType::Union => SetOperationType::Union,
            TokenType::Intersect => SetOperationType::Intersect,
            TokenType::Except => SetOperationType::Except,
            _ => {
                // Spark / Oracle `MINUS` as a synonym for `EXCEPT`.
                if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("MINUS") {
                    self.advance();
                    let all = self.match_token(TokenType::All);
                    let _ = self.match_token(TokenType::Distinct);
                    let right = self.parse_statement_inner()?;
                    return Ok(Statement::SetOperation(SetOperationStatement {
                        comments: vec![],
                        op: SetOperationType::Except,
                        all,
                        left: Box::new(left),
                        right: Box::new(right),
                        order_by: vec![],
                        limit: None,
                        offset: None,
                    }));
                }
                return Ok(left);
            }
        };
        self.advance();

        let all = self.match_token(TokenType::All);
        let _ = self.match_token(TokenType::Distinct); // UNION DISTINCT

        // DuckDB `UNION ALL BY NAME` / `UNION BY NAME` — column-name-based
        // set operation. Swallow the modifier so the inner SELECT parses.
        if self.match_token(TokenType::By) {
            if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("NAME") {
                self.advance();
            }
        }

        let right = self.parse_statement_inner()?;

        // Check for further set operations chaining
        let combined = Statement::SetOperation(SetOperationStatement {
            comments: vec![],
            op,
            all,
            left: Box::new(left),
            right: Box::new(right),
            order_by: vec![],
            limit: None,
            offset: None,
        });

        // Parse trailing ORDER BY / LIMIT / OFFSET that applies to the whole set operation
        if matches!(
            self.peek_type(),
            TokenType::Union | TokenType::Intersect | TokenType::Except
        ) {
            self.maybe_parse_set_operation(combined)
        } else {
            // Check for global ORDER BY / LIMIT
            if let Statement::SetOperation(mut sop) = combined {
                if self.match_token(TokenType::Order) {
                    self.expect(TokenType::By)?;
                    sop.order_by = self.parse_order_by_items()?;
                }
                if self.match_token(TokenType::Limit) {
                    sop.limit = Some(self.parse_expr()?);
                }
                if self.match_token(TokenType::Offset) {
                    sop.offset = Some(self.parse_expr()?);
                    // ANSI SQL:2008 / T-SQL: OFFSET n ROWS. Consume optional ROW(S).
                    let _ = self.match_token(TokenType::Rows) || self.match_keyword("ROW");
                }
                // Accept trailing LIMIT after OFFSET (OFFSET n LIMIT m ordering).
                if sop.limit.is_none() && self.match_token(TokenType::Limit) {
                    sop.limit = Some(self.parse_expr()?);
                }
                Ok(Statement::SetOperation(sop))
            } else {
                Ok(combined)
            }
        }
    }

    fn parse_select_items(&mut self) -> Result<Vec<SelectItem>> {
        let mut items = vec![self.parse_select_item()?];
        while self.match_token(TokenType::Comma) {
            // DuckDB / BigQuery / Snowflake allow a trailing comma in the
            // SELECT list before `FROM` / end of select clause. Bail out if
            // the next token can't start a select item.
            if matches!(
                self.peek_type(),
                TokenType::From
                    | TokenType::Where
                    | TokenType::Group
                    | TokenType::Order
                    | TokenType::Limit
                    | TokenType::Having
                    | TokenType::Qualify
                    | TokenType::Eof
                    | TokenType::Semicolon
                    | TokenType::RParen
                    | TokenType::Union
                    | TokenType::Intersect
                    | TokenType::Except
            ) {
                break;
            }
            items.push(self.parse_select_item()?);
        }
        Ok(items)
    }

    /// Consume DuckDB / Snowflake star modifiers — `EXCLUDE (...)`,
    /// `EXCEPT (...)`, `RENAME (...)`, `REPLACE (...)` — that may follow
    /// `*` or `t.*` in a SELECT list. Each modifier may appear at most
    /// once; we tolerate any order.
    fn swallow_star_modifiers(&mut self) {
        loop {
            let matched = self.check_keyword("EXCLUDE")
                || self.check_keyword("RENAME")
                || (self.check_keyword("REPLACE")
                    && matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::LParen)
                    ))
                || (self.peek_type() == &TokenType::Except
                    && matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::LParen)
                    ));
            // sqlfluff `SELECT * GLOB '…' FROM t` / `* SIMILAR TO '…'` /
            // `* LIKE '…'` style column-filter shorthand. Swallow the
            // operator and its pattern literal so the rest parses.
            let pattern_modifier = if matches!(self.peek_type(), TokenType::Like | TokenType::ILike)
                || (self.check_keyword("GLOB")
                    || self.check_keyword("REGEXP")
                    || self.check_keyword("RLIKE")
                    || self.check_keyword("IREGEXP")
                    || self.check_keyword("SIMILAR"))
            {
                let next_is_string = matches!(
                    self.peek_offset(1).map(|t| &t.token_type),
                    Some(TokenType::String)
                );
                let is_similar_to = self.check_keyword("SIMILAR")
                    && self
                        .peek_offset(1)
                        .map(|t| t.value.eq_ignore_ascii_case("TO"))
                        .unwrap_or(false);
                next_is_string || is_similar_to
            } else {
                false
            };
            if !matched && !pattern_modifier {
                break;
            }
            if pattern_modifier {
                // Operator keyword (and optional TO for SIMILAR TO) +
                // pattern string. We're tolerant of extra ESCAPE clause.
                self.advance(); // GLOB / LIKE / etc.
                if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("TO") {
                    self.advance();
                }
                if matches!(self.peek_type(), TokenType::String) {
                    self.advance();
                    if self.match_token(TokenType::Escape) {
                        if matches!(self.peek_type(), TokenType::String) {
                            self.advance();
                        }
                    }
                }
                continue;
            }
            self.advance(); // keyword
            if self.match_token(TokenType::LParen) {
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
            } else if self.is_name_token() {
                // EXCLUDE col (single-column without parens)
                self.advance();
            }
        }
    }

    fn parse_select_item(&mut self) -> Result<SelectItem> {
        if self.peek().token_type == TokenType::Star {
            self.advance();
            // DuckDB / Snowflake `* EXCLUDE (col, ...)`,
            // `* RENAME (a AS b, ...)`, `* REPLACE (expr AS col, ...)`.
            // Swallow the modifier so the surrounding select parses.
            self.swallow_star_modifiers();
            return Ok(SelectItem::Wildcard);
        }

        // DuckDB struct-shorthand alias-first form: `alias: expr` in a SELECT
        // list. Only fire when we see `<name> :` followed by something that
        // is not another `:` (which would form `::` cast) — i.e. a leading
        // alias-then-colon pattern. The alias may be any name-like token.
        if self.is_name_token() {
            let pos1 = self.peek_offset(1).map(|t| &t.token_type);
            let pos2 = self.peek_offset(2).map(|t| &t.token_type);
            if matches!(pos1, Some(TokenType::Colon)) && !matches!(pos2, Some(TokenType::Colon)) {
                // Save state so we can roll back if the trailing expression
                // fails to parse (avoids misclassifying obscure forms).
                let saved = self.pos;
                let alias_tok = self.advance().clone();
                self.advance(); // consume ':'
                if let Ok(expr) = self.parse_expr() {
                    return Ok(SelectItem::Expr {
                        expr,
                        alias: Some(alias_tok.value),
                        alias_quote_style: quote_style_from_char(alias_tok.quote_char),
                    });
                }
                self.pos = saved;
            }
        }

        let expr = self.parse_expr()?;

        // Check for table.* pattern
        if let Expr::QualifiedWildcard { ref table } = expr {
            self.swallow_star_modifiers();
            return Ok(SelectItem::QualifiedWildcard {
                table: table.clone(),
            });
        }

        // Hive scripting: `SELECT TRANSFORM(cols) [ROW FORMAT ...] USING
        // 'cmd' [AS (cols)] [ROW FORMAT ...] [RECORDREADER 'cls']`. The
        // tail clauses appear between the function call and `FROM`. We
        // don't model the scripting AST yet; swallow opaquely so the rest
        // of the SELECT parses.
        if matches!(
            &expr,
            Expr::Function { name, .. } if name.eq_ignore_ascii_case("TRANSFORM")
        ) {
            while !matches!(
                self.peek_type(),
                TokenType::From | TokenType::Eof | TokenType::Semicolon | TokenType::Comma
            ) {
                let v = self.peek().value.to_uppercase();
                let is_tail = self.peek_type() == &TokenType::Using
                    || self.peek_type() == &TokenType::As
                    || matches!(
                        v.as_str(),
                        "ROW"
                            | "FORMAT"
                            | "SERDE"
                            | "WITH"
                            | "SERDEPROPERTIES"
                            | "RECORDREADER"
                            | "RECORDWRITER"
                            | "FIELDS"
                            | "TERMINATED"
                            | "BY"
                            | "COLLECTION"
                            | "ITEMS"
                            | "MAP"
                            | "KEYS"
                            | "LINES"
                            | "NULL"
                            | "DEFINED"
                            | "STORED"
                            | "DELIMITED"
                            | "ESCAPED"
                            | "LOCATION"
                            | "OUTPUTFORMAT"
                            | "INPUTFORMAT"
                    );
                if !is_tail
                    && !matches!(
                        self.peek_type(),
                        TokenType::String
                            | TokenType::LParen
                            | TokenType::RParen
                            | TokenType::Identifier
                            | TokenType::Eq
                    )
                {
                    break;
                }
                self.advance();
            }
            return Ok(SelectItem::Expr {
                expr,
                alias: None,
                alias_quote_style: QuoteStyle::None,
            });
        }

        let (alias, alias_quote_style) = match self.parse_optional_alias()? {
            Some((name, qs)) => (Some(name), qs),
            None => (None, QuoteStyle::None),
        };

        Ok(SelectItem::Expr {
            expr,
            alias,
            alias_quote_style,
        })
    }

    fn parse_optional_alias(&mut self) -> Result<Option<(String, QuoteStyle)>> {
        if self.match_token(TokenType::As) {
            // After AS, also accept `@name` / `:name` as an alias. Both forms
            // appear in auto-generated SQL corpora (e.g. `AS @rpm`, `AS :minutes`)
            // where the symbol is part of the column name from the source data.
            if let Some((name, qs)) = self.try_parse_prefixed_alias()? {
                return Ok(Some((name, qs)));
            }
            // PostgreSQL / SQLite tolerate reserved-word literals as aliases
            // (`SELECT bool 't' AS true`). Accept TRUE / FALSE / NULL tokens.
            if matches!(
                self.peek_type(),
                TokenType::True | TokenType::False | TokenType::Null
            ) {
                let token = self.advance().clone();
                return Ok(Some((token.value, QuoteStyle::None)));
            }
            // DuckDB allows column aliases that collide with reserved
            // keywords (`AS matched`, `AS or`, `AS using`). After AS, take
            // whatever non-structural token appears.
            if matches!(
                self.peek_type(),
                TokenType::Matched
                    | TokenType::Or
                    | TokenType::And
                    | TokenType::Using
                    | TokenType::When
                    | TokenType::Where
                    | TokenType::Asc
                    | TokenType::Desc
                    | TokenType::Limit
                    | TokenType::Group
                    | TokenType::Having
                    | TokenType::On
                    | TokenType::Into
                    | TokenType::From
                    | TokenType::Order
                    | TokenType::Like
            ) {
                let token = self.advance().clone();
                return Ok(Some((token.value, QuoteStyle::None)));
            }
            return Ok(Some(self.expect_name_with_quote()?));
        }
        // Implicit alias
        if self.is_name_token() {
            let peeked_upper = self.peek().value.to_uppercase();
            if !matches!(
                peeked_upper.as_str(),
                "FROM"
                    | "WHERE"
                    | "GROUP"
                    | "ORDER"
                    | "LIMIT"
                    | "HAVING"
                    | "UNION"
                    | "INTERSECT"
                    | "EXCEPT"
                    | "JOIN"
                    | "INNER"
                    | "LEFT"
                    | "RIGHT"
                    | "FULL"
                    | "CROSS"
                    | "ON"
                    | "WINDOW"
                    | "QUALIFY"
                    | "INTO"
                    | "SET"
                    | "RETURNING"
                    | "PIVOT"
                    | "UNPIVOT"
                    | "PREWHERE"
                    | "SETTINGS"
                    | "FORMAT"
                    | "SAMPLE"
                    | "TABLESAMPLE"
                    | "LATERAL"
                    | "USING"
                    | "OFFSET"
                    | "FETCH"
                    | "FOR"
                    | "WITH"
                    | "OPTION"
                    | "MATCH_RECOGNIZE"
                    | "SORT"
                    | "DISTRIBUTE"
                    | "CLUSTER"
                    | "GLOBAL"
                    | "PREFERRING"
                    | "FORCE"
                    | "USE"
                    | "IGNORE"
                    | "STRAIGHT_JOIN"
                    | "DISTRIBUTED"
                    | "VALUE"
                    | "VALUES"
                    | "DEFAULT"
                    | "PARTITION"
            ) {
                let token = self.advance().clone();
                let qs = quote_style_from_char(token.quote_char);
                return Ok(Some((token.value.clone(), qs)));
            }
        }
        Ok(None)
    }

    fn parse_table_source(&mut self) -> Result<TableSource> {
        let mut source = self.parse_base_table_source()?;
        // PostgreSQL table-inheritance star: `FROM parent*` includes all
        // child tables. Swallow the trailing `*` so the table alias /
        // joins continue to parse.
        let _ = self.match_token(TokenType::Star);
        // BigQuery / Snowflake / MySQL TiDB time-travel:
        //   `<tbl> [FOR SYSTEM_TIME] AS OF [TIMESTAMP] <expr>` or
        //   `<tbl> AS OF VERSION <expr>` / `AS OF TIMESTAMP <expr>`.
        // We don't model the time-travel clause in the AST; swallow the
        // keywords and the expression so the surrounding query parses.
        if self.is_name_token()
            && self.peek().value.eq_ignore_ascii_case("FOR")
            && self
                .peek_offset(1)
                .map(|t| t.value.eq_ignore_ascii_case("SYSTEM_TIME"))
                .unwrap_or(false)
        {
            self.advance(); // FOR
            self.advance(); // SYSTEM_TIME
        }
        if self.peek_type() == &TokenType::As
            && self
                .peek_offset(1)
                .map(|t| t.value.eq_ignore_ascii_case("OF"))
                .unwrap_or(false)
        {
            self.advance(); // AS
            self.advance(); // OF
            // Optional TIMESTAMP / VERSION qualifier.
            if matches!(self.peek_type(), TokenType::Timestamp)
                || (self.is_name_token()
                    && matches!(
                        self.peek().value.to_uppercase().as_str(),
                        "VERSION" | "SCN" | "SEQUENCE"
                    ))
            {
                self.advance();
            }
            let _ = self.parse_expr()?;
        }
        // Hive / Spark / Trino `TABLESAMPLE [method] (...)` after a table
        // reference. We don't model the sample clause in the AST; just
        // consume the optional method identifier (BERNOULLI / SYSTEM /
        // RESERVOIR) and the parenthesized body so the surrounding query
        // parses. Also accept an optional `REPEATABLE (n)` trailer.
        if self.match_token(TokenType::Tablesample) {
            // Optional sampling method identifier.
            if matches!(self.peek_type(), TokenType::Identifier) {
                self.advance();
            }
            if self.match_token(TokenType::LParen) {
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
            }
            if self.check_keyword("REPEATABLE") {
                self.advance();
                if self.match_token(TokenType::LParen) {
                    let mut depth = 1;
                    while depth > 0 {
                        match self.peek_type() {
                            TokenType::LParen => depth += 1,
                            TokenType::RParen => {
                                depth -= 1;
                                if depth == 0 {
                                    self.advance();
                                    break;
                                }
                            }
                            TokenType::Eof => break,
                            _ => {}
                        }
                        self.advance();
                    }
                }
            }
            // Optional trailing alias on the sampled table — `… TABLESAMPLE
            // (…) s`. We attach it to the underlying table reference when
            // possible, otherwise just consume the identifier.
            if let TableSource::Table(ref mut tr) = source {
                if tr.alias.is_none() {
                    if let Some((name, qs)) = self.parse_optional_alias()? {
                        tr.alias = Some(name);
                        tr.alias_quote_style = qs;
                    }
                }
            }
        }
        // Check for trailing PIVOT / UNPIVOT
        let source = self.parse_pivot_or_unpivot(source)?;
        // ClickHouse: `SELECT * FROM t SAMPLE 0.1` (no parens) — and the
        // optional `OFFSET m` modifier. The keyword tokenizes as a plain
        // identifier so this also handles dialects that don't reserve it.
        if self.check_keyword("SAMPLE") {
            self.advance();
            // Accept a number, identifier, or parenthesized expression.
            if matches!(self.peek_type(), TokenType::Number) {
                self.advance();
                // Optional `/ N` ratio.
                if self.peek_type() == &TokenType::Slash {
                    self.advance();
                    if matches!(self.peek_type(), TokenType::Number) {
                        self.advance();
                    }
                }
            }
            if self.check_keyword("OFFSET") {
                self.advance();
                if matches!(self.peek_type(), TokenType::Number) {
                    self.advance();
                }
            }
        }
        Ok(source)
    }

    fn parse_base_table_source(&mut self) -> Result<TableSource> {
        // LATERAL
        if self.match_token(TokenType::Lateral) {
            let source = self.parse_table_source()?;
            return Ok(TableSource::Lateral {
                source: Box::new(source),
            });
        }

        // Spark / DuckDB / Postgres `FROM VALUES (...) [, (...)]+ [alias[(cols)]]`
        // (un-parenthesised VALUES list). Swallow the rows.
        if self.match_token(TokenType::Values) {
            // First row.
            if self.match_token(TokenType::LParen) {
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
            }
            // Additional rows.
            while self.peek_type() == &TokenType::Comma {
                let saved = self.pos;
                self.advance();
                if !self.match_token(TokenType::LParen) {
                    // Not a row — restore comma for the outer parser.
                    self.pos = saved;
                    break;
                }
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
            }
            let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                Some((name, qs)) => (Some(name), qs),
                None => (None, QuoteStyle::None),
            };
            if alias.is_some() && self.peek_type() == &TokenType::LParen {
                let saved = self.pos;
                self.advance();
                let mut ok = true;
                loop {
                    if !self.is_name_token() {
                        ok = false;
                        break;
                    }
                    self.advance();
                    if self.match_token(TokenType::RParen) {
                        break;
                    }
                    if !self.match_token(TokenType::Comma) {
                        ok = false;
                        break;
                    }
                }
                if !ok {
                    self.pos = saved;
                }
            }
            return Ok(TableSource::TableFunction {
                name: "VALUES".to_string(),
                args: vec![],
                alias,
                alias_quote_style,
            });
        }

        // UNNEST(expr)
        if self.match_token(TokenType::Unnest) {
            self.expect(TokenType::LParen)?;
            let expr = self.parse_expr()?;
            // Multi-arg form (Trino): UNNEST(a, b, c). Drop extras.
            while self.match_token(TokenType::Comma) {
                let _ = self.parse_expr()?;
            }
            self.expect(TokenType::RParen)?;
            let (mut alias, mut alias_quote_style) = match self.parse_optional_alias()? {
                Some((name, qs)) => (Some(name), qs),
                None => (None, QuoteStyle::None),
            };
            // BigQuery `WITH OFFSET [AS name]` / Postgres `WITH ORDINALITY`.
            let mut with_offset = false;
            if self.check_keyword("WITH") {
                let saved = self.pos;
                self.advance();
                if self.check_keyword("OFFSET") || self.check_keyword("ORDINALITY") {
                    self.advance();
                    with_offset = true;
                    // Optional alias after OFFSET / ORDINALITY.
                    if alias.is_none() {
                        if let Some((n, qs)) = self.parse_optional_alias()? {
                            alias = Some(n);
                            alias_quote_style = qs;
                        }
                    } else if self.is_name_token() {
                        // `UNNEST(a) id WITH OFFSET pos` — extra trailing
                        // name; absorb so we don't trip the join parser.
                        self.advance();
                    }
                } else {
                    self.pos = saved;
                }
            }
            // Optional positional column list: `AS t (n, a)`.
            if alias.is_some() && self.peek_type() == &TokenType::LParen {
                let saved = self.pos;
                self.advance();
                let mut ok = true;
                loop {
                    if !self.is_name_token() {
                        ok = false;
                        break;
                    }
                    self.advance();
                    if self.match_token(TokenType::RParen) {
                        break;
                    }
                    if !self.match_token(TokenType::Comma) {
                        ok = false;
                        break;
                    }
                }
                if !ok {
                    self.pos = saved;
                }
            }
            return Ok(TableSource::Unnest {
                expr: Box::new(expr),
                alias,
                alias_quote_style,
                with_offset,
            });
        }

        // Subquery: (SELECT ...)
        if self.peek_type() == &TokenType::LParen {
            let saved = self.pos;
            self.advance();
            // A derived-table body is a subquery when it begins with a statement
            // keyword, or with another `(` when the body is itself a
            // parenthesised (possibly set-operation) query — e.g. redundant
            // nesting `((SELECT …))` or a set operation whose branches are each
            // parenthesised `((SELECT …) EXCEPT (SELECT …))`. Delegate to the
            // recursive statement parser (the same path the top-level set-op
            // parser uses) rather than hand-counting parens, which cannot tell
            // redundant wrapping apart from a parenthesised first branch
            // (CR-014).
            let direct_subquery = matches!(
                self.peek_type(),
                TokenType::Select
                    | TokenType::With
                    | TokenType::Explain
                    | TokenType::From
                    | TokenType::Describe
                    | TokenType::Show
                    | TokenType::Table
            );
            // A `(`-led body may instead be a parenthesised join / table list
            // (`((t1 JOIN t2)) alias`). Attempt the subquery interpretation and,
            // on failure, restore to `saved` so the parenthesised-join handling
            // further below still runs.
            let paren_body = self.peek_type() == &TokenType::LParen;
            let mut subquery: Option<Statement> = None;
            if direct_subquery {
                let query = self.parse_statement_inner()?;
                // Set operations across parenthesised subqueries: `(SELECT …)
                // UNION ALL (SELECT …) [ORDER BY …] [LIMIT …]`.
                let query = self.maybe_parse_set_operation(query)?;
                self.expect(TokenType::RParen)?;
                subquery = Some(query);
            } else if paren_body {
                let attempt = self
                    .parse_statement_inner()
                    .and_then(|q| self.maybe_parse_set_operation(q));
                match attempt {
                    Ok(query) if self.peek_type() == &TokenType::RParen => {
                        self.advance();
                        subquery = Some(query);
                    }
                    _ => self.pos = saved,
                }
            }
            if let Some(query) = subquery {
                let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                    Some((name, qs)) => (Some(name), qs),
                    None => (None, QuoteStyle::None),
                };
                // Positional column-list alias: `(SELECT ...) t(c1, c2)`
                if alias.is_some() && self.peek_type() == &TokenType::LParen {
                    let saved2 = self.pos;
                    self.advance();
                    let mut ok = true;
                    loop {
                        if !self.is_name_token() {
                            ok = false;
                            break;
                        }
                        self.advance();
                        if self.match_token(TokenType::RParen) {
                            break;
                        }
                        if !self.match_token(TokenType::Comma) {
                            ok = false;
                            break;
                        }
                    }
                    if !ok {
                        self.pos = saved2;
                    }
                }
                return Ok(TableSource::Subquery {
                    query: Box::new(query),
                    alias,
                    alias_quote_style,
                });
            }
            // `(VALUES (...), (...)) alias[(cols)]` — common in DuckDB /
            // Postgres derived tables. We don't model the VALUES rows in the
            // AST as a table source; swallow the parenthesized body and
            // synthesise an empty subquery placeholder.
            if self.peek_type() == &TokenType::Values {
                // Re-advance past the values list, balancing parens (we are
                // inside the outer LParen at depth 1).
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
                let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                    Some((name, qs)) => (Some(name), qs),
                    None => (None, QuoteStyle::None),
                };
                if alias.is_some() && self.peek_type() == &TokenType::LParen {
                    let saved2 = self.pos;
                    self.advance();
                    let mut ok = true;
                    loop {
                        if !self.is_name_token() {
                            ok = false;
                            break;
                        }
                        self.advance();
                        if self.match_token(TokenType::RParen) {
                            break;
                        }
                        if !self.match_token(TokenType::Comma) {
                            ok = false;
                            break;
                        }
                    }
                    if !ok {
                        self.pos = saved2;
                    }
                }
                // Synthesise an empty values placeholder. Reuse Subquery
                // with a single-row Insert wrapper is awkward; instead,
                // wrap as a TableFunction("VALUES") with empty args.
                return Ok(TableSource::TableFunction {
                    name: "VALUES".to_string(),
                    args: vec![],
                    alias,
                    alias_quote_style,
                });
            }
            self.pos = saved;

            // MySQL / SQLite / others permit parenthesized join expressions
            // as a table source: `(t1 LEFT JOIN t2 ON …)` or comma-list
            // `(t1, t2)`. Recurse into the parens, then consume joins /
            // commas until the matching `)`. Emit the first source so the
            // surrounding query parses; trailing tables are discarded
            // (their predicates were already parsed into the JOIN node we
            // throw away — acceptance only).
            if self.peek_type() == &TokenType::LParen {
                let inner_saved = self.pos;
                self.advance();
                let after_lparen = self.pos;
                if let Ok(inner) = self.parse_table_source() {
                    let _ = self.parse_joins();
                    while self.match_token(TokenType::Comma) {
                        if self.parse_table_source().is_err() {
                            self.pos = inner_saved;
                            // Fall through to the generic parse_table_ref
                            // path below, which will surface the original
                            // error message.
                            break;
                        }
                        let _ = self.parse_joins();
                    }
                    if self.pos != inner_saved && self.match_token(TokenType::RParen) {
                        let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                            Some((name, qs)) => (Some(name), qs),
                            None => (None, QuoteStyle::None),
                        };
                        if let Some(name) = alias.clone() {
                            if let TableSource::Table(mut tr) = inner {
                                tr.alias = Some(name);
                                tr.alias_quote_style = alias_quote_style;
                                return Ok(TableSource::Table(tr));
                            }
                        }
                        return Ok(inner);
                    }
                }
                // Restore so the caller sees the LParen and emits a useful
                // error rather than silently misparsing partial state.
                self.pos = inner_saved;
                let _ = after_lparen; // suppress unused warning when build optimises
            }
        }

        // Regular table reference (possibly with function syntax)
        let table_ref = self.parse_table_ref()?;

        // MySQL / TiDB partition selector: `tbl PARTITION (p0, p1)`. Swallow
        // it so the table reference parses cleanly.
        if matches!(self.peek_type(), TokenType::Partition)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::LParen)
            )
        {
            self.advance();
            self.advance();
            while !matches!(self.peek_type(), TokenType::RParen | TokenType::Eof) {
                self.advance();
            }
            let _ = self.match_token(TokenType::RParen);
        }

        // Check if it's actually a table function: name(args...). Also
        // accept dotted qualifiers so DuckDB `schema.func(...)` /
        // `catalog.schema.func(...)` parse.
        if self.peek_type() == &TokenType::LParen {
            // SQL/PGQ `GRAPH_TABLE(graph MATCH … COLUMNS (…))`,
            // SQL/XML `XMLTABLE('xpath' PASSING expr COLUMNS …)`,
            // SQL/JSON `JSON_TABLE(expr, '$' COLUMNS (…))`. Swallow the
            // body opaquely so the rest of the query parses.
            let fname = table_ref.name.to_uppercase();
            if matches!(
                fname.as_str(),
                "GRAPH_TABLE" | "XMLTABLE" | "JSON_TABLE" | "OPENJSON" | "OPENROWSET" | "OPENXML"
            ) {
                self.advance();
                let mut depth = 1usize;
                while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        _ => {}
                    }
                    self.advance();
                }
                let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                    Some((name, qs)) => (Some(name), qs),
                    None => (None, QuoteStyle::None),
                };
                if alias.is_some() && self.peek_type() == &TokenType::LParen {
                    let saved = self.pos;
                    self.advance();
                    let mut ok = true;
                    loop {
                        if !self.is_name_token() {
                            ok = false;
                            break;
                        }
                        self.advance();
                        if self.match_token(TokenType::RParen) {
                            break;
                        }
                        if !self.match_token(TokenType::Comma) {
                            ok = false;
                            break;
                        }
                    }
                    if !ok {
                        self.pos = saved;
                    }
                }
                return Ok(TableSource::TableFunction {
                    name: match (&table_ref.catalog, &table_ref.schema) {
                        (Some(c), Some(s)) => format!("{}.{}.{}", c, s, table_ref.name),
                        (None, Some(s)) => format!("{}.{}", s, table_ref.name),
                        _ => table_ref.name,
                    },
                    args: vec![],
                    alias,
                    alias_quote_style,
                });
            }
            self.advance();
            // Hive `noop(on tbl partition by ... order by ... )` table-valued
            // function. Arguments start with the `ON` keyword and include
            // PARTITION/ORDER/CLUSTER/DISTRIBUTE/SORT BY clauses we don't
            // model. Swallow the body opaquely.
            let args = if matches!(self.peek_type(), TokenType::On) {
                let mut depth = 0usize;
                while !matches!(self.peek_type(), TokenType::Eof) {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            if depth == 0 {
                                break;
                            }
                            depth -= 1;
                        }
                        _ => {}
                    }
                    self.advance();
                }
                vec![]
            } else if self.peek_type() != &TokenType::RParen {
                self.parse_expr_list()?
            } else {
                vec![]
            };
            self.expect(TokenType::RParen)?;
            let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                Some((name, qs)) => (Some(name), qs),
                None => (None, QuoteStyle::None),
            };
            // DuckDB / Postgres positional column-list alias:
            //   range(10) t(i)   →   alias = "t", columns = (i)
            // We consume the parenthesized list but do not model it in the AST.
            if alias.is_some() && self.peek_type() == &TokenType::LParen {
                let saved = self.pos;
                self.advance();
                let mut ok = true;
                loop {
                    if !self.is_name_token() {
                        ok = false;
                        break;
                    }
                    self.advance();
                    if self.match_token(TokenType::RParen) {
                        break;
                    }
                    if !self.match_token(TokenType::Comma) {
                        ok = false;
                        break;
                    }
                }
                if !ok {
                    self.pos = saved;
                }
            }
            return Ok(TableSource::TableFunction {
                name: match (&table_ref.catalog, &table_ref.schema) {
                    (Some(c), Some(s)) => format!("{}.{}.{}", c, s, table_ref.name),
                    (None, Some(s)) => format!("{}.{}", s, table_ref.name),
                    _ => table_ref.name,
                },
                args,
                alias,
                alias_quote_style,
            });
        }

        // Also support positional column-list alias on a plain table reference:
        //   FROM tbl t(c1, c2)
        if self.peek_type() == &TokenType::LParen && table_ref.alias.is_some() {
            let saved = self.pos;
            self.advance();
            let mut ok = true;
            loop {
                if !self.is_name_token() {
                    ok = false;
                    break;
                }
                self.advance();
                if self.match_token(TokenType::RParen) {
                    break;
                }
                if !self.match_token(TokenType::Comma) {
                    ok = false;
                    break;
                }
            }
            if !ok {
                self.pos = saved;
            }
        }

        // MySQL / MariaDB index hints — `USE INDEX (idx)`, `FORCE INDEX (idx)`,
        // `IGNORE INDEX (idx)`, optionally with `FOR JOIN|ORDER BY|GROUP BY`.
        // Swallow any sequence of these so the rest of the query parses.
        loop {
            let saved = self.pos;
            let is_hint = matches!(self.peek_type(), TokenType::Use | TokenType::Ignore)
                || self.check_keyword("FORCE");
            if !is_hint {
                break;
            }
            self.advance();
            if !self.check_keyword("INDEX") && !self.check_keyword("KEY") {
                self.pos = saved;
                break;
            }
            self.advance();
            // Optional `FOR JOIN | FOR ORDER BY | FOR GROUP BY`.
            if self.match_keyword("FOR") {
                if matches!(
                    self.peek_type(),
                    TokenType::Join | TokenType::Order | TokenType::Group
                ) {
                    self.advance();
                    let _ = self.match_token(TokenType::By);
                }
            }
            if self.match_token(TokenType::LParen) {
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        }
                        TokenType::Eof => break,
                        _ => {}
                    }
                    self.advance();
                }
            }
        }

        // ClickHouse `FROM tbl [AS alias] FINAL` — swallow the FINAL modifier.
        // The token tokenizes as Identifier so check_keyword is enough.
        if self.check_keyword("FINAL") {
            self.advance();
        }

        // MySQL: `FROM t PARTITION (p0[, p1, ...])` — swallow partition
        // selector. May appear before or after the alias; we accept it
        // here (i.e., before parse_optional_alias has run).
        if matches!(self.peek_type(), TokenType::Partition)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::LParen)
            )
        {
            self.advance();
            self.advance();
            let mut depth = 1;
            while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    _ => {}
                }
                self.advance();
            }
        }

        Ok(TableSource::Table(table_ref))
    }

    /// After parsing a base table source, check if PIVOT or UNPIVOT follows.
    fn parse_pivot_or_unpivot(&mut self, source: TableSource) -> Result<TableSource> {
        if self.match_token(TokenType::Pivot) {
            self.expect(TokenType::LParen)?;
            let aggregate = self.parse_expr()?;
            // Snowflake / Databricks: optional `AS <alias>` on the aggregate
            // expression: `PIVOT (sum(sales) AS sales FOR …)`.
            if self.peek_type() == &TokenType::As
                && self
                    .peek_offset(1)
                    .map(|t| {
                        matches!(
                            t.token_type,
                            TokenType::Identifier | TokenType::String | TokenType::Number
                        )
                    })
                    .unwrap_or(false)
            {
                self.advance();
                self.advance();
            }
            // Multi-aggregate PIVOT: `PIVOT (SUM(x), COUNT(x) FOR …)`. Drop
            // the extra aggregates — we only keep the first one in the AST.
            while self.match_token(TokenType::Comma) {
                let _ = self.parse_expr()?;
                if self.peek_type() == &TokenType::As
                    && self
                        .peek_offset(1)
                        .map(|t| {
                            matches!(
                                t.token_type,
                                TokenType::Identifier | TokenType::String | TokenType::Number
                            )
                        })
                        .unwrap_or(false)
                {
                    self.advance();
                    self.advance();
                }
            }
            self.expect_keyword("FOR")?;
            // Snowflake `FOR (col1, col2) IN …` — grouped pivot key. Use the
            // first column name as the AST's for_column.
            let for_column = if self.peek_type() == &TokenType::LParen {
                self.advance();
                let first = self.expect_name()?;
                while self.match_token(TokenType::Comma) {
                    let _ = self.expect_name()?;
                }
                self.expect(TokenType::RParen)?;
                first
            } else {
                self.expect_name()?
            };
            self.expect(TokenType::In)?;
            self.expect(TokenType::LParen)?;
            let in_values = self.parse_pivot_values()?;
            self.expect(TokenType::RParen)?;
            self.expect(TokenType::RParen)?;
            let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                Some((name, qs)) => (Some(name), qs),
                None => (None, QuoteStyle::None),
            };
            return Ok(TableSource::Pivot {
                source: Box::new(source),
                aggregate: Box::new(aggregate),
                for_column,
                in_values,
                alias,
                alias_quote_style,
            });
        }
        if self.match_token(TokenType::Unpivot) {
            // BigQuery: `UNPIVOT INCLUDE|EXCLUDE NULLS (...)`.
            if self.check_keyword("INCLUDE") || self.check_keyword("EXCLUDE") {
                let saved = self.pos;
                self.advance();
                if !self.match_keyword("NULLS") {
                    self.pos = saved;
                }
            }
            self.expect(TokenType::LParen)?;
            // Snowflake/DuckDB allow a grouped value-column tuple:
            // `UNPIVOT ((col1, col2) FOR period IN (...))`. Swallow the
            // grouping parens — we only model a single value-column name.
            let value_column = if self.peek_type() == &TokenType::LParen {
                self.advance();
                let first = self.expect_name()?;
                while self.match_token(TokenType::Comma) {
                    let _ = self.expect_name()?;
                }
                self.expect(TokenType::RParen)?;
                first
            } else {
                self.expect_name()?
            };
            self.expect_keyword("FOR")?;
            let for_column = self.expect_name()?;
            self.expect(TokenType::In)?;
            self.expect(TokenType::LParen)?;
            let in_columns = self.parse_pivot_values()?;
            self.expect(TokenType::RParen)?;
            self.expect(TokenType::RParen)?;
            let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                Some((name, qs)) => (Some(name), qs),
                None => (None, QuoteStyle::None),
            };
            return Ok(TableSource::Unpivot {
                source: Box::new(source),
                value_column,
                for_column,
                in_columns,
                alias,
                alias_quote_style,
            });
        }
        Ok(source)
    }

    /// Parse comma-separated pivot values, each optionally aliased.
    fn parse_pivot_values(&mut self) -> Result<Vec<PivotValue>> {
        let mut values = Vec::new();
        loop {
            let value = self.parse_expr()?;
            // Snowflake / BigQuery permit string or numeric aliases on pivot
            // values: `(a, b) AS 'semester_1'` / `(a, b) AS 1`. Accept those
            // alongside the regular identifier alias.
            let (alias, alias_quote_style) = if self.match_token(TokenType::As)
                && matches!(self.peek_type(), TokenType::String | TokenType::Number)
            {
                let tok = self.advance().clone();
                (Some(tok.value), QuoteStyle::None)
            } else {
                match self.parse_optional_alias()? {
                    Some((name, qs)) => (Some(name), qs),
                    None => (None, QuoteStyle::None),
                }
            };
            values.push(PivotValue {
                value,
                alias,
                alias_quote_style,
            });
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }
        Ok(values)
    }

    fn parse_table_ref(&mut self) -> Result<TableRef> {
        // T-SQL table variable: `FROM @t` / `INTO @t` etc. The @ is its own
        // token; fuse with the following name into a single identifier.
        if matches!(self.peek_type(), TokenType::AtSign)
            && self
                .peek_offset(1)
                .map(|t| {
                    matches!(t.token_type, TokenType::Identifier)
                        || matches!(t.token_type, TokenType::AtSign)
                })
                .unwrap_or(false)
        {
            let mut name = String::from("@");
            self.advance();
            if matches!(self.peek_type(), TokenType::AtSign) {
                name.push('@');
                self.advance();
            }
            let n = self.advance().clone();
            name.push_str(&n.value);
            let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                Some((a, qs)) => (Some(a), qs),
                None => (None, QuoteStyle::None),
            };
            return Ok(TableRef {
                catalog: None,
                schema: None,
                name,
                alias,
                name_quote_style: QuoteStyle::None,
                alias_quote_style,
            });
        }
        let (first, first_qs) = self.expect_name_with_quote()?;

        // Check for schema.table or catalog.schema.table. We also tolerate 4+
        // part qualified names (DuckDB / SQL Server `srv.db.sch.tbl`) by
        // folding additional segments into the catalog field.
        let (catalog, schema, name, name_qs) = if self.match_token(TokenType::Dot) {
            let (second, second_qs) = self.expect_name_with_quote()?;
            if self.match_token(TokenType::Dot) {
                let (mut third, mut third_qs) = self.expect_name_with_quote()?;
                let mut catalog = first;
                let mut schema = second;
                while self.match_token(TokenType::Dot) {
                    let (next, next_qs) = self.expect_name_with_quote()?;
                    catalog.push('.');
                    catalog.push_str(&schema);
                    schema = third;
                    third = next;
                    third_qs = next_qs;
                }
                (Some(catalog), Some(schema), third, third_qs)
            } else {
                (None, Some(first), second, second_qs)
            }
        } else {
            (None, None, first, first_qs)
        };

        let (alias, alias_quote_style) = match self.parse_optional_alias()? {
            Some((name, qs)) => (Some(name), qs),
            None => (None, QuoteStyle::None),
        };

        Ok(TableRef {
            catalog,
            schema,
            name,
            alias,
            name_quote_style: name_qs,
            alias_quote_style,
        })
    }

    /// Like `parse_table_ref` but does not consume an alias.
    fn parse_table_ref_no_alias(&mut self) -> Result<TableRef> {
        let (first, first_qs) = self.expect_name_with_quote()?;

        let (catalog, schema, name, name_qs) = if self.match_token(TokenType::Dot) {
            let (second, second_qs) = self.expect_name_with_quote()?;
            if self.match_token(TokenType::Dot) {
                let (mut third, mut third_qs) = self.expect_name_with_quote()?;
                let mut catalog = first;
                let mut schema = second;
                while self.match_token(TokenType::Dot) {
                    let (next, next_qs) = self.expect_name_with_quote()?;
                    catalog.push('.');
                    catalog.push_str(&schema);
                    schema = third;
                    third = next;
                    third_qs = next_qs;
                }
                (Some(catalog), Some(schema), third, third_qs)
            } else {
                (None, Some(first), second, second_qs)
            }
        } else {
            (None, None, first, first_qs)
        };

        Ok(TableRef {
            catalog,
            schema,
            name,
            alias: None,
            name_quote_style: name_qs,
            alias_quote_style: QuoteStyle::None,
        })
    }

    fn parse_joins(&mut self) -> Result<Vec<JoinClause>> {
        let mut joins = Vec::new();
        loop {
            // Hive `LATERAL VIEW [OUTER] func(args) tbl_alias [AS col, ...]`.
            // Model as a CROSS JOIN over a table-function so the rest of the
            // query parses; the AS column list is dropped.
            if self.peek_type() == &TokenType::Lateral
                && self
                    .peek_offset(1)
                    .map(|t| t.value.eq_ignore_ascii_case("VIEW"))
                    .unwrap_or(false)
            {
                self.advance(); // LATERAL
                self.advance(); // VIEW
                let _outer = self.check_keyword("OUTER") && {
                    self.advance();
                    true
                };
                // func(args) — parse name and arg list
                let fname = self.expect_name().unwrap_or_default();
                let mut fargs = Vec::new();
                if self.match_token(TokenType::LParen) {
                    if self.peek_type() != &TokenType::RParen {
                        fargs.push(self.parse_expr()?);
                        while self.match_token(TokenType::Comma) {
                            fargs.push(self.parse_expr()?);
                        }
                    }
                    self.expect(TokenType::RParen)?;
                }
                let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                    Some((name, qs)) => (Some(name), qs),
                    None => (None, QuoteStyle::None),
                };
                // Optional `[AS] col1[, col2, ...]` column list. Hive
                // allows the AS to be omitted entirely; Spark sometimes
                // emits `tbl_name col`. Consume names while we keep seeing
                // identifier-then-comma pairs.
                let _ = self.match_token(TokenType::As);
                if self.is_name_token() {
                    self.advance();
                    while self.match_token(TokenType::Comma) {
                        if !self.is_name_token() {
                            break;
                        }
                        self.advance();
                    }
                }
                joins.push(JoinClause {
                    join_type: JoinType::Cross,
                    table: TableSource::TableFunction {
                        name: fname,
                        args: fargs,
                        alias,
                        alias_quote_style,
                    },
                    on: None,
                    using: Vec::new(),
                });
                continue;
            }
            // ClickHouse: ARRAY JOIN / LEFT ARRAY JOIN — flatten arrays as join source.
            // We model it as a CROSS JOIN over the array expression.
            let saved_array = self.pos;
            let _left_array = self.match_token(TokenType::Left);
            if self.match_token(TokenType::Array) && self.match_token(TokenType::Join) {
                // parse the array expression(s) — comma-separated
                let mut sources = Vec::new();
                loop {
                    // ClickHouse permits inline array literals as the source:
                    //   ARRAY JOIN [1,2,3] AS x, [(...), (...)] AS y
                    // Wrap as Unnest so we don't reject the syntax.
                    let src = if matches!(self.peek_type(), TokenType::LBracket) {
                        let arr = self.parse_primary()?;
                        let (alias, alias_quote_style) = match self.parse_optional_alias()? {
                            Some((name, qs)) => (Some(name), qs),
                            None => (None, QuoteStyle::None),
                        };
                        TableSource::Unnest {
                            expr: Box::new(arr),
                            alias,
                            alias_quote_style,
                            with_offset: false,
                        }
                    } else {
                        self.parse_table_source()?
                    };
                    sources.push(src);
                    if !self.match_token(TokenType::Comma) {
                        break;
                    }
                }
                for src in sources {
                    joins.push(JoinClause {
                        join_type: JoinType::Cross,
                        table: src,
                        on: None,
                        using: Vec::new(),
                    });
                }
                continue;
            } else {
                self.pos = saved_array;
            }
            // ClickHouse / Hive join strictness modifiers — consume and drop:
            //   GLOBAL? ALL | ANY | SEMI | ANTI | ASOF [LEFT|RIGHT|INNER|OUTER] JOIN
            let saved_strictness = self.pos;
            let _global_prefix = self.check_keyword("GLOBAL") && {
                self.advance();
                true
            };
            let consumed_strictness = if self.match_token(TokenType::All) {
                true
            } else if self.match_token(TokenType::Any) {
                true
            } else if self.check_keyword("SEMI")
                || self.check_keyword("ANTI")
                || self.check_keyword("ASOF")
                || self.check_keyword("PASTE")
            {
                self.advance();
                // DuckDB / ClickHouse allow compound forms like
                // `ASOF ANTI JOIN` / `ASOF SEMI JOIN` — absorb a
                // following second strictness keyword too.
                if self.check_keyword("SEMI")
                    || self.check_keyword("ANTI")
                    || self.check_keyword("ASOF")
                {
                    self.advance();
                }
                true
            } else {
                _global_prefix
            };
            // If the strictness modifier wasn't followed by a join keyword,
            // rewind so we don't accidentally consume a stray ALL/ANY (e.g.
            // `ORDER BY ALL`).
            if consumed_strictness
                && !matches!(
                    self.peek_type(),
                    TokenType::Join
                        | TokenType::Inner
                        | TokenType::Left
                        | TokenType::Right
                        | TokenType::Full
                        | TokenType::Cross
                )
            {
                self.pos = saved_strictness;
            }
            let join_type = match self.peek_type() {
                // `FROM a, b` is treated as `FROM a CROSS JOIN b`. Note the
                // SQL standard gives comma a lower precedence than explicit
                // JOIN operators (so `FROM a, b JOIN c ON ...` should be
                // `a CROSS JOIN (b JOIN c ...)`), but we flatten everything
                // into a left-deep chain. Column resolution still works for
                // the common cases since the join order is associative when
                // ON-clauses only reference adjacent tables.
                TokenType::Comma => {
                    self.advance();
                    JoinType::Cross
                }
                // `NATURAL [LEFT|RIGHT|FULL [OUTER]] JOIN tbl` — auto-equi-join
                // on shared column names. We don't model NATURAL semantics yet;
                // promote to the corresponding non-natural join type and treat
                // the implicit USING clause as empty.
                t if matches!(t, TokenType::Identifier)
                    && self.peek().value.eq_ignore_ascii_case("NATURAL") =>
                {
                    self.advance(); // NATURAL
                    let jt = match self.peek_type() {
                        TokenType::Left => {
                            self.advance();
                            let _ = self.match_token(TokenType::Outer);
                            JoinType::Left
                        }
                        TokenType::Right => {
                            self.advance();
                            let _ = self.match_token(TokenType::Outer);
                            JoinType::Right
                        }
                        TokenType::Full => {
                            self.advance();
                            let _ = self.match_token(TokenType::Outer);
                            JoinType::Full
                        }
                        TokenType::Inner => {
                            self.advance();
                            JoinType::Inner
                        }
                        _ => JoinType::Inner,
                    };
                    self.expect(TokenType::Join)?;
                    jt
                }
                // MySQL `STRAIGHT_JOIN` — non-reordered INNER JOIN.
                t if matches!(t, TokenType::Identifier)
                    && self.peek().value.eq_ignore_ascii_case("STRAIGHT_JOIN") =>
                {
                    self.advance();
                    JoinType::Inner
                }
                TokenType::Join => {
                    self.advance();
                    JoinType::Inner
                }
                TokenType::Inner => {
                    self.advance();
                    self.expect(TokenType::Join)?;
                    JoinType::Inner
                }
                TokenType::Left => {
                    self.advance();
                    let _ = self.match_token(TokenType::Outer);
                    // Hive / Spark: LEFT SEMI JOIN / LEFT ANTI JOIN
                    let _ = self.check_keyword("SEMI") && {
                        self.advance();
                        true
                    } || self.check_keyword("ANTI") && {
                        self.advance();
                        true
                    };
                    // ClickHouse: LEFT ANY|ALL JOIN
                    let _ = self.match_token(TokenType::Any) || self.match_token(TokenType::All);
                    // Some dialects (Spark/Hive variants) allow a trailing
                    // OUTER after the strictness modifier.
                    let _ = self.match_token(TokenType::Outer);
                    self.expect(TokenType::Join)?;
                    JoinType::Left
                }
                TokenType::Right => {
                    self.advance();
                    let _ = self.match_token(TokenType::Outer);
                    let _ = self.check_keyword("SEMI") && {
                        self.advance();
                        true
                    } || self.check_keyword("ANTI") && {
                        self.advance();
                        true
                    };
                    let _ = self.match_token(TokenType::Any) || self.match_token(TokenType::All);
                    let _ = self.match_token(TokenType::Outer);
                    self.expect(TokenType::Join)?;
                    JoinType::Right
                }
                TokenType::Full => {
                    self.advance();
                    let _ = self.match_token(TokenType::Outer);
                    self.expect(TokenType::Join)?;
                    JoinType::Full
                }
                TokenType::Cross => {
                    self.advance();
                    // T-SQL `CROSS APPLY <source>` ≈ `CROSS JOIN LATERAL ...`.
                    if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("APPLY") {
                        self.advance();
                        JoinType::Cross
                    } else {
                        self.expect(TokenType::Join)?;
                        JoinType::Cross
                    }
                }
                TokenType::Outer => {
                    // T-SQL `OUTER APPLY <source>` ≈ `LEFT JOIN LATERAL ... ON TRUE`.
                    self.advance();
                    if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("APPLY") {
                        self.advance();
                        JoinType::Left
                    } else {
                        break;
                    }
                }
                _ => break,
            };

            let table = self.parse_table_source()?;
            let mut on = None;
            let mut using = vec![];

            if self.match_token(TokenType::On) {
                on = Some(self.parse_expr()?);
            } else if self.match_token(TokenType::Using) {
                // ClickHouse permits a bare column name without parens:
                // `JOIN t USING k`.
                if self.match_token(TokenType::LParen) {
                    using = vec![self.expect_name()?];
                    while self.match_token(TokenType::Comma) {
                        using.push(self.expect_name()?);
                    }
                    self.expect(TokenType::RParen)?;
                } else {
                    using = vec![self.expect_name()?];
                    while self.match_token(TokenType::Comma) {
                        if !self.is_name_token() {
                            break;
                        }
                        using.push(self.expect_name()?);
                    }
                }
            }

            joins.push(JoinClause {
                join_type,
                table,
                on,
                using,
            });
        }
        Ok(joins)
    }

    fn parse_order_by_items(&mut self) -> Result<Vec<OrderByItem>> {
        let mut items = Vec::new();
        // DuckDB / Snowflake `ORDER BY ALL` shortcut.
        if self.match_token(TokenType::All) {
            let ascending = if self.match_token(TokenType::Desc) {
                false
            } else {
                let _ = self.match_token(TokenType::Asc);
                true
            };
            items.push(OrderByItem {
                expr: Expr::Wildcard,
                ascending,
                nulls_first: None,
            });
            return Ok(items);
        }
        loop {
            // MySQL: `ORDER BY BINARY col [ASC|DESC]` — BINARY here is a
            // collation modifier on the sort key. Swallow it; the rest of
            // the expression parses normally.
            if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("BINARY") {
                let saved = self.pos;
                self.advance();
                // Only consume BINARY when followed by something that can
                // start an order-by key (name, literal, paren, etc.); if it
                // looks like the end of the list, rewind.
                if matches!(
                    self.peek_type(),
                    TokenType::Comma | TokenType::Semicolon | TokenType::Eof | TokenType::RParen
                ) {
                    self.pos = saved;
                }
            }
            let expr = self.parse_expr()?;
            // ClickHouse: `ORDER BY expr AS alias`. Swallow the alias.
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            let ascending = if self.match_token(TokenType::Desc) {
                false
            } else {
                let _ = self.match_token(TokenType::Asc);
                true
            };

            let nulls_first = if self.match_token(TokenType::Nulls) {
                if self.match_token(TokenType::First) {
                    Some(true)
                } else {
                    self.expect(TokenType::Identifier)?; // LAST
                    Some(false)
                }
            } else {
                None
            };

            items.push(OrderByItem {
                expr,
                ascending,
                nulls_first,
            });
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }
        Ok(items)
    }

    fn parse_expr_list(&mut self) -> Result<Vec<Expr>> {
        let mut exprs = vec![self.parse_expr()?];
        while self.match_token(TokenType::Comma) {
            // Tolerate a trailing comma — DuckDB / PostgreSQL accept
            // `IN ('a', 'b', )` and similar list shapes.
            if matches!(self.peek_type(), TokenType::RParen | TokenType::RBracket) {
                break;
            }
            exprs.push(self.parse_expr()?);
        }
        Ok(exprs)
    }

    /// Parse a comma-separated expression list where each item may carry an
    /// inline alias (`expr AS name` or `expr name`). Used for dialects (notably
    /// ClickHouse) that permit aliases inside partition/grouping lists.
    fn parse_expr_list_allow_item_alias(&mut self) -> Result<Vec<Expr>> {
        let mut exprs = Vec::new();
        loop {
            exprs.push(self.parse_expr()?);
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            if !self.match_token(TokenType::Comma) {
                break;
            }
            if matches!(self.peek_type(), TokenType::RParen | TokenType::RBracket) {
                break;
            }
        }
        Ok(exprs)
    }

    /// Parse array-literal elements: comma-separated expressions, each
    /// optionally followed by `AS alias` (ClickHouse lets bindings
    /// appear inside `[…]`). The closing token is the caller's
    /// responsibility.
    fn parse_array_items(&mut self, close: TokenType) -> Result<Vec<Expr>> {
        if self.peek_type() == &close {
            return Ok(vec![]);
        }
        let mut items = Vec::new();
        loop {
            let expr = self.parse_expr()?;
            if self.match_token(TokenType::As) {
                let _ = self.parse_optional_alias();
            }
            items.push(expr);
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }
        Ok(items)
    }

    /// Parse a GROUP BY list, which may contain regular expressions,
    /// CUBE(...), ROLLUP(...), and GROUPING SETS(...).
    fn parse_group_by_list(&mut self) -> Result<Vec<Expr>> {
        // DuckDB / Snowflake `GROUP BY ALL` shortcut — emit a wildcard
        // marker so downstream code can recognise it. PostgreSQL also
        // allows `GROUP BY ALL <col>, <col>` (treated identically to a
        // regular GROUP BY list); fall through to the normal parser when
        // the next token is a column expression rather than a clause
        // terminator.
        if self.match_token(TokenType::All) {
            let terminates = matches!(
                self.peek_type(),
                TokenType::Comma
                    | TokenType::Semicolon
                    | TokenType::Eof
                    | TokenType::RParen
                    | TokenType::Having
                    | TokenType::Order
                    | TokenType::Limit
                    | TokenType::Offset
                    | TokenType::Window
                    | TokenType::Union
                    | TokenType::Intersect
                    | TokenType::Except
                    | TokenType::Qualify
            );
            if terminates {
                return Ok(vec![Expr::Wildcard]);
            }
            // Followed by a real grouping expression — fall through.
        }
        let mut items = vec![self.parse_group_by_item()?];
        // ClickHouse: `GROUP BY col AS alias [, …]` — swallow alias.
        if self.match_token(TokenType::As) && self.is_name_token() {
            self.advance();
        }
        // MySQL: `GROUP BY col ASC|DESC [, …]` — swallow direction.
        let _ = self.match_token(TokenType::Asc) || self.match_token(TokenType::Desc);
        while self.match_token(TokenType::Comma) {
            items.push(self.parse_group_by_item()?);
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            let _ = self.match_token(TokenType::Asc) || self.match_token(TokenType::Desc);
        }
        Ok(items)
    }

    /// Parse a single GROUP BY item: a CUBE, ROLLUP, GROUPING SETS, or regular expression.
    fn parse_group_by_item(&mut self) -> Result<Expr> {
        match self.peek_type() {
            TokenType::Cube => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let exprs = if self.peek_type() == &TokenType::RParen {
                    vec![]
                } else {
                    self.parse_group_by_element_list()?
                };
                self.expect(TokenType::RParen)?;
                Ok(Expr::Cube { exprs })
            }
            TokenType::Rollup => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let exprs = if self.peek_type() == &TokenType::RParen {
                    vec![]
                } else {
                    self.parse_group_by_element_list()?
                };
                self.expect(TokenType::RParen)?;
                Ok(Expr::Rollup { exprs })
            }
            TokenType::Grouping => {
                // Could be GROUPING SETS or GROUPING() function
                let saved = self.pos;
                self.advance();
                if self.peek_type() == &TokenType::Sets {
                    // GROUPING SETS (...)
                    self.advance();
                    self.expect(TokenType::LParen)?;
                    let sets = self.parse_grouping_sets_elements()?;
                    self.expect(TokenType::RParen)?;
                    Ok(Expr::GroupingSets { sets })
                } else {
                    // It's the GROUPING() function, backtrack and parse as expression
                    self.pos = saved;
                    self.parse_expr()
                }
            }
            _ => self.parse_expr(),
        }
    }

    /// Parse elements inside CUBE(...) or ROLLUP(...).
    /// Each element can be a single expression or a parenthesized tuple of expressions.
    fn parse_group_by_element_list(&mut self) -> Result<Vec<Expr>> {
        let mut items = vec![self.parse_group_by_element()?];
        while self.match_token(TokenType::Comma) {
            items.push(self.parse_group_by_element()?);
        }
        Ok(items)
    }

    /// Parse a single element inside CUBE/ROLLUP: either `expr` or `(expr, expr, ...)`.
    fn parse_group_by_element(&mut self) -> Result<Expr> {
        if self.peek_type() == &TokenType::LParen {
            self.advance();
            let exprs = self.parse_expr_list()?;
            self.expect(TokenType::RParen)?;
            if exprs.len() == 1 {
                Ok(Expr::Nested(Box::new(exprs.into_iter().next().unwrap())))
            } else {
                Ok(Expr::Tuple(exprs))
            }
        } else {
            let e = self.parse_expr()?;
            // ClickHouse: `GROUP BY expr AS alias`. Swallow the alias.
            if self.match_token(TokenType::As) && self.is_name_token() {
                self.advance();
            }
            Ok(e)
        }
    }

    /// Parse elements inside GROUPING SETS (...).
    /// Each element can be: (), (expr, ...), CUBE(...), ROLLUP(...), or a single expr.
    fn parse_grouping_sets_elements(&mut self) -> Result<Vec<Expr>> {
        let mut items = vec![self.parse_grouping_sets_element()?];
        while self.match_token(TokenType::Comma) {
            items.push(self.parse_grouping_sets_element()?);
        }
        Ok(items)
    }

    /// Parse a single GROUPING SETS element.
    fn parse_grouping_sets_element(&mut self) -> Result<Expr> {
        match self.peek_type() {
            TokenType::Cube => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let exprs = if self.peek_type() == &TokenType::RParen {
                    vec![]
                } else {
                    self.parse_group_by_element_list()?
                };
                self.expect(TokenType::RParen)?;
                Ok(Expr::Cube { exprs })
            }
            TokenType::Rollup => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let exprs = if self.peek_type() == &TokenType::RParen {
                    vec![]
                } else {
                    self.parse_group_by_element_list()?
                };
                self.expect(TokenType::RParen)?;
                Ok(Expr::Rollup { exprs })
            }
            TokenType::LParen => {
                self.advance();
                if self.peek_type() == &TokenType::RParen {
                    // Empty grouping set: ()
                    self.advance();
                    Ok(Expr::Tuple(vec![]))
                } else {
                    let exprs = self.parse_expr_list()?;
                    self.expect(TokenType::RParen)?;
                    if exprs.len() == 1 {
                        Ok(Expr::Nested(Box::new(exprs.into_iter().next().unwrap())))
                    } else {
                        Ok(Expr::Tuple(exprs))
                    }
                }
            }
            _ => self.parse_expr(),
        }
    }

    // ── INSERT ──────────────────────────────────────────────────────

    fn parse_insert(&mut self) -> Result<InsertStatement> {
        // Accept MySQL `REPLACE INTO ...` as a synonym for `INSERT INTO ...`.
        if !self.match_token(TokenType::Insert) {
            self.expect(TokenType::Replace)?;
        }
        // SQLite / DuckDB conflict-resolution prefix:
        //   `INSERT OR REPLACE|IGNORE|FAIL|ABORT|ROLLBACK INTO ...`.
        // Swallow opaquely; we don't model conflict resolution at the
        // statement level (ON CONFLICT covers most cases downstream).
        if self.match_token(TokenType::Or) {
            if self.match_token(TokenType::Replace) {
                // matched
            } else if self.match_token(TokenType::Ignore) {
                // matched
            } else if self.is_name_token() {
                let v = self.peek().value.to_uppercase();
                if matches!(v.as_str(), "FAIL" | "ABORT" | "ROLLBACK") {
                    self.advance();
                }
            }
        }
        // MySQL modifiers between INSERT/REPLACE and INTO:
        //   `INSERT LOW_PRIORITY|DELAYED|HIGH_PRIORITY [IGNORE] INTO ...`,
        //   `INSERT IGNORE INTO ...`. Swallow them so the rest parses.
        loop {
            if self.match_token(TokenType::Ignore) {
                continue;
            }
            if self.is_name_token() {
                let v = self.peek().value.to_uppercase();
                if matches!(v.as_str(), "LOW_PRIORITY" | "DELAYED" | "HIGH_PRIORITY") {
                    self.advance();
                    continue;
                }
            }
            break;
        }
        let _ = self.match_token(TokenType::Into);
        // Hive: `INSERT OVERWRITE [LOCAL] DIRECTORY '/path'` or
        // `INSERT OVERWRITE TABLE tbl ...`. Consume OVERWRITE (tokenized as
        // an identifier) and any DIRECTORY clause that follows.
        if self.check_keyword("OVERWRITE") {
            self.advance();
            if self.check_keyword("LOCAL") {
                self.advance();
            }
            if self.check_keyword("DIRECTORY") {
                self.advance();
                // Consume `'path'` (string) and any STORED AS / ROW FORMAT
                // clauses until we hit SELECT/WITH/LParen/VALUES/EOF.
                if matches!(self.peek_type(), TokenType::String) {
                    self.advance();
                }
                while !matches!(
                    self.peek_type(),
                    TokenType::Select
                        | TokenType::With
                        | TokenType::LParen
                        | TokenType::Values
                        | TokenType::Eof
                        | TokenType::Semicolon
                ) {
                    self.advance();
                }
            }
        }
        // Hive: `INSERT INTO TABLE tbl ...` and `INSERT OVERWRITE TABLE tbl ...`.
        let _ = self.match_token(TokenType::Table);
        let table = self.parse_table_ref()?;

        // Hive `PARTITION (k=v, ...)` between table and column list / source.
        if self.peek_type() == &TokenType::Partition {
            self.advance();
            if self.match_token(TokenType::LParen) {
                let mut depth = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth -= 1,
                        TokenType::Eof => break,
                        _ => {}
                    }
                    if depth == 0 {
                        self.advance();
                        break;
                    }
                    self.advance();
                }
            }
        }

        let columns = if self.match_token(TokenType::LParen) {
            // BigQuery / SQLFluff fixture: `INSERT INTO t (SELECT ... )` —
            // no column list, the parenthesized SELECT is the source.
            // Rewind to the `(` and let the source dispatch handle it.
            if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                self.pos -= 1;
                Vec::new()
            } else {
                // ClickHouse `INSERT INTO t (COLUMNS('.*') EXCEPT (...))` — when
                // the list contains a function call or anything other than plain
                // identifiers, fall back to a balanced-paren swallow.
                let saved = self.pos;
                let try_simple: Result<Vec<String>> = (|| {
                    let mut cols = vec![self.parse_dotted_name()?];
                    while self.match_token(TokenType::Comma) {
                        cols.push(self.parse_dotted_name()?);
                    }
                    self.expect(TokenType::RParen)?;
                    Ok(cols)
                })();
                match try_simple {
                    Ok(c) => c,
                    Err(_) => {
                        self.pos = saved;
                        let mut depth = 1_i32;
                        while depth > 0 && self.peek_type() != &TokenType::Eof {
                            match self.peek_type() {
                                TokenType::LParen => depth += 1,
                                TokenType::RParen => depth -= 1,
                                _ => {}
                            }
                            self.advance();
                        }
                        Vec::new()
                    }
                }
            }
        } else {
            vec![]
        };

        // ClickHouse `INSERT INTO t [(cols)] SETTINGS k=v[, …] VALUES …`.
        // Swallow the SETTINGS clause before the source clause so the
        // surrounding parse completes.
        if self.check_keyword("SETTINGS") {
            self.advance();
            loop {
                if !self.is_name_token() {
                    break;
                }
                self.advance(); // key
                if !self.match_token(TokenType::Eq) {
                    break;
                }
                // value: number / string / identifier / unary-signed number
                let _ = self.match_token(TokenType::Minus) || self.match_token(TokenType::Plus);
                if matches!(self.peek_type(), TokenType::Number | TokenType::String)
                    || self.is_name_token()
                {
                    self.advance();
                }
                if !self.match_token(TokenType::Comma) {
                    break;
                }
            }
        }

        let source = if self.match_token(TokenType::Values) || self.match_keyword("VALUE") {
            let mut rows = Vec::new();
            loop {
                self.expect(TokenType::LParen)?;
                // MySQL allows `VALUES ()` as an empty row to insert all
                // defaults — accept and emit as an empty row.
                let row = if self.peek_type() == &TokenType::RParen {
                    Vec::new()
                } else {
                    self.parse_expr_list()?
                };
                self.expect(TokenType::RParen)?;
                rows.push(row);
                // ClickHouse permits comma-less rows: `VALUES (1)(2)(3)`.
                if self.peek_type() == &TokenType::LParen {
                    continue;
                }
                if !self.match_token(TokenType::Comma) {
                    break;
                }
                // Trailing comma: `VALUES (1,2), (3,4),` — DuckDB / sqlfluff
                // fixture truncation. Accept and stop the row loop.
                if !matches!(self.peek_type(), TokenType::LParen) {
                    break;
                }
            }
            InsertSource::Values(rows)
        } else if matches!(
            self.peek_type(),
            TokenType::Select | TokenType::With | TokenType::LParen
        ) {
            InsertSource::Query(Box::new(self.parse_statement_inner()?))
        } else if self.match_token(TokenType::Default) {
            self.expect(TokenType::Values)?;
            InsertSource::Default
        } else if self.match_token(TokenType::Set) {
            // MySQL `INSERT INTO t SET col = val, col = val, ...`.
            // Collapse into a single-row VALUES placeholder by collecting
            // the right-hand expressions; column names are dropped.
            let mut row = Vec::new();
            loop {
                let _ = self.expect_name()?;
                self.expect(TokenType::Eq)?;
                row.push(self.parse_expr()?);
                if !self.match_token(TokenType::Comma) {
                    break;
                }
            }
            InsertSource::Values(vec![row])
        } else if self.peek_type() == &TokenType::From {
            // DuckDB `INSERT INTO t FROM source` shorthand for
            // `INSERT INTO t SELECT * FROM source`. Synthesize a SELECT *
            // statement so the existing query path handles it.
            self.advance();
            let from = Some(FromClause {
                source: self.parse_table_source()?,
            });
            let joins = self.parse_joins()?;
            let stmt = Statement::Select(SelectStatement {
                comments: vec![],
                ctes: vec![],
                distinct: false,
                top: None,
                columns: vec![SelectItem::Wildcard],
                from,
                joins,
                where_clause: None,
                group_by: vec![],
                having: None,
                order_by: vec![],
                limit: None,
                offset: None,
                fetch_first: None,
                qualify: None,
                window_definitions: vec![],
            });
            InsertSource::Query(Box::new(stmt))
        } else if self.peek().value.eq_ignore_ascii_case("FORMAT") {
            // ClickHouse `INSERT INTO t FORMAT name <raw payload>`.
            // Swallow the format name and the remainder of the statement
            // as opaque bytes; we cannot parse JSONEachRow / TabSeparated
            // payloads, but we should not reject the statement.
            self.advance();
            let _ = self.expect_name();
            while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon) {
                self.advance();
            }
            InsertSource::Default
        } else {
            return Err(SqlglotError::ParserError {
                message: "Expected VALUES, SELECT, or DEFAULT VALUES after INSERT".into(),
            });
        };

        // MySQL 8.0.19+ row alias: `INSERT INTO t (cols) VALUES (...) AS
        // alias [(col_alias, ...)] ON DUPLICATE KEY UPDATE ...`. Swallow
        // the alias so the ON DUPLICATE clause parses.
        if self.peek_type() == &TokenType::As
            && self
                .peek_offset(1)
                .map(|t| {
                    matches!(
                        t.token_type,
                        TokenType::Identifier
                            | TokenType::Key
                            | TokenType::Year
                            | TokenType::Month
                            | TokenType::Day
                            | TokenType::Hour
                            | TokenType::Minute
                            | TokenType::Second
                    ) || t
                        .value
                        .chars()
                        .next()
                        .is_some_and(|c| c.is_alphabetic() || c == '_')
                })
                .unwrap_or(false)
        {
            self.advance(); // AS
            self.advance(); // alias name
            if self.match_token(TokenType::LParen) {
                let mut depth = 1_i32;
                while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth -= 1,
                        _ => {}
                    }
                    self.advance();
                }
            }
        }

        // MySQL `ON DUPLICATE KEY UPDATE col=val, ...`. Swallow the clause.
        if self.peek_type() == &TokenType::On
            && self
                .peek_offset(1)
                .map(|t| t.value.eq_ignore_ascii_case("DUPLICATE"))
                .unwrap_or(false)
        {
            self.advance();
            self.advance();
            // KEY UPDATE
            if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("KEY") {
                self.advance();
            }
            if self.match_token(TokenType::Update) {
                // assignments until end-of-statement
                loop {
                    let _ = self.expect_name();
                    if !self.match_token(TokenType::Eq) {
                        break;
                    }
                    let _ = self.parse_expr();
                    if !self.match_token(TokenType::Comma) {
                        break;
                    }
                }
            }
        }

        // ON CONFLICT
        let on_conflict = if self.match_token(TokenType::On) {
            if self.match_token(TokenType::Conflict) {
                let columns = if self.match_token(TokenType::LParen) {
                    self.parse_parenthesized_raw_items()?
                } else {
                    vec![]
                };
                self.expect(TokenType::Do)?;
                let action = if self.match_token(TokenType::Nothing) {
                    ConflictAction::DoNothing
                } else {
                    self.expect(TokenType::Update)?;
                    self.expect(TokenType::Set)?;
                    let mut assignments = Vec::new();
                    loop {
                        let col = self.expect_name()?;
                        self.expect(TokenType::Eq)?;
                        let val = self.parse_expr()?;
                        assignments.push((col, val));
                        if !self.match_token(TokenType::Comma) {
                            break;
                        }
                    }
                    ConflictAction::DoUpdate(assignments)
                };
                // Postgres / DuckDB allow `ON CONFLICT (...) DO UPDATE SET
                // ... WHERE predicate` to limit the update. Swallow the
                // predicate opaquely.
                if self.match_token(TokenType::Where) {
                    let _ = self.parse_expr()?;
                }
                Some(OnConflict { columns, action })
            } else {
                None
            }
        } else {
            None
        };

        let returning = if self.match_token(TokenType::Returning) {
            self.parse_select_items()?
        } else {
            vec![]
        };

        Ok(InsertStatement {
            comments: vec![],
            table,
            columns,
            source,
            on_conflict,
            returning,
        })
    }

    // ── UPDATE ──────────────────────────────────────────────────────

    fn parse_update(&mut self) -> Result<UpdateStatement> {
        self.expect(TokenType::Update)?;
        let table = self.parse_table_ref()?;
        // MySQL multi-table UPDATE: `UPDATE t1, t2 [, ...] SET ...`.
        // Swallow the additional table refs (we keep only the first as
        // the primary target).
        while self.match_token(TokenType::Comma) {
            let _ = self.parse_table_ref()?;
        }
        // PG SQL:2011 temporal `UPDATE t FOR PORTION OF col FROM a TO b
        // [AS alias] SET ...`. Swallow the qualifier verbatim.
        if self.check_keyword("FOR")
            && self
                .peek_offset(1)
                .map(|t| t.value.eq_ignore_ascii_case("PORTION"))
                .unwrap_or(false)
        {
            while !matches!(
                self.peek_type(),
                TokenType::Set | TokenType::Eof | TokenType::Semicolon
            ) {
                self.advance();
            }
        }
        // MySQL `UPDATE t PARTITION (p0[, p1]) SET ...` — swallow.
        if matches!(self.peek_type(), TokenType::Partition)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::LParen)
            )
        {
            self.advance();
            self.advance();
            let mut depth = 1;
            while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    _ => {}
                }
                self.advance();
            }
        }
        // MySQL multi-table UPDATE: `UPDATE t1 [LEFT|RIGHT|INNER|CROSS] JOIN
        // t2 ON ... SET ...`. Swallow the joins so the existing single-target
        // update parses; the joined tables are dropped from the AST.
        let _ = self.parse_joins();
        self.expect(TokenType::Set)?;

        let mut assignments = Vec::new();
        loop {
            // Accept qualified LHS like `alias.col` (Oracle, T-SQL idiom),
            // and PG/Snowflake subscripts/field access on the LHS such as
            // `arr[1] = …`, `arr[1:3] = …`, `obj['k']`, `(a,b) = …`.
            // Accept LHS row-tuple `(a, b, c) = (rhs)` (PostgreSQL).
            if self.peek_type() == &TokenType::LParen {
                let saved = self.pos;
                self.advance();
                let mut depth = 1;
                while depth > 0 && self.peek_type() != &TokenType::Eof {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth -= 1,
                        _ => {}
                    }
                    self.advance();
                }
                if self.peek_type() == &TokenType::Eq {
                    self.advance();
                    let val = self.parse_expr()?;
                    assignments.push(("__tuple__".to_string(), val));
                    if !self.match_token(TokenType::Comma) {
                        break;
                    }
                    continue;
                }
                self.pos = saved;
            }
            let mut col = self.expect_name()?;
            while self.match_token(TokenType::Dot) {
                col.push('.');
                col.push_str(&self.expect_name()?);
            }
            // Swallow `[index]` / `[a:b]` subscripts in the LHS — we don't
            // model array-element assignment in the AST.
            while self.peek_type() == &TokenType::LBracket {
                self.advance();
                let mut depth = 1;
                while depth > 0 && self.peek_type() != &TokenType::Eof {
                    match self.peek_type() {
                        TokenType::LBracket => depth += 1,
                        TokenType::RBracket => depth -= 1,
                        _ => {}
                    }
                    self.advance();
                }
            }
            self.expect(TokenType::Eq)?;
            let val = self.parse_expr()?;
            assignments.push((col, val));
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }

        let from = if self.match_token(TokenType::From) {
            Some(FromClause {
                source: self.parse_table_source()?,
            })
        } else {
            None
        };

        let where_clause = if self.match_token(TokenType::Where) {
            Some(self.parse_expr()?)
        } else {
            None
        };

        // Teradata `PREFERRING <expr> [PARTITION BY <list>]` skyline
        // clause on UPDATE. Swallow up to a known terminator.
        if self.check_keyword("PREFERRING") {
            self.advance();
            loop {
                match self.peek_type() {
                    TokenType::Eof
                    | TokenType::Semicolon
                    | TokenType::RParen
                    | TokenType::Returning => break,
                    _ => self.advance(),
                };
            }
        }

        // MySQL: `UPDATE … [ORDER BY …] [LIMIT N]`. Swallow.
        if self.match_token(TokenType::Order) {
            self.expect(TokenType::By)?;
            let _ = self.parse_order_by_items()?;
        }
        if self.match_token(TokenType::Limit) {
            let _ = self.parse_expr()?;
        }

        let returning = if self.match_token(TokenType::Returning) {
            self.parse_select_items()?
        } else {
            vec![]
        };

        Ok(UpdateStatement {
            comments: vec![],
            table,
            assignments,
            from,
            where_clause,
            returning,
        })
    }

    // ── DELETE ──────────────────────────────────────────────────────

    fn parse_delete(&mut self) -> Result<DeleteStatement> {
        self.expect(TokenType::Delete)?;
        // MySQL multi-table form: `DELETE t1[, t2, ...] FROM <join expr>`.
        // Swallow the leading table-alias list (we don't model it) before
        // the mandatory FROM.
        let mut multi_table = false;
        if !matches!(self.peek_type(), TokenType::From) {
            let saved = self.pos;
            if self.is_name_token() {
                self.advance();
                let _ = self.match_token(TokenType::Dot);
                if self.is_name_token() {
                    self.advance();
                }
                while self.match_token(TokenType::Comma) {
                    if !self.is_name_token() {
                        break;
                    }
                    self.advance();
                    let _ = self.match_token(TokenType::Dot);
                    if self.is_name_token() {
                        self.advance();
                    }
                }
                if matches!(self.peek_type(), TokenType::From) {
                    multi_table = true;
                } else {
                    self.pos = saved;
                }
            }
        }
        // BigQuery / some Snowflake forms allow `DELETE <table> WHERE …`
        // (FROM optional). If FROM is missing but the next token starts a
        // table-ref, treat it as the implicit FROM target.
        let from_optional = !matches!(self.peek_type(), TokenType::From);
        if !from_optional {
            self.expect(TokenType::From)?;
        }
        let table = self.parse_table_ref()?;
        // MySQL: `DELETE FROM t PARTITION (p0[, p1, ...])` — swallow
        // partition selector.
        if matches!(self.peek_type(), TokenType::Partition)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::LParen)
            )
        {
            self.advance();
            self.advance();
            let mut depth = 1;
            while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    _ => {}
                }
                self.advance();
            }
        }
        if multi_table {
            // Swallow JOIN clauses, additional comma-joined tables, and
            // any opaque tail up to USING / WHERE / RETURNING / ; / EOF.
            loop {
                if matches!(
                    self.peek_type(),
                    TokenType::Where
                        | TokenType::Using
                        | TokenType::Returning
                        | TokenType::Semicolon
                        | TokenType::Eof
                ) {
                    break;
                }
                self.advance();
            }
        }

        let using = if self.match_token(TokenType::Using) {
            Some(FromClause {
                source: self.parse_table_source()?,
            })
        } else {
            None
        };

        // Teradata `PREFERRING <expr> [PARTITION BY <list>]` skyline
        // clause on DELETE.
        if self.check_keyword("PREFERRING") {
            self.advance();
            loop {
                match self.peek_type() {
                    TokenType::Eof
                    | TokenType::Semicolon
                    | TokenType::Where
                    | TokenType::Returning
                    | TokenType::RParen => break,
                    _ => self.advance(),
                };
            }
        }

        let where_clause = if self.match_token(TokenType::Where) {
            Some(self.parse_expr()?)
        } else {
            None
        };

        // MySQL: `DELETE FROM tbl [WHERE ...] [ORDER BY ...] [LIMIT N]`.
        // Swallow ORDER BY and LIMIT modifiers — we don't model them on
        // DeleteStatement yet.
        if self.match_token(TokenType::Order) {
            self.expect(TokenType::By)?;
            let _ = self.parse_order_by_items()?;
        }
        if self.match_token(TokenType::Limit) {
            let _ = self.parse_expr()?;
        }

        let returning = if self.match_token(TokenType::Returning) {
            self.parse_select_items()?
        } else {
            vec![]
        };

        Ok(DeleteStatement {
            comments: vec![],
            table,
            using,
            where_clause,
            returning,
        })
    }

    // ── MERGE ───────────────────────────────────────────────────────

    fn parse_merge(&mut self) -> Result<MergeStatement> {
        self.expect(TokenType::Merge)?;
        let _ = self.match_token(TokenType::Into);
        let target = self.parse_table_ref()?;

        self.expect(TokenType::Using)?;
        let source = self.parse_table_source()?;

        // DuckDB supports `MERGE INTO t USING src USING (cols)` as a
        // shorthand for the ON condition (column-equality join, akin to
        // SQL USING for JOINs). Swallow the column list opaquely and
        // synthesize a trivial truthy ON expression so downstream parsing
        // continues. We don't model USING-style MERGE in the AST yet.
        let on = if self.match_token(TokenType::Using) {
            self.expect(TokenType::LParen)?;
            let _ = self.expect_name()?;
            while self.match_token(TokenType::Comma) {
                let _ = self.expect_name()?;
            }
            self.expect(TokenType::RParen)?;
            Expr::Boolean(true)
        } else {
            self.expect(TokenType::On)?;
            self.parse_expr()?
        };

        let mut clauses = Vec::new();
        while self.match_token(TokenType::When) {
            clauses.push(self.parse_merge_clause()?);
        }

        if clauses.is_empty() {
            return Err(SqlglotError::ParserError {
                message: "MERGE requires at least one WHEN clause".into(),
            });
        }

        // OUTPUT clause (T-SQL extension)
        let output = if self.match_keyword("OUTPUT") {
            self.parse_select_items()?
        } else {
            vec![]
        };

        // PostgreSQL: `MERGE … RETURNING <select_list>`. We don't yet model
        // RETURNING for MERGE, so swallow the items and discard them.
        if self.match_token(TokenType::Returning) {
            let _ = self.parse_select_items()?;
        }

        Ok(MergeStatement {
            comments: vec![],
            target,
            source,
            on,
            clauses,
            output,
        })
    }

    fn parse_merge_clause(&mut self) -> Result<MergeClause> {
        let kind = if self.match_token(TokenType::Not) {
            self.expect(TokenType::Matched)?;
            if self.match_keyword("BY") {
                if self.match_keyword("SOURCE") {
                    MergeClauseKind::NotMatchedBySource
                } else {
                    // BY TARGET is the default / explicit form
                    let _ = self.match_keyword("TARGET");
                    MergeClauseKind::NotMatched
                }
            } else {
                MergeClauseKind::NotMatched
            }
        } else {
            self.expect(TokenType::Matched)?;
            MergeClauseKind::Matched
        };

        let condition = if self.match_token(TokenType::And) {
            Some(self.parse_expr()?)
        } else {
            None
        };

        self.expect(TokenType::Then)?;

        let action = self.parse_merge_action(&kind)?;

        Ok(MergeClause {
            kind,
            condition,
            action,
        })
    }

    fn parse_merge_action(&mut self, kind: &MergeClauseKind) -> Result<MergeAction> {
        if self.match_token(TokenType::Update) {
            self.expect(TokenType::Set)?;
            let mut assignments = Vec::new();
            loop {
                let mut col = self.expect_name()?;
                // Support dotted column names like target.col
                while self.match_token(TokenType::Dot) {
                    col.push('.');
                    col.push_str(&self.expect_name()?);
                }
                self.expect(TokenType::Eq)?;
                let val = self.parse_expr()?;
                assignments.push((col, val));
                if !self.match_token(TokenType::Comma) {
                    break;
                }
            }
            Ok(MergeAction::Update(assignments))
        } else if self.match_token(TokenType::Insert) {
            // INSERT ROW (BigQuery)
            if self.match_keyword("ROW") {
                return Ok(MergeAction::InsertRow);
            }

            let columns = if self.match_token(TokenType::LParen) {
                let mut cols = vec![self.expect_name()?];
                while self.match_token(TokenType::Comma) {
                    cols.push(self.expect_name()?);
                }
                self.expect(TokenType::RParen)?;
                cols
            } else {
                vec![]
            };

            self.expect(TokenType::Values)?;
            self.expect(TokenType::LParen)?;
            let values = self.parse_expr_list()?;
            self.expect(TokenType::RParen)?;

            Ok(MergeAction::Insert { columns, values })
        } else if self.match_token(TokenType::Delete) {
            Ok(MergeAction::Delete)
        } else {
            Err(SqlglotError::ParserError {
                message: format!(
                    "Expected UPDATE, INSERT, or DELETE after WHEN {} THEN",
                    match kind {
                        MergeClauseKind::Matched => "MATCHED",
                        MergeClauseKind::NotMatched => "NOT MATCHED",
                        MergeClauseKind::NotMatchedBySource => "NOT MATCHED BY SOURCE",
                    }
                ),
            })
        }
    }

    // ── CREATE ──────────────────────────────────────────────────────

    fn parse_create(&mut self) -> Result<Statement> {
        self.expect(TokenType::Create)?;

        let or_replace = if self.check_keyword("OR") {
            self.advance();
            self.expect(TokenType::Replace)?;
            true
        } else {
            false
        };

        let temporary = self.match_token(TokenType::Temporary) || self.match_token(TokenType::Temp);

        let materialized = self.match_token(TokenType::Materialized);

        if self.match_token(TokenType::View) {
            return self
                .parse_create_view(or_replace, materialized)
                .map(Statement::CreateView);
        }

        self.expect(TokenType::Table)?;

        let if_not_exists = if self.match_token(TokenType::If) {
            self.expect(TokenType::Not)?;
            self.expect(TokenType::Exists)?;
            true
        } else {
            false
        };

        let table = self.parse_table_ref_no_alias()?;

        // CREATE TABLE ... AS SELECT ...
        if self.match_token(TokenType::As) {
            let query = self.parse_statement_inner()?;
            // Greenplum / Citus / etc. trailing `DISTRIBUTED BY (...)` /
            // `DISTRIBUTED RANDOMLY` / `DISTRIBUTED REPLICATED`. Swallow.
            if self.check_keyword("DISTRIBUTED") {
                self.advance();
                if self.check_keyword("BY") || matches!(self.peek_type(), TokenType::By) {
                    self.advance();
                    if self.match_token(TokenType::LParen) {
                        let mut depth = 1;
                        while depth > 0 {
                            match self.peek_type() {
                                TokenType::LParen => depth += 1,
                                TokenType::RParen => {
                                    depth -= 1;
                                    if depth == 0 {
                                        self.advance();
                                        break;
                                    }
                                }
                                TokenType::Eof => break,
                                _ => {}
                            }
                            self.advance();
                        }
                    }
                } else if self.is_name_token() {
                    // RANDOMLY / REPLICATED — single keyword
                    self.advance();
                }
            }
            return Ok(Statement::CreateTable(CreateTableStatement {
                comments: vec![],
                if_not_exists,
                temporary,
                table,
                columns: vec![],
                constraints: vec![],
                as_select: Some(Box::new(query)),
            }));
        }

        self.expect(TokenType::LParen)?;

        let mut columns = Vec::new();
        let mut constraints = Vec::new();

        loop {
            // Check for table-level constraints
            if matches!(
                self.peek_type(),
                TokenType::Primary
                    | TokenType::Unique
                    | TokenType::Foreign
                    | TokenType::Check
                    | TokenType::Constraint
            ) {
                constraints.push(self.parse_table_constraint()?);
            } else if self.peek_type() != &TokenType::RParen {
                columns.push(self.parse_column_def()?);
            }

            if !self.match_token(TokenType::Comma) {
                break;
            }
        }
        self.expect(TokenType::RParen)?;

        // Tolerate dialect-specific trailing clauses (ClickHouse `ENGINE = X`,
        // `ORDER BY (...)`, `PARTITION BY ...`, `SETTINGS ...`, MySQL
        // `ENGINE=InnoDB DEFAULT CHARSET=utf8`, etc.) by consuming tokens
        // until the next statement boundary. Respects paren depth so a
        // top-level `;` inside `ORDER BY (a, b)` is not mistaken for end.
        self.skip_trailing_options();

        Ok(Statement::CreateTable(CreateTableStatement {
            comments: vec![],
            if_not_exists,
            temporary,
            table,
            columns,
            constraints,
            as_select: None,
        }))
    }

    /// Discard tokens up to (but not including) a top-level `;` or EOF.
    /// Used to skip dialect-specific tail clauses we don't model in the AST
    /// (CREATE TABLE engines, options, etc.).
    fn skip_trailing_options(&mut self) {
        let mut depth: i32 = 0;
        loop {
            match self.peek_type() {
                TokenType::Eof => break,
                TokenType::Semicolon if depth == 0 => break,
                TokenType::LParen => {
                    depth += 1;
                    self.advance();
                }
                TokenType::RParen => {
                    depth -= 1;
                    if depth < 0 {
                        break;
                    }
                    self.advance();
                }
                _ => {
                    self.advance();
                }
            }
        }
    }

    fn parse_create_view(
        &mut self,
        or_replace: bool,
        materialized: bool,
    ) -> Result<CreateViewStatement> {
        let if_not_exists = if self.match_token(TokenType::If) {
            self.expect(TokenType::Not)?;
            self.expect(TokenType::Exists)?;
            true
        } else {
            false
        };

        // Parse name without alias (so AS is not consumed as an alias)
        let name = self.parse_table_ref_no_alias()?;

        let columns = if self.match_token(TokenType::LParen) {
            let mut cols = vec![self.expect_name()?];
            while self.match_token(TokenType::Comma) {
                cols.push(self.expect_name()?);
            }
            self.expect(TokenType::RParen)?;
            cols
        } else {
            vec![]
        };

        self.expect(TokenType::As)?;
        let query = self.parse_statement_inner()?;

        Ok(CreateViewStatement {
            comments: vec![],
            name,
            columns,
            query: Box::new(query),
            or_replace,
            materialized,
            if_not_exists,
        })
    }

    fn parse_table_constraint(&mut self) -> Result<TableConstraint> {
        let name = if self.match_token(TokenType::Constraint) {
            Some(self.expect_name()?)
        } else {
            None
        };

        if self.match_token(TokenType::Primary) {
            self.expect(TokenType::Key)?;
            self.expect(TokenType::LParen)?;
            let columns = self.parse_name_list()?;
            self.expect(TokenType::RParen)?;
            // TiDB / MySQL: `PRIMARY KEY (cols) GLOBAL|LOCAL` index scope
            // modifier and `USING BTREE|HASH` index-type modifier.
            if self.is_name_token()
                && matches!(
                    self.peek().value.to_uppercase().as_str(),
                    "GLOBAL" | "LOCAL"
                )
            {
                self.advance();
            }
            if self.match_token(TokenType::Using) && self.is_name_token() {
                self.advance();
            }
            self.swallow_constraint_modifiers();
            Ok(TableConstraint::PrimaryKey { name, columns })
        } else if self.match_token(TokenType::Unique) {
            let _ = self.match_token(TokenType::Index) || self.match_token(TokenType::Key);
            // Optional index name before `(`.
            if !matches!(self.peek_type(), TokenType::LParen) && self.is_name_token() {
                self.advance();
            }
            self.expect(TokenType::LParen)?;
            let columns = self.parse_name_list()?;
            self.expect(TokenType::RParen)?;
            if self.is_name_token()
                && matches!(
                    self.peek().value.to_uppercase().as_str(),
                    "GLOBAL" | "LOCAL"
                )
            {
                self.advance();
            }
            if self.match_token(TokenType::Using) && self.is_name_token() {
                self.advance();
            }
            self.swallow_constraint_modifiers();
            Ok(TableConstraint::Unique { name, columns })
        } else if self.match_token(TokenType::Foreign) {
            self.expect(TokenType::Key)?;
            self.expect(TokenType::LParen)?;
            let columns = self.parse_name_list()?;
            self.expect(TokenType::RParen)?;
            self.expect(TokenType::References)?;
            let ref_table = self.parse_table_ref()?;
            self.expect(TokenType::LParen)?;
            let ref_columns = self.parse_name_list()?;
            self.expect(TokenType::RParen)?;

            // PG / ANSI `MATCH FULL | PARTIAL | SIMPLE` clause — swallow.
            if self.check_keyword("MATCH") {
                self.advance();
                if self.is_name_token() {
                    self.advance();
                }
            }

            let mut on_delete = None;
            let mut on_update = None;
            // Accept ON DELETE / ON UPDATE clauses in any order. Match the
            // ON keyword only when the following token is DELETE / UPDATE
            // so a misplaced ON UPDATE doesn't consume the bare ON token
            // and orphan the rest of the action list.
            while self.peek_type() == &TokenType::On {
                let next = self.peek_offset(1).map(|t| &t.token_type);
                if matches!(next, Some(TokenType::Delete)) {
                    self.advance();
                    self.advance();
                    on_delete = Some(self.parse_referential_action()?);
                } else if matches!(next, Some(TokenType::Update)) {
                    self.advance();
                    self.advance();
                    on_update = Some(self.parse_referential_action()?);
                } else {
                    break;
                }
            }

            self.swallow_constraint_modifiers();
            Ok(TableConstraint::ForeignKey {
                name,
                columns,
                ref_table,
                ref_columns,
                on_delete,
                on_update,
            })
        } else if self.match_token(TokenType::Check) {
            self.expect(TokenType::LParen)?;
            let expr = self.parse_expr()?;
            self.expect(TokenType::RParen)?;
            self.swallow_constraint_modifiers();
            Ok(TableConstraint::Check { name, expr })
        } else {
            Err(SqlglotError::ParserError {
                message: "Expected constraint type".into(),
            })
        }
    }

    /// Swallow trailing constraint modifiers shared by FK / CHECK / PK /
    /// UNIQUE: `NOT VALID`, `[NOT] ENFORCED`, `DEFERRABLE`, `NOT DEFERRABLE`,
    /// `INITIALLY DEFERRED | IMMEDIATE`, `NO INHERIT`. Best-effort — we
    /// don't model them in the AST.
    fn swallow_constraint_modifiers(&mut self) {
        loop {
            if self.check_keyword("NOT")
                && self
                    .peek_offset(1)
                    .map(|t| t.value.to_uppercase())
                    .as_deref()
                    .is_some_and(|v| matches!(v, "VALID" | "ENFORCED" | "DEFERRABLE"))
            {
                self.advance();
                self.advance();
                continue;
            }
            if self.check_keyword("ENFORCED")
                || self.check_keyword("DEFERRABLE")
                || self.check_keyword("CLUSTERED")
                || self.check_keyword("NONCLUSTERED")
                || self.check_keyword("INVISIBLE")
                || self.check_keyword("VISIBLE")
            {
                self.advance();
                continue;
            }
            if self.check_keyword("INITIALLY") {
                self.advance();
                if self.is_name_token() {
                    self.advance();
                }
                continue;
            }
            if self.check_keyword("NO")
                && self
                    .peek_offset(1)
                    .map(|t| t.value.eq_ignore_ascii_case("INHERIT"))
                    .unwrap_or(false)
            {
                self.advance();
                self.advance();
                continue;
            }
            break;
        }
    }

    fn parse_referential_action(&mut self) -> Result<ReferentialAction> {
        if self.match_token(TokenType::Cascade) {
            Ok(ReferentialAction::Cascade)
        } else if self.match_token(TokenType::Restrict) {
            Ok(ReferentialAction::Restrict)
        } else if self.match_token(TokenType::Set) {
            if self.match_token(TokenType::Null) {
                Ok(ReferentialAction::SetNull)
            } else if self.match_token(TokenType::Default) {
                Ok(ReferentialAction::SetDefault)
            } else {
                Err(SqlglotError::ParserError {
                    message: "Expected NULL or DEFAULT after SET".into(),
                })
            }
        } else if self.check_keyword("NO") {
            self.advance();
            self.expect(TokenType::Identifier)?; // ACTION
            Ok(ReferentialAction::NoAction)
        } else {
            Err(SqlglotError::ParserError {
                message: "Expected referential action (CASCADE, RESTRICT, SET NULL, SET DEFAULT, NO ACTION)".into(),
            })
        }
    }

    fn parse_name_list(&mut self) -> Result<Vec<String>> {
        let mut names = vec![self.expect_name()?];
        while self.match_token(TokenType::Comma) {
            names.push(self.expect_name()?);
        }
        Ok(names)
    }

    /// Parse a dotted column reference for INSERT column lists:
    /// `name` or `parent.child` (ClickHouse nested columns).
    fn parse_dotted_name(&mut self) -> Result<String> {
        let mut name = self.expect_name()?;
        while self.peek_type() == &TokenType::Dot {
            let next = self.peek_offset(1).map(|t| t.token_type.clone());
            let next_is_namelike = matches!(
                next,
                Some(TokenType::Identifier)
                    | Some(TokenType::Star)
                    | Some(TokenType::Int)
                    | Some(TokenType::BigInt)
                    | Some(TokenType::Text)
                    | Some(TokenType::Date)
                    | Some(TokenType::Timestamp)
            );
            if !next_is_namelike {
                break;
            }
            self.advance(); // .
            if self.peek_type() == &TokenType::Star {
                name.push('.');
                name.push('*');
                self.advance();
                break;
            }
            let part = self.expect_name()?;
            name.push('.');
            name.push_str(&part);
        }
        Ok(name)
    }

    fn parse_column_def(&mut self) -> Result<ColumnDef> {
        let name = self.expect_name()?;
        let data_type = self.parse_data_type()?;

        let mut nullable = None;
        let mut default = None;
        let mut primary_key = false;
        let mut unique = false;
        let mut auto_increment = false;
        let mut collation = None;
        let mut comment = None;

        loop {
            if self.match_token(TokenType::Not) {
                self.expect(TokenType::Null)?;
                nullable = Some(false);
            } else if self.peek_type() == &TokenType::Null {
                self.advance();
                nullable = Some(true);
            } else if self.peek_type() == &TokenType::As
                && matches!(
                    self.peek_offset(1).map(|t| &t.token_type),
                    Some(TokenType::LParen)
                )
            {
                // SQLite / MySQL generated-column shorthand:
                //   `col TYPE AS (expr) [STORED|VIRTUAL|PERSISTENT]`.
                // Swallow AS, the parenthesised expression (depth-balanced),
                // and the optional storage-kind keyword.
                self.advance(); // AS
                self.advance(); // (
                let mut depth: i32 = 1;
                while depth > 0 {
                    match self.peek_type() {
                        TokenType::LParen => {
                            depth += 1;
                            self.advance();
                        }
                        TokenType::RParen => {
                            depth -= 1;
                            self.advance();
                        }
                        TokenType::Eof => break,
                        _ => {
                            self.advance();
                        }
                    }
                }
                if self.is_name_token()
                    && matches!(
                        self.peek().value.to_uppercase().as_str(),
                        "STORED" | "VIRTUAL" | "PERSISTENT" | "PERSISTED"
                    )
                {
                    self.advance();
                }
            } else if self.match_token(TokenType::Default) {
                // SQL Server / IBM `DEFAULT NEXT VALUE FOR seq[.qual]`.
                if self.is_name_token()
                    && self.peek().value.eq_ignore_ascii_case("NEXT")
                    && self
                        .peek_offset(1)
                        .map(|t| t.value.eq_ignore_ascii_case("VALUE"))
                        .unwrap_or(false)
                    && self
                        .peek_offset(2)
                        .map(|t| t.value.eq_ignore_ascii_case("FOR"))
                        .unwrap_or(false)
                {
                    self.advance();
                    self.advance();
                    self.advance();
                    let mut seq = self.expect_name()?;
                    while self.match_token(TokenType::Dot) {
                        seq.push('.');
                        seq.push_str(&self.expect_name()?);
                    }
                    default = Some(Expr::Function {
                        name: "NEXT_VALUE_FOR".to_string(),
                        args: vec![Expr::Column {
                            table: None,
                            name: seq,
                            quote_style: QuoteStyle::None,
                            table_quote_style: QuoteStyle::None,
                        }],
                        distinct: false,
                        filter: None,
                        over: None,
                        order_by: Vec::new(),
                        within_group: false,
                    });
                } else {
                    default = Some(self.parse_expr()?);
                }
            } else if self.match_token(TokenType::Primary) {
                self.expect(TokenType::Key)?;
                primary_key = true;
            } else if self.match_token(TokenType::Unique) {
                unique = true;
            } else if self.match_token(TokenType::AutoIncrement) {
                auto_increment = true;
            } else if self.match_token(TokenType::Collate) {
                collation = Some(self.expect_name()?);
            } else if self.match_token(TokenType::Comment) {
                let tok = self.expect(TokenType::String)?;
                comment = Some(tok.value);
            } else if self.match_token(TokenType::References) {
                // Inline foreign key — skip for now
                let _ = self.parse_table_ref()?;
                if self.match_token(TokenType::LParen) {
                    while !self.match_token(TokenType::RParen) {
                        self.advance();
                    }
                }
            } else if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("GENERATED") {
                // SQL:2003 / MySQL / PG / SQL Server identity / computed
                // column: `GENERATED ALWAYS AS (expr) [VIRTUAL|STORED]`,
                // `GENERATED ALWAYS AS IDENTITY [(...)]`,
                // `GENERATED BY DEFAULT AS IDENTITY [(...)]`. Swallow up
                // through the trailing parenthesised body if present and
                // let the next loop iteration pick up VIRTUAL/STORED.
                self.advance();
                if self.is_name_token()
                    && (self.peek().value.eq_ignore_ascii_case("ALWAYS")
                        || self.peek().value.eq_ignore_ascii_case("BY"))
                {
                    self.advance();
                    if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("DEFAULT") {
                        self.advance();
                    }
                }
                if self.match_token(TokenType::As) {
                    if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("IDENTITY") {
                        self.advance();
                    } else if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("ROW")
                    {
                        // SQL Server `GENERATED AS ROW START | END`.
                        self.advance();
                        if self.is_name_token() {
                            self.advance();
                        }
                    }
                }
                if self.peek_type() == &TokenType::LParen {
                    let mut depth = 0_i32;
                    self.advance();
                    depth += 1;
                    while depth > 0 {
                        match self.peek_type() {
                            TokenType::LParen => depth += 1,
                            TokenType::RParen => {
                                depth -= 1;
                                if depth == 0 {
                                    self.advance();
                                    break;
                                }
                            }
                            TokenType::Eof => break,
                            _ => {}
                        }
                        self.advance();
                    }
                }
            } else if self.is_name_token()
                && matches!(
                    self.peek().value.to_uppercase().as_str(),
                    "CODEC"
                        | "TTL"
                        | "MATERIALIZED"
                        | "ALIAS"
                        | "EPHEMERAL"
                        | "PERSISTED"
                        | "PERSISTENT"
                        | "VIRTUAL"
                        | "STORED"
                        | "ENCODE"
                        | "ENCRYPT"
                        | "MASKED"
                        | "INVISIBLE"
                        | "VISIBLE"
                        | "ENFORCED"
                        | "OPTIONS"
                        | "COMPRESSION"
                        | "SORTKEY"
                        | "DISTKEY"
                        | "CHARSET"
                        | "CHARACTER"
                        | "SRID"
                        | "FORMAT"
                        | "TAG"
                        | "MASKING"
                )
            {
                // ClickHouse / Snowflake / Redshift column modifiers. Consume
                // the keyword and the optional parenthesised body (`CODEC(...)`,
                // `TTL expr`, etc.) so the rest of the column def parses.
                self.advance();
                if self.peek_type() == &TokenType::LParen {
                    let mut depth = 0_i32;
                    self.advance();
                    depth += 1;
                    while depth > 0 {
                        match self.peek_type() {
                            TokenType::LParen => depth += 1,
                            TokenType::RParen => {
                                depth -= 1;
                                if depth == 0 {
                                    self.advance();
                                    break;
                                }
                            }
                            TokenType::Eof => break,
                            _ => {}
                        }
                        self.advance();
                    }
                } else {
                    // Best-effort: swallow an expression up to comma /
                    // top-level RParen / column-def boundary, balancing
                    // nested parens (e.g. `TTL toDate('2000-01-02')`,
                    // `ALIAS arrayResize(emptyArrayUInt32(), length(\`Arr.C2\`))`).
                    let mut depth: i32 = 0;
                    loop {
                        match self.peek_type() {
                            TokenType::LParen => {
                                depth += 1;
                                self.advance();
                            }
                            TokenType::RParen => {
                                if depth == 0 {
                                    break;
                                }
                                depth -= 1;
                                self.advance();
                            }
                            TokenType::Comma if depth == 0 => break,
                            TokenType::Eof => break,
                            _ => {
                                self.advance();
                            }
                        }
                    }
                }
            } else {
                break;
            }
        }

        Ok(ColumnDef {
            name,
            data_type,
            nullable,
            default,
            primary_key,
            unique,
            auto_increment,
            collation,
            comment,
        })
    }

    fn parse_data_type(&mut self) -> Result<DataType> {
        let token = self.peek().clone();
        // DuckDB / Spark template syntax: `${var}` (or `?` placeholder) used
        // where a data type is expected. Lower to `Unknown(name)` so the
        // surrounding expression parses.
        if matches!(token.token_type, TokenType::Parameter) {
            self.advance();
            return Ok(DataType::Unknown(token.value));
        }
        let type_result = match &token.token_type {
            TokenType::Int | TokenType::Integer => {
                self.advance();
                Ok(DataType::Int)
            }
            TokenType::BigInt => {
                self.advance();
                Ok(DataType::BigInt)
            }
            TokenType::SmallInt => {
                self.advance();
                Ok(DataType::SmallInt)
            }
            TokenType::TinyInt => {
                self.advance();
                Ok(DataType::TinyInt)
            }
            TokenType::Float => {
                self.advance();
                Ok(DataType::Float)
            }
            TokenType::Double => {
                self.advance();
                let _ = self.match_keyword("PRECISION");
                Ok(DataType::Double)
            }
            TokenType::Real => {
                self.advance();
                Ok(DataType::Real)
            }
            TokenType::Decimal | TokenType::Numeric => {
                let is_numeric = token.token_type == TokenType::Numeric;
                self.advance();
                let (precision, scale) = self.parse_type_params()?;
                if is_numeric {
                    Ok(DataType::Numeric { precision, scale })
                } else {
                    Ok(DataType::Decimal { precision, scale })
                }
            }
            TokenType::Varchar => {
                self.advance();
                let len = self.parse_single_type_param()?;
                Ok(DataType::Varchar(len))
            }
            TokenType::Char => {
                self.advance();
                let len = self.parse_single_type_param()?;
                Ok(DataType::Char(len))
            }
            TokenType::Text => {
                self.advance();
                Ok(DataType::Text)
            }
            TokenType::Boolean => {
                self.advance();
                Ok(DataType::Boolean)
            }
            TokenType::Date => {
                self.advance();
                Ok(DataType::Date)
            }
            TokenType::Timestamp => {
                self.advance();
                let precision = self.parse_single_type_param()?;
                let with_tz = if self.match_keyword("WITH") {
                    let _ = self.match_keyword("LOCAL");
                    let _ = self.match_keyword("TIME");
                    let _ = self.match_keyword("ZONE");
                    true
                } else if self.match_keyword("WITHOUT") {
                    let _ = self.match_keyword("TIME");
                    let _ = self.match_keyword("ZONE");
                    false
                } else {
                    false
                };
                Ok(DataType::Timestamp { precision, with_tz })
            }
            TokenType::TimestampTz => {
                self.advance();
                let precision = self.parse_single_type_param()?;
                Ok(DataType::Timestamp {
                    precision,
                    with_tz: true,
                })
            }
            TokenType::Time => {
                self.advance();
                let precision = self.parse_single_type_param()?;
                Ok(DataType::Time { precision })
            }
            TokenType::Interval => {
                self.advance();
                Ok(DataType::Interval)
            }
            TokenType::Blob => {
                self.advance();
                Ok(DataType::Blob)
            }
            TokenType::Bytea => {
                self.advance();
                Ok(DataType::Bytea)
            }
            TokenType::Json => {
                self.advance();
                Ok(DataType::Json)
            }
            TokenType::Jsonb => {
                self.advance();
                Ok(DataType::Jsonb)
            }
            TokenType::Uuid => {
                self.advance();
                Ok(DataType::Uuid)
            }
            TokenType::Array => {
                self.advance();
                if self.match_token(TokenType::Lt) {
                    let inner = self.parse_data_type()?;
                    self.expect(TokenType::Gt)?;
                    Ok(DataType::Array(Some(Box::new(inner))))
                } else {
                    Ok(DataType::Array(None))
                }
            }
            TokenType::Struct => {
                self.advance();
                // STRUCT<a INT, b STRING> (Hive/Spark) or STRUCT(a INT, b INT) (DuckDB).
                // Swallow the body — we don't model named struct fields in the AST.
                let close = if self.match_token(TokenType::Lt) {
                    Some(TokenType::Gt)
                } else if self.match_token(TokenType::LParen) {
                    Some(TokenType::RParen)
                } else {
                    None
                };
                if let Some(close_tok) = close {
                    let mut depth = 1_i32;
                    while depth > 0 {
                        if self.peek_type() == &TokenType::Eof {
                            break;
                        }
                        if self.peek_type() == &close_tok {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        } else if matches!(self.peek_type(), TokenType::Lt | TokenType::LParen)
                            && (self.peek_type() == &TokenType::Lt && close_tok == TokenType::Gt
                                || self.peek_type() == &TokenType::LParen
                                    && close_tok == TokenType::RParen)
                        {
                            depth += 1;
                        }
                        self.advance();
                    }
                }
                Ok(DataType::Unknown("STRUCT".to_string()))
            }
            TokenType::Map => {
                self.advance();
                let close = if self.match_token(TokenType::Lt) {
                    Some(TokenType::Gt)
                } else if self.match_token(TokenType::LParen) {
                    Some(TokenType::RParen)
                } else {
                    None
                };
                if let Some(close_tok) = close {
                    let mut depth = 1_i32;
                    while depth > 0 {
                        if self.peek_type() == &TokenType::Eof {
                            break;
                        }
                        if self.peek_type() == &close_tok {
                            depth -= 1;
                            if depth == 0 {
                                self.advance();
                                break;
                            }
                        } else if (self.peek_type() == &TokenType::Lt && close_tok == TokenType::Gt)
                            || (self.peek_type() == &TokenType::LParen
                                && close_tok == TokenType::RParen)
                        {
                            depth += 1;
                        }
                        self.advance();
                    }
                }
                Ok(DataType::Unknown("MAP".to_string()))
            }
            TokenType::Identifier => {
                let name = token.value.to_uppercase();
                self.advance();
                match name.as_str() {
                    "STRING" => Ok(DataType::String),
                    "BINARY" => {
                        let len = self.parse_single_type_param()?;
                        Ok(DataType::Binary(len))
                    }
                    "VARBINARY" => {
                        let len = self.parse_single_type_param()?;
                        Ok(DataType::Varbinary(len))
                    }
                    "DATETIME" => Ok(DataType::DateTime),
                    "BYTES" => Ok(DataType::Bytes),
                    "VARIANT" => Ok(DataType::Variant),
                    "OBJECT" => Ok(DataType::Object),
                    "XML" => Ok(DataType::Xml),
                    "INET" => Ok(DataType::Inet),
                    "CIDR" => Ok(DataType::Cidr),
                    "MACADDR" => Ok(DataType::Macaddr),
                    "BIT" => {
                        // Postgres `BIT VARYING(n)` is the same as VARBIT.
                        // Swallow the VARYING keyword if present and parse
                        // the length normally.
                        if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("VARYING")
                        {
                            self.advance();
                            let len = self.parse_single_type_param()?;
                            return Ok(DataType::Varbinary(len));
                        }
                        let len = self.parse_single_type_param()?;
                        Ok(DataType::Bit(len))
                    }
                    "MONEY" => Ok(DataType::Money),
                    "SERIAL" => Ok(DataType::Serial),
                    "BIGSERIAL" => Ok(DataType::BigSerial),
                    "SMALLSERIAL" => Ok(DataType::SmallSerial),
                    "REGCLASS" => Ok(DataType::Regclass),
                    "REGTYPE" => Ok(DataType::Regtype),
                    "HSTORE" => Ok(DataType::Hstore),
                    "GEOGRAPHY" => Ok(DataType::Geography),
                    "GEOMETRY" => Ok(DataType::Geometry),
                    "SUPER" => Ok(DataType::Super),
                    _ => Ok(DataType::Unknown(name)),
                }
            }
            _ => {
                // Fallback: accept any keyword-like token as an unknown
                // data type by its textual value. Covers PostgreSQL `cube`,
                // `lseg`, `path`, `polygon`, and any vendor-specific type
                // name that happens to collide with a TokenType variant.
                let v = token.value.clone();
                if !v.is_empty() && v.chars().all(|c| c.is_ascii_alphanumeric() || c == '_') {
                    self.advance();
                    Ok(DataType::Unknown(v.to_uppercase()))
                } else {
                    Err(SqlglotError::ParserError {
                        message: format!("Expected data type, got {:?}", token.token_type),
                    })
                }
            }
        };

        // PostgreSQL opt_array_bounds: typename[], typename[N], typename[][]...
        let mut dt = type_result?;
        while self.match_token(TokenType::LBracket) {
            // Consume optional integer bound (PostgreSQL ignores it but accepts it)
            let _ = self.match_token(TokenType::Number);
            self.expect(TokenType::RBracket)?;
            dt = DataType::Array(Some(Box::new(dt)));
        }
        // ClickHouse parameterized types: `DateTime('Asia/Dubai')`,
        // `Nullable(String)`, `Array(Int32)`, `Enum8('a' = 1, 'b' = 2)`,
        // `Decimal(9, 2)`, etc. The base type was already produced — swallow
        // the parenthesized parameter list so the surrounding expression
        // continues to parse.
        if self.peek_type() == &TokenType::LParen {
            let saved = self.pos;
            self.advance();
            let mut depth = 1;
            let mut ok = true;
            while depth > 0 {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    TokenType::Eof => {
                        ok = false;
                        break;
                    }
                    _ => {}
                }
                self.advance();
            }
            if !ok {
                self.pos = saved;
            }
        }
        Ok(dt)
    }

    fn parse_type_params(&mut self) -> Result<(Option<u32>, Option<u32>)> {
        if self.match_token(TokenType::LParen) {
            let p: Option<u32> = self.expect(TokenType::Number)?.value.parse().ok();
            let s = if self.match_token(TokenType::Comma) {
                self.expect(TokenType::Number)?.value.parse().ok()
            } else {
                None
            };
            self.expect(TokenType::RParen)?;
            Ok((p, s))
        } else {
            Ok((None, None))
        }
    }

    fn parse_single_type_param(&mut self) -> Result<Option<u32>> {
        if self.match_token(TokenType::LParen) {
            // Handle TSQL MAX keyword (e.g. VARBINARY(MAX), VARCHAR(MAX))
            if self.check_keyword("MAX") {
                self.advance(); // consume MAX
                self.expect(TokenType::RParen)?;
                return Ok(None);
            }
            let n: Option<u32> = self.expect(TokenType::Number)?.value.parse().ok();
            self.expect(TokenType::RParen)?;
            Ok(n)
        } else {
            Ok(None)
        }
    }

    // ── DROP ────────────────────────────────────────────────────────

    fn parse_drop(&mut self) -> Result<Statement> {
        self.expect(TokenType::Drop)?;

        if self.match_token(TokenType::Materialized) {
            self.expect(TokenType::View)?;
            let if_exists = if self.match_token(TokenType::If) {
                self.expect(TokenType::Exists)?;
                true
            } else {
                false
            };
            let name = self.parse_table_ref()?;
            // MySQL/MariaDB allow comma-list — swallow the rest.
            while self.match_token(TokenType::Comma) {
                let _ = self.parse_table_ref()?;
            }
            // Trailing CASCADE / RESTRICT.
            let _ = self.match_token(TokenType::Cascade) || self.match_token(TokenType::Restrict);
            return Ok(Statement::DropView(DropViewStatement {
                comments: vec![],
                name,
                if_exists,
                materialized: true,
            }));
        }

        if self.match_token(TokenType::View) {
            let if_exists = if self.match_token(TokenType::If) {
                self.expect(TokenType::Exists)?;
                true
            } else {
                false
            };
            let name = self.parse_table_ref()?;
            while self.match_token(TokenType::Comma) {
                let _ = self.parse_table_ref()?;
            }
            let _ = self.match_token(TokenType::Cascade) || self.match_token(TokenType::Restrict);
            return Ok(Statement::DropView(DropViewStatement {
                comments: vec![],
                name,
                if_exists,
                materialized: false,
            }));
        }

        // DROP <kind> ... — preserve as a Command for non-TABLE/VIEW drops
        // (FUNCTION, PROCEDURE, SCHEMA, DATABASE, INDEX, ROLE, USER, …).
        if self.peek_type() != &TokenType::Table {
            // Already consumed DROP; capture the remainder.
            let body = self.consume_raw_to_statement_end();
            return Ok(Statement::Command(CommandStatement {
                comments: vec![],
                kind: "DROP".to_string(),
                body,
            }));
        }

        self.expect(TokenType::Table)?;

        let if_exists = if self.match_token(TokenType::If) {
            self.expect(TokenType::Exists)?;
            true
        } else {
            false
        };

        let table = self.parse_table_ref()?;
        // MySQL / MariaDB: `DROP TABLE [IF EXISTS] t1, t2, …`. Swallow the
        // extra table names so the statement parses.
        while self.match_token(TokenType::Comma) {
            let _ = self.parse_table_ref()?;
        }
        let cascade = self.match_token(TokenType::Cascade);
        // Tolerate Doris / StarRocks / Oracle trailing modifiers on DROP TABLE
        // (`FORCE`, `PURGE`, `RESTRICT`).
        while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon) {
            if self.is_name_token()
                && matches!(
                    self.peek().value.to_uppercase().as_str(),
                    "FORCE" | "PURGE" | "RESTRICT"
                )
            {
                self.advance();
            } else if matches!(self.peek_type(), TokenType::Restrict) {
                self.advance();
            } else {
                break;
            }
        }

        Ok(Statement::DropTable(DropTableStatement {
            comments: vec![],
            if_exists,
            table,
            cascade,
        }))
    }

    // ── ALTER TABLE ─────────────────────────────────────────────────

    fn parse_alter_table(&mut self) -> Result<AlterTableStatement> {
        self.expect(TokenType::Alter)?;
        self.expect(TokenType::Table)?;
        let table = self.parse_table_ref_no_alias()?;

        let mut actions = Vec::new();
        loop {
            let action = self.parse_alter_action()?;
            actions.push(action);
            if !self.match_token(TokenType::Comma) {
                break;
            }
        }

        Ok(AlterTableStatement {
            comments: vec![],
            table,
            actions,
        })
    }

    fn parse_alter_action(&mut self) -> Result<AlterTableAction> {
        // Hive multi-partition continuation after a comma:
        // `ALTER TABLE t DROP PARTITION (a), PARTITION (b)`. Swallow the
        // bare PARTITION clause.
        if self.peek_type() == &TokenType::Partition {
            self.advance();
            let mut depth: i32 = 0;
            while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                && (depth > 0 || !matches!(self.peek_type(), TokenType::Comma))
            {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => depth = depth.saturating_sub(1),
                    _ => {}
                }
                self.advance();
            }
            return Ok(AlterTableAction::DropColumn {
                name: String::new(),
                if_exists: false,
            });
        }
        if self.match_keyword("ADD") {
            if matches!(
                self.peek_type(),
                TokenType::Constraint
                    | TokenType::Primary
                    | TokenType::Unique
                    | TokenType::Foreign
                    | TokenType::Check
            ) {
                let constraint = self.parse_table_constraint()?;
                self.swallow_constraint_modifiers();
                Ok(AlterTableAction::AddConstraint(constraint))
            } else if self.check_keyword("EXCLUDE") {
                // PG `ADD EXCLUDE [USING method] (col WITH op [, ...]) [WHERE
                // (predicate)] [DEFERRABLE …]` — swallow opaquely until we
                // hit a top-level statement boundary or comma.
                let mut depth: i32 = 0;
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    && (depth > 0 || !matches!(self.peek_type(), TokenType::Comma))
                {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth = depth.saturating_sub(1),
                        _ => {}
                    }
                    self.advance();
                }
                Ok(AlterTableAction::DropColumn {
                    name: String::new(),
                    if_exists: false,
                })
            } else if self.check_keyword("INDEX")
                || self.check_keyword("KEY")
                || self.check_keyword("PROJECTION")
                || self.check_keyword("STATISTICS")
            {
                // ClickHouse / MySQL `ADD INDEX [name] expr TYPE x GRANULARITY n
                // [AFTER y]`, `ADD KEY ...`, `ADD PROJECTION ...`. The body
                // is heterogeneous; swallow it opaquely up to the next
                // top-level Comma / Semicolon / EOF.
                let mut depth: i32 = 0;
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    && (depth > 0 || !matches!(self.peek_type(), TokenType::Comma))
                {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth = depth.saturating_sub(1),
                        _ => {}
                    }
                    self.advance();
                }
                Ok(AlterTableAction::DropColumn {
                    name: String::new(),
                    if_exists: false,
                })
            } else if self.check_keyword("COLUMNS") {
                // Hive / Spark / Databricks `ALTER TABLE … ADD COLUMNS
                // (col type [, col type]*)` or the comma-list form
                // `ADD COLUMNS col type, col type`. Swallow opaquely.
                self.advance();
                let mut depth: i32 = 0;
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    && (depth > 0 || !matches!(self.peek_type(), TokenType::Comma))
                {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth = depth.saturating_sub(1),
                        _ => {}
                    }
                    self.advance();
                    if depth == 0
                        && matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    {
                        break;
                    }
                }
                Ok(AlterTableAction::DropColumn {
                    name: String::new(),
                    if_exists: false,
                })
            } else {
                let _ = self.match_keyword("COLUMN");
                let col = self.parse_column_def()?;
                // ClickHouse: `ADD COLUMN name type AFTER other` / `FIRST` —
                // consume the placement modifier so the rest of the action
                // list parses.
                if self.check_keyword("AFTER") {
                    self.advance();
                    if self.is_name_token() {
                        self.advance();
                    }
                } else if self.check_keyword("FIRST") {
                    self.advance();
                }
                Ok(AlterTableAction::AddColumn(col))
            }
        } else if self.match_token(TokenType::Drop) {
            // Hive: `DROP IF EXISTS PARTITION (…), PARTITION (…)`. The
            // optional `IF EXISTS` precedes PARTITION.
            if self.peek_type() == &TokenType::If
                && self
                    .peek_offset(1)
                    .map(|t| matches!(t.token_type, TokenType::Exists))
                    .unwrap_or(false)
                && self
                    .peek_offset(2)
                    .map(|t| matches!(t.token_type, TokenType::Partition))
                    .unwrap_or(false)
            {
                self.advance(); // IF
                self.advance(); // EXISTS
            }
            // MySQL / TiDB: `DROP INDEX|KEY name`, `DROP PRIMARY KEY`,
            // `DROP FOREIGN KEY name`, `DROP CONSTRAINT name`,
            // `DROP PARTITION (...)`, `DROP CHECK name`. We don't have a
            // dedicated AST node for these, so swallow them to end-of-action.
            if matches!(
                self.peek_type(),
                TokenType::Index
                    | TokenType::Primary
                    | TokenType::Foreign
                    | TokenType::Constraint
                    | TokenType::Check
                    | TokenType::Partition
                    | TokenType::Unique
            ) || self.check_keyword("KEY")
                || self.check_keyword("FEATURE")
                || self.check_keyword("PROJECTION")
                || self.check_keyword("STATISTICS")
                || self.check_keyword("INDEX")
                || self.check_keyword("DISTRIBUTION")
            {
                let mut depth: i32 = 0;
                while !matches!(self.peek_type(), TokenType::Eof | TokenType::Semicolon)
                    && (depth > 0 || !matches!(self.peek_type(), TokenType::Comma))
                {
                    match self.peek_type() {
                        TokenType::LParen => depth += 1,
                        TokenType::RParen => depth = depth.saturating_sub(1),
                        _ => {}
                    }
                    self.advance();
                }
                return Ok(AlterTableAction::DropColumn {
                    name: String::new(),
                    if_exists: false,
                });
            }
            let _ = self.match_keyword("COLUMN");
            let if_exists = if self.match_token(TokenType::If) {
                self.expect(TokenType::Exists)?;
                true
            } else {
                false
            };
            let mut name = self.expect_name()?;
            // ClickHouse `DROP COLUMN nested.col` — accept dotted suffixes;
            // we collapse them into the column name string for now.
            while self.peek_type() == &TokenType::Dot {
                self.advance();
                if !self.is_name_token() {
                    break;
                }
                name.push('.');
                name.push_str(&self.peek().value);
                self.advance();
            }
            Ok(AlterTableAction::DropColumn { name, if_exists })
        } else if self.match_keyword("RENAME") {
            if self.match_keyword("COLUMN") {
                let old_name = self.expect_name()?;
                self.expect(TokenType::Identifier)?; // TO
                let new_name = self.expect_name()?;
                Ok(AlterTableAction::RenameColumn { old_name, new_name })
            } else if self.match_keyword("TO") {
                let mut new_name = self.expect_name()?;
                while self.match_token(TokenType::Dot) {
                    new_name.push('.');
                    new_name.push_str(&self.expect_name()?);
                }
                Ok(AlterTableAction::RenameTable { new_name })
            } else {
                Err(SqlglotError::ParserError {
                    message: "Expected COLUMN or TO after RENAME".into(),
                })
            }
        } else {
            Err(SqlglotError::ParserError {
                message: "Expected ADD, DROP, or RENAME in ALTER TABLE".into(),
            })
        }
    }

    /// Try [`parse_alter_table`]; on failure, rewind and capture the entire
    /// `ALTER …` statement verbatim as a [`Statement::Command`]. This covers
    /// the long tail of vendor-specific ALTER forms — MySQL `ALTER TABLE …
    /// CONVERT TO CHARACTER SET … COLLATE …`, Hive `ALTER TABLE … PARTITION
    /// (…) COMPACT 'major'`, T-SQL `ALTER TABLE … WITH (…) CHECK CONSTRAINT
    /// …`, etc. (Gap 5)
    fn parse_alter_or_command(&mut self) -> Result<Statement> {
        let saved = self.pos;
        let saved_comments = self.pending_comments.clone();
        match self.parse_alter_table() {
            Ok(stmt) => Ok(Statement::AlterTable(stmt)),
            Err(_) => {
                self.pos = saved;
                self.pending_comments = saved_comments;
                self.parse_command_kind("ALTER")
            }
        }
    }

    /// Try [`parse_create`]; on failure, rewind and capture the entire
    /// `CREATE …` statement verbatim as a [`Statement::Command`]. Also
    /// handles the `CREATE TABLE t AS VALUES (…)` form (Gap 7) and rarer
    /// `CREATE OPERATOR / AGGREGATE / SEQUENCE / FUNCTION / TEXT SEARCH
    /// CONFIGURATION / …` (Gap 4).
    fn parse_create_or_command(&mut self) -> Result<Statement> {
        let saved = self.pos;
        let saved_comments = self.pending_comments.clone();
        match self.parse_create() {
            Ok(stmt) => Ok(stmt),
            Err(_) => {
                self.pos = saved;
                self.pending_comments = saved_comments;
                self.parse_command_kind("CREATE")
            }
        }
    }

    // ── TRUNCATE ────────────────────────────────────────────────────

    fn parse_truncate(&mut self) -> Result<TruncateStatement> {
        self.expect(TokenType::Truncate)?;
        let _ = self.match_token(TokenType::Table);
        let table = self.parse_table_ref()?;
        Ok(TruncateStatement {
            comments: vec![],
            table,
        })
    }

    // ── Transaction ─────────────────────────────────────────────────

    fn parse_transaction(&mut self) -> Result<TransactionStatement> {
        match self.peek_type() {
            TokenType::Begin => {
                self.advance();
                let _ = self.match_token(TokenType::Transaction);
                let _ = self.match_keyword("WORK");
                Ok(TransactionStatement::Begin)
            }
            TokenType::Commit => {
                self.advance();
                let _ = self.match_token(TokenType::Transaction);
                let _ = self.match_keyword("WORK");
                // SQL-standard COMMIT [WORK] [AND [NO] CHAIN]
                if self.match_token(TokenType::And) {
                    let _ = self.match_token(TokenType::Not);
                    let _ = self.match_keyword("NO");
                    let _ = self.match_keyword("CHAIN");
                }
                Ok(TransactionStatement::Commit)
            }
            TokenType::Rollback => {
                self.advance();
                let _ = self.match_token(TokenType::Transaction);
                let _ = self.match_keyword("WORK");
                if self.match_keyword("TO") {
                    let _ = self.match_token(TokenType::Savepoint);
                    let name = self.expect_name()?;
                    Ok(TransactionStatement::RollbackTo(name))
                } else {
                    // ROLLBACK [WORK] [AND [NO] CHAIN]
                    if self.match_token(TokenType::And) {
                        let _ = self.match_token(TokenType::Not);
                        let _ = self.match_keyword("NO");
                        let _ = self.match_keyword("CHAIN");
                    }
                    Ok(TransactionStatement::Rollback)
                }
            }
            TokenType::Savepoint => {
                self.advance();
                let name = self.expect_name()?;
                Ok(TransactionStatement::Savepoint(name))
            }
            _ => Err(SqlglotError::ParserError {
                message: "Expected transaction statement".into(),
            }),
        }
    }

    // ── EXPLAIN ─────────────────────────────────────────────────────

    fn parse_explain(&mut self) -> Result<ExplainStatement> {
        self.expect(TokenType::Explain)?;
        let analyze = self.match_token(TokenType::Analyze);
        // PostgreSQL `EXPLAIN (VERBOSE, COSTS OFF, ...)` option block, plus
        // unparenthesized `VERBOSE` / `FORMAT TEXT|JSON|YAML`.
        if self.match_token(TokenType::LParen) {
            let mut depth = 1;
            while depth > 0 {
                match self.peek_type() {
                    TokenType::Eof => break,
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        depth -= 1;
                        if depth == 0 {
                            self.advance();
                            break;
                        }
                    }
                    _ => {}
                }
                self.advance();
            }
        } else {
            // Optional bare keywords: VERBOSE / FORMAT [=] <name|string>
            loop {
                if self.check_keyword("VERBOSE") {
                    self.advance();
                    continue;
                }
                if self.check_keyword("FORMAT") {
                    self.advance();
                    let _ = self.match_token(TokenType::Eq);
                    // Format name can be an identifier (TEXT/JSON/YAML/XML/...)
                    // or a string literal (`'plan_tree'`).
                    if matches!(self.peek_type(), TokenType::String | TokenType::Identifier)
                        || self.is_name_token()
                    {
                        self.advance();
                    }
                    continue;
                }
                break;
            }
            // Hive / Spark EXPLAIN modifiers: EXTENDED, LOCKS, AUTHORIZATION,
            // DEPENDENCY, VECTORIZATION [ONLY] [SUMMARY|OPERATOR|EXPRESSION|DETAIL],
            // CBO, AST, REWRITE, FORMATTED, LOGICAL, NODE. Also ClickHouse
            // `EXPLAIN indexes=1 actions=1 …` bare options. Consume any
            // identifier-like tokens (and optional `= value`) until we hit a
            // statement-starting keyword.
            loop {
                match self.peek_type() {
                    TokenType::Select
                    | TokenType::With
                    | TokenType::Insert
                    | TokenType::Update
                    | TokenType::Delete
                    | TokenType::Merge
                    | TokenType::Create
                    | TokenType::Drop
                    | TokenType::Alter
                    | TokenType::Truncate
                    | TokenType::LParen
                    | TokenType::Eof
                    | TokenType::Semicolon => break,
                    TokenType::Identifier => {
                        self.advance();
                        if self.match_token(TokenType::Eq) {
                            // value: number, string, or identifier
                            if matches!(self.peek_type(), TokenType::Number | TokenType::String)
                                || self.is_name_token()
                            {
                                self.advance();
                            }
                        }
                        // Optional comma between options
                        // (ClickHouse `dump_tree = 1, dump_ast = 1 …`).
                        let _ = self.match_token(TokenType::Comma);
                    }
                    _ => {
                        // Also accept unreserved keyword-style modifiers
                        // (ONLY, FORMATTED, EXTENDED, etc. that tokenize as
                        // their own variants). Bail when we hit anything
                        // that isn't a plain name token.
                        if self.is_name_token() {
                            self.advance();
                        } else {
                            break;
                        }
                    }
                }
            }
        }
        let statement = self.parse_statement_inner()?;
        Ok(ExplainStatement {
            comments: vec![],
            analyze,
            statement: Box::new(statement),
        })
    }

    // ── USE ─────────────────────────────────────────────────────────

    fn parse_use(&mut self) -> Result<UseStatement> {
        self.expect(TokenType::Use)?;
        // Optional kind: USE DATABASE / SCHEMA / CATALOG / WAREHOUSE / ROLE
        // (DuckDB / Snowflake / Spark). Swallow the leading keyword.
        let _ = matches!(self.peek_type(), TokenType::Database | TokenType::Schema) && {
            self.advance();
            true
        } || (self.is_name_token()
            && matches!(
                self.peek().value.to_uppercase().as_str(),
                "CATALOG" | "WAREHOUSE" | "ROLE"
            )
            && {
                self.advance();
                true
            });
        // `USE default` (Hive): `default` is a keyword, accept it as a name.
        let mut name = if matches!(self.peek_type(), TokenType::Default) {
            let v = self.peek().value.clone();
            self.advance();
            v
        } else if self.is_name_token()
            && self.peek().value.eq_ignore_ascii_case("IDENTIFIER")
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::LParen)
            )
        {
            // Snowflake / Databricks IDENTIFIER('name') indirection —
            // swallow the call and use a synthetic name.
            self.advance(); // IDENTIFIER
            self.advance(); // (
            let mut depth: i32 = 1;
            while depth > 0 {
                match self.peek_type() {
                    TokenType::LParen => {
                        depth += 1;
                        self.advance();
                    }
                    TokenType::RParen => {
                        depth -= 1;
                        self.advance();
                    }
                    TokenType::Eof => break,
                    _ => {
                        self.advance();
                    }
                }
            }
            "IDENTIFIER".to_string()
        } else {
            self.expect_name()?
        };
        while self.match_token(TokenType::Dot) {
            name.push('.');
            if matches!(self.peek_type(), TokenType::Default) {
                name.push_str(&self.peek().value);
                self.advance();
            } else {
                name.push_str(&self.expect_name()?);
            }
        }
        Ok(UseStatement {
            comments: vec![],
            name,
        })
    }

    // ══════════════════════════════════════════════════════════════
    // Expression parsing (precedence climbing)
    // ══════════════════════════════════════════════════════════════

    fn parse_expr(&mut self) -> Result<Expr> {
        // DuckDB lambda: `lambda x: body` or `lambda x, y: body`. Lower to a
        // `Function("lambda", [name(s), body])` placeholder so the call parses.
        if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("lambda") {
            let saved = self.pos;
            self.advance();
            let mut names: Vec<Expr> = Vec::new();
            let mut ok = self.is_name_token();
            while ok {
                let n = self.advance().clone();
                names.push(Expr::Column {
                    table: None,
                    name: n.value.clone(),
                    table_quote_style: QuoteStyle::None,
                    quote_style: QuoteStyle::None,
                });
                if !self.match_token(TokenType::Comma) {
                    break;
                }
                if !self.is_name_token() {
                    ok = false;
                    break;
                }
            }
            if ok && self.match_token(TokenType::Colon) {
                let body = self.parse_expr()?;
                let mut args = names;
                args.push(body);
                return Ok(Expr::Function {
                    name: "lambda".to_string(),
                    args,
                    distinct: false,
                    filter: None,
                    over: None,
                    order_by: Vec::new(),
                    within_group: false,
                });
            }
            self.pos = saved;
        }
        // DuckDB / PostgreSQL named-argument prefix `name := value` and
        // BigQuery `name => value` — discard the name so the surrounding
        // function call parses. Only triggered when the lookahead clearly
        // matches the named-arg shape.
        if self.is_name_token() {
            let next = self.peek_offset(1).map(|t| &t.token_type);
            let after = self.peek_offset(2).map(|t| &t.token_type);
            if matches!(next, Some(TokenType::Colon)) && matches!(after, Some(TokenType::Eq)) {
                self.advance();
                self.advance();
                self.advance();
            } else if matches!(next, Some(TokenType::DoubleArrow)) {
                self.advance();
                self.advance();
            } else if matches!(next, Some(TokenType::Eq)) && matches!(after, Some(TokenType::Gt)) {
                // `name => value` tokenized as `Eq Gt` (no DoubleArrow merge).
                self.advance();
                self.advance();
                self.advance();
            }
        }
        let cond = self.parse_or_expr()?;
        // MySQL session-variable assignment in expression position:
        // `@var := expr`. Tokenized as `Colon Eq`. Lower to `BinaryOp Eq`
        // so the surrounding query parses.
        if matches!(self.peek_type(), TokenType::Colon)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::Eq)
            )
        {
            self.advance();
            self.advance();
            let rhs = self.parse_expr()?;
            return Ok(Expr::BinaryOp {
                left: Box::new(cond),
                op: BinaryOperator::Eq,
                right: Box::new(rhs),
            });
        }
        // ClickHouse C-style ternary: `cond ? then : else`. Tokenized as
        // `Parameter('?')` followed later by `Colon`. Lower to a CASE.
        if matches!(self.peek_type(), TokenType::Parameter) && self.peek().value == "?" {
            self.advance();
            let then_branch = self.parse_or_expr()?;
            if self.match_token(TokenType::Colon) {
                let else_branch = self.parse_expr()?;
                return Ok(Expr::Case {
                    operand: None,
                    when_clauses: vec![(cond, then_branch)],
                    else_clause: Some(Box::new(else_branch)),
                });
            }
        }
        Ok(cond)
    }

    fn parse_or_expr(&mut self) -> Result<Expr> {
        let mut left = self.parse_and_expr()?;
        while self.match_token(TokenType::Or) {
            let right = self.parse_and_expr()?;
            left = Expr::BinaryOp {
                left: Box::new(left),
                op: BinaryOperator::Or,
                right: Box::new(right),
            };
        }
        Ok(left)
    }

    fn parse_and_expr(&mut self) -> Result<Expr> {
        let mut left = self.parse_not_expr()?;
        while self.match_token(TokenType::And) {
            let right = self.parse_not_expr()?;
            left = Expr::BinaryOp {
                left: Box::new(left),
                op: BinaryOperator::And,
                right: Box::new(right),
            };
        }
        Ok(left)
    }

    fn parse_not_expr(&mut self) -> Result<Expr> {
        if self.match_token(TokenType::Not) {
            let expr = self.parse_not_expr()?;
            Ok(Expr::UnaryOp {
                op: UnaryOperator::Not,
                expr: Box::new(expr),
            })
        } else {
            self.parse_comparison()
        }
    }

    fn parse_comparison(&mut self) -> Result<Expr> {
        let mut left = self.parse_addition()?;

        loop {
            // ClickHouse distributed predicates: `expr GLOBAL [NOT] IN (...)`
            // and `expr GLOBAL JOIN ...`. The keyword tokenizes as a plain
            // identifier — swallow it so the following predicate parses.
            if self.check_keyword("GLOBAL") {
                let next = self.peek_offset(1).map(|t| &t.token_type);
                if matches!(next, Some(TokenType::In) | Some(TokenType::Not)) {
                    self.advance();
                }
            }
            // ANSI / Postgres `period1 OVERLAPS period2` — model as Eq for
            // acceptance purposes.
            if self.check_keyword("OVERLAPS") {
                self.advance();
                let right = self.parse_addition()?;
                left = Expr::BinaryOp {
                    left: Box::new(left),
                    op: BinaryOperator::Eq,
                    right: Box::new(right),
                };
                continue;
            }
            // MySQL JSON `value MEMBER OF (json_array_expr)` — model as Eq.
            if self.check_keyword("MEMBER")
                && self
                    .peek_offset(1)
                    .map(|t| t.value.eq_ignore_ascii_case("OF"))
                    .unwrap_or(false)
            {
                self.advance();
                self.advance();
                let right = self.parse_addition()?;
                left = Expr::BinaryOp {
                    left: Box::new(left),
                    op: BinaryOperator::Eq,
                    right: Box::new(right),
                };
                continue;
            }
            // PostgreSQL geometric and full-text operators that tokenize as
            // multi-character sequences our tokenizer doesn't fuse:
            //   `<->`  (distance)         tokens: Lt, Arrow
            //   `&&` `&<` `&>`            (array / range overlap)
            //   `@@`                      (text search match)
            //   `|>` `<|`                 (range left/right of)
            // Lower all of them to a generic Eq so the surrounding
            // expression parses; the bench only cares about acceptance.
            {
                let p0 = self.peek_type().clone();
                let p1 = self.peek_offset(1).map(|t| t.token_type.clone());
                let p2 = self.peek_offset(2).map(|t| t.token_type.clone());
                let p1v = self
                    .peek_offset(1)
                    .map(|t| t.value.clone())
                    .unwrap_or_default();
                let consume_count = match (&p0, &p1, &p2) {
                    // <-> distance
                    (TokenType::Lt, Some(TokenType::Arrow), _) => 2,
                    // && overlap
                    (TokenType::BitwiseAnd, Some(TokenType::BitwiseAnd), _) => 2,
                    // &<| / &>| geometric variants
                    (TokenType::BitwiseAnd, Some(TokenType::Lt), Some(TokenType::BitwiseOr))
                    | (TokenType::BitwiseAnd, Some(TokenType::Gt), Some(TokenType::BitwiseOr)) => 3,
                    // &< / &>
                    (TokenType::BitwiseAnd, Some(TokenType::Lt), _)
                    | (TokenType::BitwiseAnd, Some(TokenType::Gt), _) => 2,
                    // @@ and @?
                    (TokenType::AtSign, Some(TokenType::AtSign), _) => 2,
                    // |> and <|
                    (TokenType::BitwiseOr, Some(TokenType::Gt), _)
                    | (TokenType::Lt, Some(TokenType::BitwiseOr), _) => 2,
                    // <<| / >>|
                    (TokenType::ShiftLeft, Some(TokenType::BitwiseOr), _)
                    | (TokenType::ShiftRight, Some(TokenType::BitwiseOr), _) => 2,
                    // ^@ starts_with operator
                    (TokenType::BitwiseXor, Some(TokenType::AtSign), _) => 2,
                    _ if matches!(p0, TokenType::AtSign)
                        && matches!(p1, Some(TokenType::Parameter))
                        && p1v == "?" =>
                    {
                        2
                    }
                    _ => 0,
                };
                if consume_count > 0 {
                    for _ in 0..consume_count {
                        self.advance();
                    }
                    let right = self.parse_addition()?;
                    left = Expr::BinaryOp {
                        left: Box::new(left),
                        op: BinaryOperator::Eq,
                        right: Box::new(right),
                    };
                    continue;
                }
            }
            let op = match self.peek_type() {
                TokenType::Eq => Some(BinaryOperator::Eq),
                TokenType::Neq => Some(BinaryOperator::Neq),
                TokenType::Lt => Some(BinaryOperator::Lt),
                TokenType::Gt => Some(BinaryOperator::Gt),
                TokenType::LtEq => {
                    // Hive / MySQL `<=>` null-safe equality tokenizes as `Lte Gt`.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::Gt)
                    ) {
                        self.advance();
                        self.advance();
                        let right = self.parse_addition()?;
                        left = Expr::BinaryOp {
                            left: Box::new(left),
                            op: BinaryOperator::Eq,
                            right: Box::new(right),
                        };
                        continue;
                    }
                    Some(BinaryOperator::LtEq)
                }
                TokenType::GtEq => Some(BinaryOperator::GtEq),
                TokenType::AtArrow => Some(BinaryOperator::AtArrow),
                TokenType::ArrowAt => Some(BinaryOperator::ArrowAt),
                // PostgreSQL geometric / regex operators starting with `~`:
                //   ~=, ~<, ~>, ~<=, ~>=, ~~, ~~*, !~, !~*. We lower all of
                //   them to a generic Eq comparison so the surrounding
                //   expression parses; the bench only cares about acceptance.
                TokenType::BitwiseNot => {
                    self.advance();
                    // Optional follow-up: =, <, >, <=, >=, ~, ~*, *.
                    let _ = match self.peek_type() {
                        TokenType::Eq
                        | TokenType::Lt
                        | TokenType::Gt
                        | TokenType::LtEq
                        | TokenType::GtEq
                        | TokenType::Star
                        | TokenType::BitwiseNot => {
                            self.advance();
                            // Allow `~~*` (LIKE-like, case-insensitive).
                            if self.peek_type() == &TokenType::Star {
                                self.advance();
                            }
                            true
                        }
                        _ => false,
                    };
                    let right = self.parse_addition()?;
                    left = Expr::BinaryOp {
                        left: Box::new(left),
                        op: BinaryOperator::Eq,
                        right: Box::new(right),
                    };
                    continue;
                }
                _ => None,
            };

            if let Some(op) = op {
                self.advance();
                // ClickHouse / SQLite accept `==` as a synonym for `=`.
                if matches!(op, BinaryOperator::Eq) && self.peek_type() == &TokenType::Eq {
                    self.advance();
                }
                if matches!(self.peek_type(), TokenType::Any | TokenType::Some) {
                    self.advance();
                    self.expect(TokenType::LParen)?;
                    let right = if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                        Expr::Subquery(Box::new(self.parse_statement_inner()?))
                    } else {
                        self.parse_expr()?
                    };
                    self.expect(TokenType::RParen)?;
                    left = Expr::AnyOp {
                        expr: Box::new(left),
                        op,
                        right: Box::new(right),
                    };
                } else if self.peek_type() == &TokenType::All {
                    self.advance();
                    self.expect(TokenType::LParen)?;
                    let right = if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                        Expr::Subquery(Box::new(self.parse_statement_inner()?))
                    } else {
                        self.parse_expr()?
                    };
                    self.expect(TokenType::RParen)?;
                    left = Expr::AllOp {
                        expr: Box::new(left),
                        op,
                        right: Box::new(right),
                    };
                } else {
                    let right = self.parse_addition()?;
                    left = Expr::BinaryOp {
                        left: Box::new(left),
                        op,
                        right: Box::new(right),
                    };
                }
            } else if self.peek_type() == &TokenType::Is {
                self.advance();
                let negated = self.match_token(TokenType::Not);
                if self.match_token(TokenType::True) {
                    left = Expr::IsBool {
                        expr: Box::new(left),
                        value: true,
                        negated,
                    };
                } else if self.match_token(TokenType::False) {
                    left = Expr::IsBool {
                        expr: Box::new(left),
                        value: false,
                        negated,
                    };
                } else if self.match_token(TokenType::Distinct) {
                    // SQL-standard `IS [NOT] DISTINCT FROM y` — null-safe
                    // comparison. We lower it to `(x <> y OR (x IS NULL) <>
                    // (y IS NULL))` for `DISTINCT FROM` (negated == false) and
                    // its inverse for `NOT DISTINCT FROM`. To keep the AST
                    // simple, model both as a binary inequality / equality
                    // wrapped in BinaryOp so the surrounding query parses.
                    self.expect(TokenType::From)?;
                    let right = self.parse_addition()?;
                    let op = if negated {
                        BinaryOperator::Eq
                    } else {
                        BinaryOperator::Neq
                    };
                    left = Expr::BinaryOp {
                        left: Box::new(left),
                        op,
                        right: Box::new(right),
                    };
                } else if matches!(self.peek_type(), TokenType::Json | TokenType::Jsonb)
                    || self.peek().value.eq_ignore_ascii_case("DOCUMENT")
                    || self.peek().value.eq_ignore_ascii_case("UNKNOWN")
                {
                    // PG / Db2 / SQL:2016 `expr IS [NOT] JSON [VALUE|ARRAY|
                    // OBJECT|SCALAR] [WITH|WITHOUT UNIQUE [KEYS]]`,
                    // `IS [NOT] DOCUMENT`, `IS [NOT] UNKNOWN`. We don't model
                    // these — fold to IsNull as a placeholder so the surrounding
                    // expression parses.
                    self.advance();
                    // Optional JSON kind keyword.
                    if matches!(
                        self.peek().value.to_uppercase().as_str(),
                        "VALUE" | "ARRAY" | "OBJECT" | "SCALAR"
                    ) && self.is_name_token()
                    {
                        self.advance();
                    }
                    // Optional `WITH|WITHOUT UNIQUE [KEYS]`.
                    if matches!(
                        self.peek().value.to_uppercase().as_str(),
                        "WITH" | "WITHOUT"
                    ) && self.is_name_token()
                    {
                        self.advance();
                        if self.peek().value.eq_ignore_ascii_case("UNIQUE") {
                            self.advance();
                            if self.peek().value.eq_ignore_ascii_case("KEYS") {
                                self.advance();
                            }
                        }
                    }
                    left = Expr::IsNull {
                        expr: Box::new(left),
                        negated,
                    };
                } else {
                    self.expect(TokenType::Null)?;
                    left = Expr::IsNull {
                        expr: Box::new(left),
                        negated,
                    };
                }
            } else if matches!(
                self.peek_type(),
                TokenType::Not
                    | TokenType::In
                    | TokenType::Like
                    | TokenType::ILike
                    | TokenType::Between
            ) {
                // Peek ahead: if NOT, only consume it if followed by IN/LIKE/ILIKE/BETWEEN
                if self.peek_type() == &TokenType::Not {
                    let saved_pos = self.pos;
                    self.advance(); // consume NOT
                    if !matches!(
                        self.peek_type(),
                        TokenType::In | TokenType::Like | TokenType::ILike | TokenType::Between
                    ) {
                        // NOT is not part of a comparison predicate — restore position
                        self.pos = saved_pos;
                        break;
                    }
                    // NOT was consumed, negated = true
                }
                let negated =
                    self.pos > 0 && self.tokens[self.pos - 1].token_type == TokenType::Not;

                if self.match_token(TokenType::In) {
                    // ClickHouse: `x IN [1, 2, 3]` — array literal directly
                    // after IN. Parse the array as the RHS and model as a
                    // single-element InList so downstream code emits IN (…).
                    if matches!(self.peek_type(), TokenType::LBracket) {
                        let rhs = self.parse_primary()?;
                        left = Expr::InList {
                            expr: Box::new(left),
                            list: vec![rhs],
                            negated,
                        };
                        continue;
                    }
                    // ClickHouse: `x IN funcCall(...)` / `x IN tableName` —
                    // bare function call or identifier as RHS. Parse a
                    // single primary expression and wrap as InList.
                    if !matches!(self.peek_type(), TokenType::LParen) {
                        let rhs = self.parse_primary()?;
                        left = Expr::InList {
                            expr: Box::new(left),
                            list: vec![rhs],
                            negated,
                        };
                        continue;
                    }
                    self.expect(TokenType::LParen)?;
                    // Check for subquery
                    if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                        let subquery = self.parse_statement_inner()?;
                        // ClickHouse accepts `IN ((SELECT ...) AS alias)`.
                        if self.match_token(TokenType::As) && self.is_name_token() {
                            self.advance();
                        } else if self.is_name_token() {
                            // also tolerate alias without AS
                            self.advance();
                        }
                        self.expect(TokenType::RParen)?;
                        left = Expr::InSubquery {
                            expr: Box::new(left),
                            subquery: Box::new(subquery),
                            negated,
                        };
                    } else {
                        let list = self.parse_expr_list()?;
                        self.expect(TokenType::RParen)?;
                        left = Expr::InList {
                            expr: Box::new(left),
                            list,
                            negated,
                        };
                    }
                } else if self.match_token(TokenType::Like) {
                    let pattern = self.parse_addition()?;
                    let escape = if self.match_token(TokenType::Escape) {
                        Some(Box::new(self.parse_primary()?))
                    } else {
                        None
                    };
                    left = Expr::Like {
                        expr: Box::new(left),
                        pattern: Box::new(pattern),
                        negated,
                        escape,
                    };
                } else if self.match_token(TokenType::ILike) {
                    let pattern = self.parse_addition()?;
                    let escape = if self.match_token(TokenType::Escape) {
                        Some(Box::new(self.parse_primary()?))
                    } else {
                        None
                    };
                    left = Expr::ILike {
                        expr: Box::new(left),
                        pattern: Box::new(pattern),
                        negated,
                        escape,
                    };
                } else if self.match_token(TokenType::Between) {
                    let low = self.parse_addition()?;
                    self.expect(TokenType::And)?;
                    let high = self.parse_addition()?;
                    left = Expr::Between {
                        expr: Box::new(left),
                        low: Box::new(low),
                        high: Box::new(high),
                        negated,
                    };
                } else {
                    break;
                }
            } else if self.check_keyword("SIMILAR") {
                // SIMILAR TO pattern [ESCAPE escape_char]
                self.advance(); // consume SIMILAR
                self.expect_keyword("TO")?;
                let pattern = self.parse_addition()?;
                let escape = if self.match_token(TokenType::Escape) {
                    Some(Box::new(self.parse_primary()?))
                } else {
                    None
                };
                left = Expr::SimilarTo {
                    expr: Box::new(left),
                    pattern: Box::new(pattern),
                    negated: false,
                    escape,
                };
            } else if self.peek_type() == &TokenType::Not && self.check_keyword_offset("SIMILAR", 1)
            {
                // NOT SIMILAR TO pattern [ESCAPE escape_char]
                self.advance(); // consume NOT
                self.advance(); // consume SIMILAR
                self.expect_keyword("TO")?;
                let pattern = self.parse_addition()?;
                let escape = if self.match_token(TokenType::Escape) {
                    Some(Box::new(self.parse_primary()?))
                } else {
                    None
                };
                left = Expr::SimilarTo {
                    expr: Box::new(left),
                    pattern: Box::new(pattern),
                    negated: true,
                    escape,
                };
            } else if self.check_keyword("REGEXP")
                || self.check_keyword("RLIKE")
                || self.check_keyword("GLOB")
                || self.check_keyword("IREGEXP")
            {
                // MySQL / Hive `expr REGEXP pat`, `expr RLIKE pat`, and
                // SQLite / DuckDB `expr GLOB pat`. Modeled as a Like with
                // no escape.
                self.advance();
                let pattern = self.parse_addition()?;
                left = Expr::Like {
                    expr: Box::new(left),
                    pattern: Box::new(pattern),
                    negated: false,
                    escape: None,
                };
            } else if self.peek_type() == &TokenType::Not
                && (self.check_keyword_offset("REGEXP", 1)
                    || self.check_keyword_offset("RLIKE", 1)
                    || self.check_keyword_offset("GLOB", 1)
                    || self.check_keyword_offset("IREGEXP", 1))
            {
                self.advance();
                self.advance();
                let pattern = self.parse_addition()?;
                left = Expr::Like {
                    expr: Box::new(left),
                    pattern: Box::new(pattern),
                    negated: true,
                    escape: None,
                };
            } else {
                break;
            }
        }

        Ok(left)
    }

    fn parse_addition(&mut self) -> Result<Expr> {
        let mut left = self.parse_multiplication()?;
        loop {
            let op = match self.peek_type() {
                TokenType::Plus => Some(BinaryOperator::Plus),
                TokenType::Minus => Some(BinaryOperator::Minus),
                TokenType::Concat => Some(BinaryOperator::Concat),
                TokenType::BitwiseOr => {
                    // Don't consume `|` when it is the start of `|>`; that
                    // is handled at comparison level (PG range/geom op).
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::Gt)
                    ) {
                        None
                    } else {
                        Some(BinaryOperator::BitwiseOr)
                    }
                }
                TokenType::BitwiseXor => {
                    // Preserve PostgreSQL `^@` for comparison-level handling.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::AtSign)
                    ) {
                        None
                    } else {
                        Some(BinaryOperator::BitwiseXor)
                    }
                }
                TokenType::ShiftLeft => {
                    // Preserve PostgreSQL `<<|` for comparison-level handling.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::BitwiseOr)
                    ) {
                        None
                    } else {
                        Some(BinaryOperator::ShiftLeft)
                    }
                }
                TokenType::ShiftRight => {
                    // Preserve PostgreSQL `>>|` for comparison-level handling.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::BitwiseOr)
                    ) {
                        None
                    } else {
                        Some(BinaryOperator::ShiftRight)
                    }
                }
                _ => None,
            };
            if let Some(op) = op {
                self.advance();
                // Oracle SQL*Plus continuation: `2359-\n,'AR'` keeps the
                // trailing `-` in the token stream. If the operator has no
                // valid right operand (next token is a delimiter), rewind
                // and treat the `-` as a no-op so the surrounding INSERT /
                // tuple keeps parsing.
                if matches!(op, BinaryOperator::Minus | BinaryOperator::Plus)
                    && matches!(
                        self.peek_type(),
                        TokenType::Comma
                            | TokenType::RParen
                            | TokenType::RBracket
                            | TokenType::Eof
                            | TokenType::Semicolon
                    )
                {
                    continue;
                }
                let right = self.parse_multiplication()?;
                left = Expr::BinaryOp {
                    left: Box::new(left),
                    op,
                    right: Box::new(right),
                };
            } else {
                break;
            }
        }
        Ok(left)
    }

    fn parse_multiplication(&mut self) -> Result<Expr> {
        let mut left = self.parse_unary()?;
        loop {
            let op = match self.peek_type() {
                TokenType::Star => Some(BinaryOperator::Multiply),
                TokenType::Slash => {
                    // DuckDB / Python-style integer division `//` — consume
                    // both slashes and lower to Divide so the surrounding
                    // expression parses.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::Slash)
                    ) {
                        self.advance();
                        self.advance();
                        let right = self.parse_unary()?;
                        left = Expr::BinaryOp {
                            left: Box::new(left),
                            op: BinaryOperator::Divide,
                            right: Box::new(right),
                        };
                        continue;
                    }
                    Some(BinaryOperator::Divide)
                }
                TokenType::Percent2 => Some(BinaryOperator::Modulo),
                TokenType::BitwiseAnd => {
                    // Don't consume the first `&` when it is the start of a
                    // multi-char PG operator (`&&`, `&<`, `&>`); leave it for
                    // the comparison-level handler.
                    if matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::BitwiseAnd) | Some(TokenType::Lt) | Some(TokenType::Gt)
                    ) {
                        None
                    } else {
                        Some(BinaryOperator::BitwiseAnd)
                    }
                }
                _ => {
                    // MySQL / ClickHouse keyword operators `DIV` (integer
                    // divide) and `MOD` (modulo). Treated as multiplicative.
                    if self.check_keyword("DIV") {
                        Some(BinaryOperator::Divide)
                    } else if self.check_keyword("MOD") {
                        Some(BinaryOperator::Modulo)
                    } else {
                        None
                    }
                }
            };
            if let Some(op) = op {
                self.advance();
                let right = self.parse_unary()?;
                left = Expr::BinaryOp {
                    left: Box::new(left),
                    op,
                    right: Box::new(right),
                };
            } else {
                break;
            }
        }
        Ok(left)
    }

    fn parse_unary(&mut self) -> Result<Expr> {
        match self.peek_type() {
            TokenType::Minus => {
                self.advance();
                let expr = self.parse_unary()?;
                Ok(Expr::UnaryOp {
                    op: UnaryOperator::Minus,
                    expr: Box::new(expr),
                })
            }
            TokenType::Plus => {
                self.advance();
                let expr = self.parse_unary()?;
                Ok(Expr::UnaryOp {
                    op: UnaryOperator::Plus,
                    expr: Box::new(expr),
                })
            }
            TokenType::BitwiseNot => {
                self.advance();
                let expr = self.parse_unary()?;
                Ok(Expr::UnaryOp {
                    op: UnaryOperator::BitwiseNot,
                    expr: Box::new(expr),
                })
            }
            _ => self.parse_postfix(),
        }
    }

    /// Parse postfix operators: `::type`, `[index]`, `->`, `->>`
    fn parse_postfix(&mut self) -> Result<Expr> {
        let mut expr = self.parse_primary()?;

        loop {
            if self.match_token(TokenType::DoubleColon) {
                // PostgreSQL-style cast: expr::type
                let data_type = self.parse_data_type()?;
                expr = Expr::Cast {
                    expr: Box::new(expr),
                    data_type,
                };
            } else if self.match_token(TokenType::LBracket) {
                // DuckDB list slicing: expr[start:end] or expr[:end] or expr[start:].
                // We model both index and slice as ArrayIndex (the slice
                // expression is discarded — the bench cares only about parse
                // acceptance).
                if self.match_token(TokenType::RBracket) {
                    // ClickHouse JSON empty subscript: `arr.k1[]` projects
                    // through every element. Treat as `ArrayIndex` against
                    // `NULL` so the surrounding expression parses.
                    expr = Expr::ArrayIndex {
                        expr: Box::new(expr),
                        index: Box::new(Expr::Null),
                    };
                } else if self.match_token(TokenType::Colon) {
                    // [:end] or [:end:step]
                    if !matches!(self.peek_type(), TokenType::RBracket | TokenType::Colon) {
                        let _ = self.parse_expr()?;
                    }
                    if self.match_token(TokenType::Colon)
                        && !matches!(self.peek_type(), TokenType::RBracket)
                    {
                        let _ = self.parse_expr()?;
                    }
                    self.expect(TokenType::RBracket)?;
                    expr = Expr::ArrayIndex {
                        expr: Box::new(expr),
                        index: Box::new(Expr::Null),
                    };
                } else {
                    let index = self.parse_expr()?;
                    if self.match_token(TokenType::Colon) {
                        // [start:end] / [start:] / [start:end:step] / [start::step]
                        if !matches!(self.peek_type(), TokenType::RBracket | TokenType::Colon) {
                            let _ = self.parse_expr()?;
                        }
                        if self.match_token(TokenType::Colon)
                            && !matches!(self.peek_type(), TokenType::RBracket)
                        {
                            let _ = self.parse_expr()?;
                        }
                    }
                    self.expect(TokenType::RBracket)?;
                    expr = Expr::ArrayIndex {
                        expr: Box::new(expr),
                        index: Box::new(index),
                    };
                }
            } else if self.match_token(TokenType::Arrow) {
                let path = self.parse_primary()?;
                expr = Expr::JsonAccess {
                    expr: Box::new(expr),
                    path: Box::new(path),
                    as_text: false,
                };
            } else if self.match_token(TokenType::DoubleArrow) {
                let path = self.parse_primary()?;
                expr = Expr::JsonAccess {
                    expr: Box::new(expr),
                    path: Box::new(path),
                    as_text: true,
                };
            } else if self.peek_type() == &TokenType::Colon
                && self
                    .peek_offset(1)
                    .map(|t| matches!(t.token_type, TokenType::Identifier))
                    .unwrap_or(false)
                && matches!(
                    expr,
                    Expr::Column { .. }
                        | Expr::JsonAccess { .. }
                        | Expr::Cast { .. }
                        | Expr::ArrayIndex { .. }
                )
            {
                // Snowflake VARIANT path accessor: `col:key`, `col:a:b`,
                // `col:a.b`. Treat each `:<name>` as a JSON access. We avoid
                // ambiguity with bind parameters (`:name`) by gating on a
                // preceding identifier-style expression.
                self.advance(); // :
                let part = self.advance().clone();
                expr = Expr::JsonAccess {
                    expr: Box::new(expr),
                    path: Box::new(Expr::StringLiteral(part.value)),
                    as_text: false,
                };
            } else if self.match_token(TokenType::Collate) {
                // Postgres / Spark `expr COLLATE collation_name` — we don't
                // model collations in the AST; consume the collation name
                // and continue. Accept any identifier-or-keyword name token.
                if self.is_name_token() || matches!(self.peek_type(), TokenType::String) {
                    self.advance();
                }
            } else if self.check_keyword("AT")
                && self
                    .peek_offset(1)
                    .map(|t| t.value.eq_ignore_ascii_case("TIME"))
                    .unwrap_or(false)
                && self
                    .peek_offset(2)
                    .map(|t| t.value.eq_ignore_ascii_case("ZONE"))
                    .unwrap_or(false)
            {
                // PostgreSQL / DuckDB: `expr AT TIME ZONE 'tz'`. Swallow the
                // suffix; the timezone-shifted value attaches to `expr`.
                self.advance(); // AT
                self.advance(); // TIME
                self.advance(); // ZONE
                let _ = self.parse_primary()?;
            } else if self.check_keyword("EXPORT_STATE")
                && matches!(expr, Expr::Function { .. } | Expr::TypedFunction { .. })
            {
                // DuckDB postfix `agg(...) EXPORT_STATE` returning the
                // serialized aggregate state instead of its final value.
                self.advance();
            } else if self.peek_type() == &TokenType::Dot
                && matches!(
                    self.peek_offset(1).map(|t| &t.token_type),
                    Some(TokenType::Colon | TokenType::BitwiseXor)
                )
            {
                // ClickHouse typed/subobject access after complex expressions:
                //   `expr.:Int64`, `expr.^a`, `expr.:`Array(Nullable(Int64))``.
                self.advance(); // .
                let _ = self.match_token(TokenType::BitwiseXor);
                let _ = self.match_token(TokenType::Colon);
                if self.is_name_token()
                    || self.is_data_type_token()
                    || matches!(self.peek_type(), TokenType::Null | TokenType::Identifier)
                {
                    let part = self.advance().clone();
                    expr = Expr::JsonAccess {
                        expr: Box::new(expr),
                        path: Box::new(Expr::StringLiteral(part.value)),
                        as_text: false,
                    };
                } else {
                    return Err(SqlglotError::UnexpectedToken {
                        token: self.peek().clone(),
                    });
                }
            } else if self.peek_type() == &TokenType::Dot
                && matches!(
                    self.peek_offset(1).map(|t| &t.token_type),
                    Some(TokenType::Number)
                )
            {
                // ClickHouse tuple element access: `t.1`, `t[1].2`. Model as
                // an ArrayIndex on a numeric literal so the surrounding
                // expression parses.
                self.advance(); // .
                let n = self.advance().clone();
                expr = Expr::ArrayIndex {
                    expr: Box::new(expr),
                    index: Box::new(Expr::Number(n.value)),
                };
            } else if self.peek_type() == &TokenType::Dot
                && self
                    .peek_offset(1)
                    .map(|t| matches!(t.token_type, TokenType::Identifier))
                    .unwrap_or(false)
            {
                // Postfix field access after a non-primary expression
                // (e.g. `arr[].field`, `arr.k1[].k2.k3`). Also handles
                // DuckDB method-call style `expr.method(args)` by
                // rewriting to `method(expr, args)`.
                self.advance(); // .
                let part = self.advance().clone();
                if self.match_token(TokenType::LParen) {
                    let mut args = vec![expr];
                    if self.peek_type() != &TokenType::RParen {
                        args.push(self.parse_function_arg()?);
                        while self.match_token(TokenType::Comma) {
                            args.push(self.parse_function_arg()?);
                        }
                    }
                    self.expect(TokenType::RParen)?;
                    expr = Expr::Function {
                        name: part.value,
                        args,
                        distinct: false,
                        within_group: false,
                        order_by: vec![],
                        filter: None,
                        over: None,
                    };
                } else {
                    expr = Expr::JsonAccess {
                        expr: Box::new(expr),
                        path: Box::new(Expr::StringLiteral(part.value)),
                        as_text: false,
                    };
                }
            } else if matches!(expr, Expr::Function { .. })
                && self.peek_type() == &TokenType::LParen
            {
                // ClickHouse combinator-style application: `f(a)(b)` —
                // apply the result of `f(a)` to `(b)`. We model this as a
                // nested function call where the outer call's name is the
                // serialized inner function-call expression — we just pack
                // both arg lists into a single Function node so the parse
                // does not stop here.
                // apply the result of `f(a)` to `(b)`. We model this as a
                // nested function call where the outer call's name is the
                // serialized inner function-call expression — we just pack
                // both arg lists into a single Function node so the parse
                // does not stop here.
                self.advance();
                let extra_args = if self.peek_type() != &TokenType::RParen {
                    let mut a = vec![self.parse_function_arg()?];
                    while self.match_token(TokenType::Comma) {
                        a.push(self.parse_function_arg()?);
                    }
                    a
                } else {
                    vec![]
                };
                self.expect(TokenType::RParen)?;
                if let Expr::Function {
                    name,
                    mut args,
                    distinct,
                    filter,
                    over,
                    order_by,
                    within_group,
                } = expr
                {
                    args.extend(extra_args);
                    expr = Expr::Function {
                        name,
                        args,
                        distinct,
                        filter,
                        over,
                        order_by,
                        within_group,
                    };
                } else {
                    unreachable!();
                }
            } else {
                break;
            }
        }

        // Check for window function: expr OVER (...)
        // BigQuery / DuckDB / ClickHouse / Snowflake: window-function nulls
        // modifier outside the call: `first_value(x) IGNORE NULLS OVER (...)`
        // or `first_value(x) RESPECT NULLS`. Swallow opaquely.
        if (self.peek().value.eq_ignore_ascii_case("IGNORE")
            || self.peek().value.eq_ignore_ascii_case("RESPECT"))
            && self
                .peek_offset(1)
                .map(|t| t.token_type == TokenType::Null || t.value.eq_ignore_ascii_case("NULLS"))
                .unwrap_or(false)
        {
            self.advance();
            self.advance();
        }
        if self.match_token(TokenType::Over) {
            let spec = if self.match_token(TokenType::LParen) {
                let ws = self.parse_window_spec()?;
                self.expect(TokenType::RParen)?;
                ws
            } else {
                // Named window reference
                let wref = self.expect_name()?;
                WindowSpec {
                    window_ref: Some(wref),
                    partition_by: vec![],
                    order_by: vec![],
                    frame: None,
                }
            };
            match expr {
                Expr::Function {
                    name,
                    args,
                    distinct,
                    filter,
                    order_by,
                    within_group,
                    ..
                } => {
                    expr = Expr::Function {
                        name,
                        args,
                        distinct,
                        filter,
                        over: Some(spec),
                        order_by,
                        within_group,
                    };
                }
                Expr::TypedFunction { func, filter, .. } => {
                    expr = Expr::TypedFunction {
                        func,
                        filter,
                        over: Some(spec),
                    };
                }
                _ => {}
            }
        }

        // FILTER (WHERE ...) for aggregate functions
        if self.match_token(TokenType::Filter) {
            self.expect(TokenType::LParen)?;
            self.expect(TokenType::Where)?;
            let filter_expr = self.parse_expr()?;
            self.expect(TokenType::RParen)?;
            match expr {
                Expr::Function {
                    name,
                    args,
                    distinct,
                    over,
                    order_by,
                    within_group,
                    ..
                } => {
                    expr = Expr::Function {
                        name,
                        args,
                        distinct,
                        filter: Some(Box::new(filter_expr)),
                        over,
                        order_by,
                        within_group,
                    };
                }
                Expr::TypedFunction { func, over, .. } => {
                    expr = Expr::TypedFunction {
                        func,
                        filter: Some(Box::new(filter_expr)),
                        over,
                    };
                }
                _ => {}
            }
            // PostgreSQL / DuckDB: `agg(x) FILTER (WHERE …) OVER (…)`.
            // Parse the trailing OVER clause after FILTER so window-call
            // aggregates with filters still resolve.
            if self.match_token(TokenType::Over) {
                let spec = if self.match_token(TokenType::LParen) {
                    let ws = self.parse_window_spec()?;
                    self.expect(TokenType::RParen)?;
                    ws
                } else {
                    let wref = self.expect_name()?;
                    WindowSpec {
                        window_ref: Some(wref),
                        partition_by: vec![],
                        order_by: vec![],
                        frame: None,
                    }
                };
                match expr {
                    Expr::Function {
                        name,
                        args,
                        distinct,
                        filter,
                        order_by,
                        within_group,
                        ..
                    } => {
                        expr = Expr::Function {
                            name,
                            args,
                            distinct,
                            filter,
                            over: Some(spec),
                            order_by,
                            within_group,
                        };
                    }
                    Expr::TypedFunction { func, filter, .. } => {
                        expr = Expr::TypedFunction {
                            func,
                            filter,
                            over: Some(spec),
                        };
                    }
                    _ => {}
                }
            }
        }

        Ok(expr)
    }

    fn parse_window_spec(&mut self) -> Result<WindowSpec> {
        let window_ref = if self.is_name_token()
            && !matches!(
                self.peek_type(),
                TokenType::Partition | TokenType::Order | TokenType::Rows | TokenType::Range
            ) {
            let saved = self.pos;
            let name = self.expect_name()?;
            // Check if it's actually a keyword we need
            if matches!(
                self.peek_type(),
                TokenType::RParen
                    | TokenType::Partition
                    | TokenType::Order
                    | TokenType::Rows
                    | TokenType::Range
            ) {
                Some(name)
            } else {
                self.pos = saved;
                None
            }
        } else {
            None
        };

        let partition_by = if self.match_token(TokenType::Partition) {
            self.expect(TokenType::By)?;
            self.parse_expr_list_allow_item_alias()?
        } else if self.is_name_token()
            && (self.peek().value.eq_ignore_ascii_case("DISTRIBUTE")
                || self.peek().value.eq_ignore_ascii_case("CLUSTER"))
        {
            // Hive `DISTRIBUTE BY` / `CLUSTER BY` inside OVER(...) — treat
            // as PARTITION BY.
            self.advance();
            self.expect(TokenType::By)?;
            self.parse_expr_list_allow_item_alias()?
        } else {
            vec![]
        };

        let order_by = if self.match_token(TokenType::Order) {
            self.expect(TokenType::By)?;
            self.parse_order_by_items()?
        } else if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("SORT") {
            // Hive `SORT BY` inside OVER(...) — treat as ORDER BY.
            self.advance();
            self.expect(TokenType::By)?;
            self.parse_order_by_items()?
        } else {
            vec![]
        };

        let frame = if matches!(self.peek_type(), TokenType::Rows | TokenType::Range) {
            Some(self.parse_window_frame()?)
        } else {
            None
        };

        Ok(WindowSpec {
            window_ref,
            partition_by,
            order_by,
            frame,
        })
    }

    fn parse_window_frame(&mut self) -> Result<WindowFrame> {
        let kind = if self.match_token(TokenType::Rows) {
            WindowFrameKind::Rows
        } else if self.match_token(TokenType::Range) {
            WindowFrameKind::Range
        } else {
            WindowFrameKind::Rows
        };

        if self.match_keyword("BETWEEN") {
            let start = self.parse_window_frame_bound()?;
            self.expect(TokenType::And)?;
            let end = self.parse_window_frame_bound()?;
            // SQL:2011 / DuckDB frame exclusion clause:
            //   `EXCLUDE CURRENT ROW | EXCLUDE GROUP | EXCLUDE TIES |
            //    EXCLUDE NO OTHERS`. Swallow opaquely; we don't model it.
            if self.check_keyword("EXCLUDE") {
                self.advance();
                if self.check_keyword("CURRENT") {
                    self.advance();
                    let _ = self.match_keyword("ROW");
                } else if self.check_keyword("NO") {
                    self.advance();
                    let _ = self.match_keyword("OTHERS");
                } else if self.check_keyword("GROUP") || self.check_keyword("TIES") {
                    self.advance();
                }
            }
            Ok(WindowFrame {
                kind,
                start,
                end: Some(end),
            })
        } else {
            let start = self.parse_window_frame_bound()?;
            if self.check_keyword("EXCLUDE") {
                self.advance();
                if self.check_keyword("CURRENT") {
                    self.advance();
                    let _ = self.match_keyword("ROW");
                } else if self.check_keyword("NO") {
                    self.advance();
                    let _ = self.match_keyword("OTHERS");
                } else if self.check_keyword("GROUP") || self.check_keyword("TIES") {
                    self.advance();
                }
            }
            Ok(WindowFrame {
                kind,
                start,
                end: None,
            })
        }
    }

    fn parse_window_frame_bound(&mut self) -> Result<WindowFrameBound> {
        if self.check_keyword("CURRENT") {
            self.advance();
            let _ = self.match_keyword("ROW");
            Ok(WindowFrameBound::CurrentRow)
        } else if self.match_token(TokenType::Unbounded) {
            if self.match_token(TokenType::Preceding) {
                Ok(WindowFrameBound::Preceding(None))
            } else {
                self.expect(TokenType::Following)?;
                Ok(WindowFrameBound::Following(None))
            }
        } else {
            let n = self.parse_expr()?;
            if self.match_token(TokenType::Preceding) {
                Ok(WindowFrameBound::Preceding(Some(Box::new(n))))
            } else {
                self.expect(TokenType::Following)?;
                Ok(WindowFrameBound::Following(Some(Box::new(n))))
            }
        }
    }

    fn parse_primary(&mut self) -> Result<Expr> {
        let token = self.peek().clone();

        // DuckDB / Spark leading-dot float literal: `.5`, `.25`. The
        // tokenizer emits `Dot` then `Number`; glue them back together.
        if matches!(token.token_type, TokenType::Dot)
            && matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::Number)
            )
        {
            self.advance();
            let n = self.peek().value.clone();
            self.advance();
            return Ok(Expr::Number(format!("0.{}", n)));
        }

        match &token.token_type {
            TokenType::Number => {
                self.advance();
                // Trailing-dot fractional literal: `10.` — accept the dot as
                // part of the number when it isn't followed by something that
                // would be a member access (column reference like `t.col` or
                // tuple element access).
                let mut value = token.value;
                if self.peek_type() == &TokenType::Dot {
                    let after = self.peek_offset(1).map(|t| &t.token_type);
                    let looks_like_member = matches!(
                        after,
                        Some(TokenType::Identifier)
                            | Some(TokenType::Number)
                            | Some(TokenType::Star)
                    );
                    if !looks_like_member {
                        self.advance();
                        value.push('.');
                    }
                }
                // Spark / Hive float suffixes: `10.0F`, `20L`, `3.14D`, `5BD`.
                // Swallow the suffix identifier so the literal parses.
                if self.is_name_token() {
                    let v = self.peek().value.as_str();
                    if matches!(v, "F" | "f" | "L" | "l" | "D" | "d" | "BD" | "bd") {
                        self.advance();
                    }
                }
                Ok(Expr::Number(value))
            }
            TokenType::HexString => {
                self.advance();
                Ok(Expr::Number(token.value))
            }
            TokenType::String => {
                self.advance();
                // ANSI / Oracle interval literal: `'1-2' YEAR TO MONTH`,
                // `'12 03:04:05.6' DAY TO SECOND(2)`. After a bare string,
                // accept an optional interval qualifier and swallow it so
                // the surrounding expression parses. Skip this when the
                // previous token was `INTERVAL` — that has its own path.
                let prev_was_interval = self
                    .pos
                    .checked_sub(2)
                    .and_then(|i| self.tokens.get(i))
                    .map(|t| matches!(t.token_type, TokenType::Interval))
                    .unwrap_or(false);
                if !prev_was_interval
                    && matches!(
                        self.peek_type(),
                        TokenType::Year
                            | TokenType::Month
                            | TokenType::Day
                            | TokenType::Hour
                            | TokenType::Minute
                            | TokenType::Second
                    )
                {
                    self.advance();
                    if self.match_token(TokenType::LParen) {
                        // qualifier precision: `SECOND(2)`
                        if matches!(self.peek_type(), TokenType::Number) {
                            self.advance();
                            if self.match_token(TokenType::Comma) {
                                if matches!(self.peek_type(), TokenType::Number) {
                                    self.advance();
                                }
                            }
                        }
                        let _ = self.match_token(TokenType::RParen);
                    }
                    if self.is_name_token() && self.peek().value.eq_ignore_ascii_case("TO") {
                        self.advance();
                        if matches!(
                            self.peek_type(),
                            TokenType::Year
                                | TokenType::Month
                                | TokenType::Day
                                | TokenType::Hour
                                | TokenType::Minute
                                | TokenType::Second
                        ) {
                            self.advance();
                            if self.match_token(TokenType::LParen) {
                                if matches!(self.peek_type(), TokenType::Number) {
                                    self.advance();
                                }
                                let _ = self.match_token(TokenType::RParen);
                            }
                        }
                    }
                    return Ok(Expr::Cast {
                        expr: Box::new(Expr::StringLiteral(token.value)),
                        data_type: DataType::Interval,
                    });
                }
                // SQL-92 / MySQL: adjacent string literals concatenate
                // (`'a' 'b'` → `'ab'`). Also fold in identifier-quoted
                // strings the lexer surfaces when MySQL ANSI_QUOTES is off
                // (`"a" "b" "c"` reaches us as a String followed by quoted
                // identifiers). Greedily consume any run of immediately
                // following String / quoted-Identifier tokens.
                let mut combined = token.value;
                loop {
                    let next = self.peek();
                    if matches!(next.token_type, TokenType::String) {
                        combined.push_str(&next.value);
                        self.advance();
                        continue;
                    }
                    if matches!(next.token_type, TokenType::Identifier)
                        && (next.quote_char == '"' || next.quote_char == '\'')
                    {
                        combined.push_str(&next.value);
                        self.advance();
                        continue;
                    }
                    break;
                }
                Ok(Expr::StringLiteral(combined))
            }
            TokenType::NationalString => {
                self.advance();
                Ok(Expr::NationalStringLiteral(token.value))
            }
            TokenType::True => {
                self.advance();
                Ok(Expr::Boolean(true))
            }
            TokenType::False => {
                self.advance();
                Ok(Expr::Boolean(false))
            }
            TokenType::Null => {
                self.advance();
                Ok(Expr::Null)
            }
            TokenType::Default => {
                self.advance();
                // MySQL `DEFAULT(col)` — emit as function call so the
                // surrounding tuple parses.
                if self.peek_type() == &TokenType::LParen {
                    self.advance();
                    let args = if self.peek_type() != &TokenType::RParen {
                        let mut a = vec![self.parse_function_arg()?];
                        while self.match_token(TokenType::Comma) {
                            a.push(self.parse_function_arg()?);
                        }
                        a
                    } else {
                        vec![]
                    };
                    self.expect(TokenType::RParen)?;
                    return Ok(Expr::Function {
                        name: "DEFAULT".to_string(),
                        args,
                        distinct: false,
                        filter: None,
                        over: None,
                        order_by: Vec::new(),
                        within_group: false,
                    });
                }
                Ok(Expr::Default)
            }
            TokenType::Star => {
                self.advance();
                Ok(Expr::Wildcard)
            }
            // ClickHouse / various: `values` used as a column name inside
            // expressions (e.g. `arrayExists(x -> x > 5, values)`). Accept
            // it as a bare column reference when it isn't followed by `(`.
            TokenType::Values
                if self.peek_offset(1).map(|t| &t.token_type) != Some(&TokenType::LParen) =>
            {
                self.advance();
                Ok(Expr::Column {
                    table: None,
                    name: token.value,
                    quote_style: QuoteStyle::None,
                    table_quote_style: QuoteStyle::None,
                })
            }
            TokenType::Parameter => {
                self.advance();
                Ok(Expr::Parameter(token.value))
            }

            // ── `@var`, `@@global_var`, `:var` style placeholders ──
            //
            // MySQL/T-SQL session and global variables tokenize as a bare
            // `@` (or `:`) followed by an identifier. We glue the prefix and
            // following name into a single `Parameter` expression so the
            // surrounding query parses.
            TokenType::AtSign | TokenType::Colon => {
                self.advance();
                let mut name = match token.token_type {
                    TokenType::AtSign => String::from("@"),
                    TokenType::Colon => String::from(":"),
                    _ => unreachable!(),
                };
                // T-SQL `@@global` — second `@`.
                if matches!(token.token_type, TokenType::AtSign)
                    && self.peek_type() == &TokenType::AtSign
                {
                    name.push('@');
                    self.advance();
                }
                // Name part: identifier-or-keyword, number, or none.
                // T-SQL accepts reserved keywords after `@` (e.g. `@limit`,
                // `@order`). Accept any token that "looks like" a name.
                if self.is_name_token()
                    || matches!(
                        self.peek_type(),
                        TokenType::Limit
                            | TokenType::Offset
                            | TokenType::Order
                            | TokenType::Group
                            | TokenType::Having
                            | TokenType::Where
                            | TokenType::From
                            | TokenType::Select
                            | TokenType::Insert
                            | TokenType::Update
                            | TokenType::Delete
                            | TokenType::Union
                            | TokenType::Intersect
                            | TokenType::Except
                            | TokenType::Join
                            | TokenType::Inner
                            | TokenType::Cross
                            | TokenType::On
                            | TokenType::As
                            | TokenType::Distinct
                            | TokenType::Default
                            | TokenType::Null
                            | TokenType::True
                            | TokenType::False
                            | TokenType::Date
                            | TokenType::Time
                            | TokenType::Timestamp
                            | TokenType::Year
                            | TokenType::Month
                            | TokenType::Day
                            | TokenType::Hour
                            | TokenType::Minute
                            | TokenType::Second
                    )
                {
                    let nt = self.advance().clone();
                    name.push_str(&nt.value);
                } else if matches!(self.peek_type(), TokenType::Number | TokenType::Int) {
                    let nt = self.advance().clone();
                    name.push_str(&nt.value);
                }
                Ok(Expr::Parameter(name))
            }

            // ── DuckDB / BigQuery struct literal: `{ key: expr, ... }` ──
            //
            // We capture the values as positional `STRUCT(...)` arguments
            // (keys are syntactically optional). This keeps surrounding
            // expressions parseable; the original AST shape is not preserved
            // because there is no dedicated struct-literal variant yet.
            TokenType::LBrace => {
                self.advance();
                let mut args = Vec::new();
                if self.peek_type() != &TokenType::RBrace {
                    loop {
                        // Optional `key:` prefix — discard the key, keep value.
                        if self.is_name_token()
                            && self
                                .peek_offset(1)
                                .is_some_and(|t| t.token_type == TokenType::Colon)
                        {
                            self.advance(); // key
                            self.advance(); // colon
                        } else if self.peek_type() == &TokenType::String
                            && self
                                .peek_offset(1)
                                .is_some_and(|t| t.token_type == TokenType::Colon)
                        {
                            self.advance(); // string key
                            self.advance(); // colon
                        }
                        let value = self.parse_expr()?;
                        args.push(value);
                        if !self.match_token(TokenType::Comma) {
                            break;
                        }
                    }
                }
                self.expect(TokenType::RBrace)?;
                Ok(Expr::Function {
                    name: "STRUCT".to_string(),
                    args,
                    distinct: false,
                    filter: None,
                    over: None,
                    order_by: Vec::new(),
                    within_group: false,
                })
            }

            // ── CAST ────────────────────────────────────────────────
            TokenType::Cast
                if self
                    .peek_offset(1)
                    .is_some_and(|t| t.token_type == TokenType::LParen) =>
            {
                self.advance();
                self.expect(TokenType::LParen)?;
                let expr = self.parse_expr()?;
                // Standard form: `CAST(expr AS type)`. ClickHouse also accepts
                // `CAST(expr, 'TypeName')` with a string literal type.
                let data_type = if self.match_token(TokenType::As) {
                    self.parse_data_type()?
                } else if self.match_token(TokenType::Comma) {
                    if matches!(self.peek_type(), TokenType::String) {
                        let s = self.peek().value.clone();
                        self.advance();
                        DataType::Unknown(s)
                    } else {
                        self.parse_data_type()?
                    }
                } else {
                    self.expect(TokenType::As)?; // produce the canonical error
                    self.parse_data_type()?
                };
                // BigQuery: `CAST(expr AS type FORMAT 'fmt' [AT TIME ZONE …])`.
                if self.check_keyword("FORMAT") {
                    self.advance();
                    let _ = self.parse_expr();
                    if self.check_keyword("AT")
                        && self
                            .peek_offset(1)
                            .map(|t| t.value.eq_ignore_ascii_case("TIME"))
                            .unwrap_or(false)
                        && self
                            .peek_offset(2)
                            .map(|t| t.value.eq_ignore_ascii_case("ZONE"))
                            .unwrap_or(false)
                    {
                        self.advance();
                        self.advance();
                        self.advance();
                        let _ = self.parse_expr();
                    }
                }
                self.expect(TokenType::RParen)?;
                Ok(Expr::Cast {
                    expr: Box::new(expr),
                    data_type,
                })
            }

            // ── EXTRACT ─────────────────────────────────────────────
            TokenType::Extract => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let field = self.parse_datetime_field()?;
                self.expect(TokenType::From)?;
                let expr = self.parse_expr()?;
                // BigQuery: `EXTRACT(field FROM ts AT TIME ZONE 'tz')`.
                // Swallow the trailing timezone clause so the function
                // parses; we lose the explicit zone but keep the AST.
                if self.check_keyword("AT")
                    && self
                        .peek_offset(1)
                        .map(|t| t.value.eq_ignore_ascii_case("TIME"))
                        .unwrap_or(false)
                    && self
                        .peek_offset(2)
                        .map(|t| t.value.eq_ignore_ascii_case("ZONE"))
                        .unwrap_or(false)
                {
                    self.advance(); // AT
                    self.advance(); // TIME
                    self.advance(); // ZONE
                    let _ = self.parse_expr();
                }
                self.expect(TokenType::RParen)?;
                Ok(Expr::Extract {
                    field,
                    expr: Box::new(expr),
                })
            }

            // ── CASE ────────────────────────────────────────────────
            TokenType::Case => self.parse_case_expr(),

            // ── EXISTS ──────────────────────────────────────────────
            TokenType::Exists => {
                self.advance();
                self.expect(TokenType::LParen)?;
                let subquery = self.parse_statement_inner()?;
                self.expect(TokenType::RParen)?;
                Ok(Expr::Exists {
                    subquery: Box::new(subquery),
                    negated: false,
                })
            }

            // ── NOT EXISTS ──────────────────────────────────────────
            TokenType::Not
                if {
                    let next_pos = self.pos + 1;
                    next_pos < self.tokens.len()
                        && self.tokens[next_pos].token_type == TokenType::Exists
                } =>
            {
                self.advance(); // NOT
                self.advance(); // EXISTS
                self.expect(TokenType::LParen)?;
                let subquery = self.parse_statement_inner()?;
                self.expect(TokenType::RParen)?;
                Ok(Expr::Exists {
                    subquery: Box::new(subquery),
                    negated: true,
                })
            }

            // ── INTERVAL ────────────────────────────────────────────
            TokenType::Interval => {
                self.advance();
                // ClickHouse accepts arithmetic in the value position
                // (e.g. `INTERVAL number - 15 MONTH`). Parse an additive
                // expression instead of a single primary so the trailing
                // unit keyword is reached cleanly.
                let value = self.parse_addition()?;
                let unit = self.try_parse_datetime_field();
                // ANSI / Spark composite ranges: `INTERVAL '0-0' YEAR TO MONTH`,
                // `INTERVAL '15:40' HOUR TO MINUTE` etc. Swallow the trailing
                // `TO <unit>` clause; we keep only the leading unit.
                if self.check_keyword("TO") {
                    let saved = self.pos;
                    self.advance();
                    if self.try_parse_datetime_field().is_none() {
                        self.pos = saved;
                    }
                }
                // PostgreSQL fractional precision on the trailing unit:
                //   `INTERVAL '1.234' SECOND(2)`, `INTERVAL '…' MINUTE TO SECOND(2)`.
                // Swallow the `(N)` after the unit.
                if self.peek_type() == &TokenType::LParen
                    && self
                        .peek_offset(1)
                        .map(|t| matches!(t.token_type, TokenType::Number))
                        .unwrap_or(false)
                    && self
                        .peek_offset(2)
                        .map(|t| matches!(t.token_type, TokenType::RParen))
                        .unwrap_or(false)
                {
                    self.advance();
                    self.advance();
                    self.advance();
                }
                Ok(Expr::Interval {
                    value: Box::new(value),
                    unit,
                })
            }

            // ── Parenthesized expression or subquery ────────────────
            TokenType::LParen => {
                self.advance();
                // Check for subquery
                if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                    let subquery = self.parse_statement_inner()?;
                    self.expect(TokenType::RParen)?;
                    Ok(Expr::Subquery(Box::new(subquery)))
                } else {
                    let expr = self.parse_expr()?;
                    // ClickHouse: `(expr AS alias)` — swallow the alias.
                    if self.match_token(TokenType::As) && self.is_name_token() {
                        self.advance();
                    }
                    // Tuple: (a, b, c) — also accept ClickHouse trailing
                    // comma `(a,)`, `(a, b,)`.
                    if self.match_token(TokenType::Comma) {
                        let mut items = vec![expr];
                        if self.peek_type() == &TokenType::RParen {
                            self.advance();
                            return Ok(Expr::Tuple(items));
                        }
                        let next = self.parse_expr()?;
                        if self.match_token(TokenType::As) && self.is_name_token() {
                            self.advance();
                        }
                        items.push(next);
                        while self.match_token(TokenType::Comma) {
                            if self.peek_type() == &TokenType::RParen {
                                break;
                            }
                            let n = self.parse_expr()?;
                            if self.match_token(TokenType::As) && self.is_name_token() {
                                self.advance();
                            }
                            items.push(n);
                        }
                        self.expect(TokenType::RParen)?;
                        Ok(Expr::Tuple(items))
                    } else {
                        self.expect(TokenType::RParen)?;
                        Ok(Expr::Nested(Box::new(expr)))
                    }
                }
            }

            // ── DuckDB MAP literal: `MAP { 'k': v, ... }` ──────────
            // Captured as a `MAP(...)` function call with the values as
            // positional arguments; keys are discarded for now.
            TokenType::Map
                if self
                    .peek_offset(1)
                    .map(|t| matches!(t.token_type, TokenType::LBrace))
                    .unwrap_or(false) =>
            {
                self.advance(); // MAP
                self.advance(); // {
                let mut args = Vec::new();
                if self.peek_type() != &TokenType::RBrace {
                    loop {
                        // Optional `key:` prefix — keep the value only.
                        let saved = self.pos;
                        let _ = self.parse_expr()?;
                        if self.match_token(TokenType::Colon) {
                            let v = self.parse_expr()?;
                            args.push(v);
                        } else {
                            self.pos = saved;
                            let v = self.parse_expr()?;
                            args.push(v);
                        }
                        if !self.match_token(TokenType::Comma) {
                            break;
                        }
                    }
                }
                self.expect(TokenType::RBrace)?;
                Ok(Expr::Function {
                    name: "MAP".to_string(),
                    args,
                    distinct: false,
                    filter: None,
                    over: None,
                    order_by: Vec::new(),
                    within_group: false,
                })
            }

            // ── Array literal: ARRAY[...] ──────────────────────────
            TokenType::Array => {
                self.advance();
                if self.match_token(TokenType::LBracket) {
                    let items = self.parse_array_items(TokenType::RBracket)?;
                    self.expect(TokenType::RBracket)?;
                    Ok(Expr::ArrayLiteral(items))
                } else if self.match_token(TokenType::LParen) {
                    // ARRAY(SELECT ...) for subqueries, or Hive
                    // `ARRAY(expr, expr, ...)` for inline array literals.
                    if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
                        let subquery = self.parse_statement_inner()?;
                        self.expect(TokenType::RParen)?;
                        Ok(Expr::Subquery(Box::new(subquery)))
                    } else {
                        let items = self.parse_array_items(TokenType::RParen)?;
                        self.expect(TokenType::RParen)?;
                        Ok(Expr::ArrayLiteral(items))
                    }
                } else {
                    Ok(Expr::Column {
                        table: None,
                        name: "ARRAY".to_string(),
                        quote_style: QuoteStyle::None,
                        table_quote_style: QuoteStyle::None,
                    })
                }
            }

            // ── Bracket array literal: [...] ────────────────────────
            TokenType::LBracket => {
                self.advance();
                let items = self.parse_array_items(TokenType::RBracket)?;
                // DuckDB list comprehension: `[expr FOR x IN list [IF cond]]`.
                // Swallow the comprehension tail opaquely; we keep the
                // initial expression as the AST representation.
                if self.peek().value.eq_ignore_ascii_case("FOR") {
                    let mut depth = 1_i32;
                    while depth > 0 && !matches!(self.peek_type(), TokenType::Eof) {
                        match self.peek_type() {
                            TokenType::LBracket | TokenType::LParen => depth += 1,
                            TokenType::RBracket => {
                                depth -= 1;
                                if depth == 0 {
                                    break;
                                }
                            }
                            TokenType::RParen => depth -= 1,
                            _ => {}
                        }
                        self.advance();
                    }
                }
                self.expect(TokenType::RBracket)?;
                Ok(Expr::ArrayLiteral(items))
            }

            // ── Identifier: column ref, function call, or qualified name ─
            _ if self.is_name_token() || self.is_data_type_token() => {
                let name_token = self.advance().clone();
                let name = name_token.value.clone();
                let name_qs = quote_style_from_char(name_token.quote_char);

                // ── ANSI typed string literals: DATE 'x', TIMESTAMP 'x', TIME 'x' ──
                if matches!(
                    name_token.token_type,
                    TokenType::Date
                        | TokenType::Timestamp
                        | TokenType::TimestampTz
                        | TokenType::Time
                ) {
                    // PG / ANSI `TIMESTAMP [WITH [LOCAL] TIME ZONE] 'lit'`
                    // and `TIMESTAMP WITHOUT TIME ZONE 'lit'`. Swallow the
                    // optional timezone modifier so the string literal
                    // attaches to the right typed-literal form.
                    let mut explicit_tz: Option<bool> = None;
                    if matches!(
                        name_token.token_type,
                        TokenType::Timestamp | TokenType::Time
                    ) && self.peek_type() == &TokenType::With
                    {
                        let saved = self.pos;
                        self.advance(); // WITH
                        let _ = self.match_keyword("LOCAL");
                        if self.check_keyword("TIME")
                            && self
                                .peek_offset(1)
                                .map(|t| t.value.eq_ignore_ascii_case("ZONE"))
                                .unwrap_or(false)
                        {
                            self.advance(); // TIME
                            self.advance(); // ZONE
                            explicit_tz = Some(true);
                        } else {
                            self.pos = saved;
                        }
                    } else if matches!(
                        name_token.token_type,
                        TokenType::Timestamp | TokenType::Time
                    ) && self.check_keyword("WITHOUT")
                    {
                        let saved = self.pos;
                        self.advance(); // WITHOUT
                        if self.check_keyword("TIME")
                            && self
                                .peek_offset(1)
                                .map(|t| t.value.eq_ignore_ascii_case("ZONE"))
                                .unwrap_or(false)
                        {
                            self.advance();
                            self.advance();
                            explicit_tz = Some(false);
                        } else {
                            self.pos = saved;
                        }
                    }

                    if self.peek_type() == &TokenType::String {
                        let value_token = self.advance().clone();
                        let data_type = match name_token.token_type {
                            TokenType::Date => DataType::Date,
                            TokenType::Timestamp => DataType::Timestamp {
                                precision: None,
                                with_tz: explicit_tz.unwrap_or(false),
                            },
                            TokenType::TimestampTz => DataType::Timestamp {
                                precision: None,
                                with_tz: true,
                            },
                            TokenType::Time => DataType::Time { precision: None },
                            _ => unreachable!(),
                        };
                        return Ok(Expr::Cast {
                            expr: Box::new(Expr::StringLiteral(value_token.value)),
                            data_type,
                        });
                    }
                }

                // ── ANSI / PG generic typed string literal: `TYPE 'lit'` ──
                // (e.g. `bool 'true'`, `int4 '42'`, `varchar 'x'`). When the
                // current token is a data-type keyword (not already handled
                // above) and a String literal follows, fold the pair into a
                // Cast so the surrounding expression parses.
                if self.is_data_type_token_kind(&name_token.token_type)
                    && self.peek_type() == &TokenType::String
                {
                    let value_token = self.advance().clone();
                    let data_type = match name_token.token_type {
                        TokenType::Boolean => DataType::Boolean,
                        TokenType::Int | TokenType::Integer => DataType::Int,
                        TokenType::BigInt => DataType::BigInt,
                        TokenType::SmallInt => DataType::SmallInt,
                        TokenType::TinyInt => DataType::TinyInt,
                        TokenType::Float => DataType::Float,
                        TokenType::Double => DataType::Double,
                        TokenType::Real => DataType::Real,
                        TokenType::Decimal => DataType::Decimal {
                            precision: None,
                            scale: None,
                        },
                        TokenType::Numeric => DataType::Numeric {
                            precision: None,
                            scale: None,
                        },
                        TokenType::Varchar => DataType::Varchar(None),
                        TokenType::Char => DataType::Char(None),
                        TokenType::Text => DataType::Text,
                        TokenType::Json => DataType::Json,
                        TokenType::Jsonb => DataType::Jsonb,
                        TokenType::Uuid => DataType::Uuid,
                        TokenType::Bytea => DataType::Bytea,
                        TokenType::Blob => DataType::Blob,
                        _ => DataType::Unknown(name.clone()),
                    };
                    return Ok(Expr::Cast {
                        expr: Box::new(Expr::StringLiteral(value_token.value)),
                        data_type,
                    });
                }

                // PostgreSQL geometric / network / OID type aliases used as
                // typed-literal prefixes (e.g. `box '(1,2,3,4)'`,
                // `point '(1,2)'`, `inet '127.0.0.1'`). Recognize a curated
                // list of bare identifiers followed by a String literal and
                // fold the pair into a Cast(Unknown(name)).
                if name_qs == QuoteStyle::None
                    && self.peek_type() == &TokenType::String
                    && matches!(
                        name.to_ascii_lowercase().as_str(),
                        "box"
                            | "point"
                            | "circle"
                            | "line"
                            | "lseg"
                            | "path"
                            | "polygon"
                            | "inet"
                            | "cidr"
                            | "macaddr"
                            | "macaddr8"
                            | "money"
                            | "regclass"
                            | "regtype"
                            | "regproc"
                            | "regprocedure"
                            | "regrole"
                            | "regnamespace"
                            | "regoperator"
                            | "regoper"
                            | "oid"
                            | "xml"
                            | "tsvector"
                            | "tsquery"
                            | "jsonpath"
                            | "name"
                            | "bit"
                            | "varbit"
                            | "interval"
                            | "bool"
                            | "int2"
                            | "int4"
                            | "int8"
                            | "float4"
                            | "float8"
                    )
                {
                    let value_token = self.advance().clone();
                    return Ok(Expr::Cast {
                        expr: Box::new(Expr::StringLiteral(value_token.value)),
                        data_type: DataType::Unknown(name.clone()),
                    });
                }

                // ── Bare niladic temporal keywords: CURRENT_TIME, CURRENT_DATE,
                //    CURRENT_TIMESTAMP, LOCALTIMESTAMP (no parens) ──
                // ANSI SQL allows these without parentheses. Materialize them
                // as typed functions so the generator can emit dialect-specific
                // forms (e.g. TSQL requires CAST(GETDATE() AS TIME) rather than
                // a bare CURRENT_TIME reserved word).
                if name_qs == QuoteStyle::None && self.peek_type() != &TokenType::LParen {
                    let upper = name.to_ascii_uppercase();
                    let typed = match upper.as_str() {
                        "CURRENT_DATE" => Some(TypedFunction::CurrentDate),
                        "CURRENT_TIME" => Some(TypedFunction::CurrentTime),
                        "CURRENT_TIMESTAMP" | "LOCALTIMESTAMP" => {
                            Some(TypedFunction::CurrentTimestamp)
                        }
                        _ => None,
                    };
                    if let Some(tf) = typed {
                        return Ok(Expr::TypedFunction {
                            func: tf,
                            filter: None,
                            over: None,
                        });
                    }
                }

                // Function call: name(...)
                if self.peek_type() == &TokenType::LParen {
                    self.advance();

                    // TRY_CAST / SAFE_CAST / TRY_TO_TIMESTAMP / … — same shape
                    // as `CAST(expr AS type)`. Lower to `Expr::Cast` when the
                    // body matches; fall back to ordinary function call when
                    // it does not (e.g. comma-separated args).
                    if matches!(name.to_ascii_uppercase().as_str(), "TRY_CAST" | "SAFE_CAST") {
                        let save = self.pos;
                        let inner = self.parse_expr()?;
                        if self.match_token(TokenType::As) {
                            let dt = self.parse_data_type()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Cast {
                                expr: Box::new(inner),
                                data_type: dt,
                            });
                        }
                        self.pos = save;
                    }

                    // Special: COUNT(*), COUNT(DISTINCT x)
                    let distinct = self.match_token(TokenType::Distinct);
                    // ANSI / ClickHouse `agg(ALL …)` — `ALL` is the opposite
                    // of DISTINCT and the default. Swallow so the args parse.
                    if !distinct {
                        let _ = self.match_token(TokenType::All);
                    }

                    // Standard SQL syntactic forms for string functions:
                    //   SUBSTRING(expr FROM start [FOR len])
                    //   SUBSTRING(expr FOR len)
                    //   TRIM([LEADING|TRAILING|BOTH] [chars] FROM expr)
                    //   POSITION(needle IN haystack)
                    //   OVERLAY(expr PLACING str FROM start [FOR len])
                    let upper_name = name.to_ascii_uppercase();
                    if !distinct && self.peek_type() != &TokenType::RParen {
                        match upper_name.as_str() {
                            "SUBSTRING" | "SUBSTR" => {
                                let saved = self.pos;
                                let first = self.parse_expr()?;
                                if self.match_token(TokenType::From) {
                                    let start = self.parse_expr()?;
                                    let length = if self.check_keyword("FOR") {
                                        self.advance();
                                        Some(self.parse_expr()?)
                                    } else {
                                        None
                                    };
                                    self.expect(TokenType::RParen)?;
                                    let mut a = vec![first, start];
                                    if let Some(l) = length {
                                        a.push(l);
                                    }
                                    return Ok(Expr::Function {
                                        name: name.clone(),
                                        args: a,
                                        distinct: false,
                                        filter: None,
                                        over: None,
                                        order_by: Vec::new(),
                                        within_group: false,
                                    });
                                } else if self.check_keyword("FOR") {
                                    self.advance();
                                    let len = self.parse_expr()?;
                                    self.expect(TokenType::RParen)?;
                                    return Ok(Expr::Function {
                                        name: name.clone(),
                                        args: vec![first, len],
                                        distinct: false,
                                        filter: None,
                                        over: None,
                                        order_by: Vec::new(),
                                        within_group: false,
                                    });
                                }
                                self.pos = saved;
                            }
                            "TRIM" => {
                                let saved = self.pos;
                                if self.check_keyword("LEADING")
                                    || self.check_keyword("TRAILING")
                                    || self.check_keyword("BOTH")
                                {
                                    self.advance();
                                }
                                if self.peek_type() == &TokenType::From {
                                    self.advance();
                                    let expr = self.parse_expr()?;
                                    self.expect(TokenType::RParen)?;
                                    return Ok(Expr::Function {
                                        name: name.clone(),
                                        args: vec![expr],
                                        distinct: false,
                                        filter: None,
                                        over: None,
                                        order_by: Vec::new(),
                                        within_group: false,
                                    });
                                }
                                let chars = self.parse_expr()?;
                                if self.match_token(TokenType::From) {
                                    let expr = self.parse_expr()?;
                                    self.expect(TokenType::RParen)?;
                                    return Ok(Expr::Function {
                                        name: name.clone(),
                                        args: vec![expr, chars],
                                        distinct: false,
                                        filter: None,
                                        over: None,
                                        order_by: Vec::new(),
                                        within_group: false,
                                    });
                                }
                                self.pos = saved;
                            }
                            "POSITION" => {
                                let saved = self.pos;
                                let needle = self.parse_expr()?;
                                if self.match_token(TokenType::In) {
                                    let haystack = self.parse_expr()?;
                                    self.expect(TokenType::RParen)?;
                                    return Ok(Expr::Function {
                                        name: name.clone(),
                                        args: vec![needle, haystack],
                                        distinct: false,
                                        filter: None,
                                        over: None,
                                        order_by: Vec::new(),
                                        within_group: false,
                                    });
                                }
                                self.pos = saved;
                            }
                            "OVERLAY" => {
                                let saved = self.pos;
                                let target = self.parse_expr()?;
                                if self.check_keyword("PLACING") {
                                    self.advance();
                                    let placing = self.parse_expr()?;
                                    if self.match_token(TokenType::From) {
                                        let from = self.parse_expr()?;
                                        let len = if self.check_keyword("FOR") {
                                            self.advance();
                                            Some(self.parse_expr()?)
                                        } else {
                                            None
                                        };
                                        self.expect(TokenType::RParen)?;
                                        let mut a = vec![target, placing, from];
                                        if let Some(l) = len {
                                            a.push(l);
                                        }
                                        return Ok(Expr::Function {
                                            name: name.clone(),
                                            args: a,
                                            distinct: false,
                                            filter: None,
                                            over: None,
                                            order_by: Vec::new(),
                                            within_group: false,
                                        });
                                    }
                                }
                                self.pos = saved;
                            }
                            _ => {}
                        }
                    }

                    // MySQL's GROUP_CONCAT has bespoke grammar
                    // (ORDER BY ..., SEPARATOR ...) — parse it into a typed
                    // expression so the structure is preserved across dialects.
                    if name.eq_ignore_ascii_case("GROUP_CONCAT") {
                        let expr = self.parse_group_concat_call(distinct)?;
                        self.expect(TokenType::RParen)?;
                        return Ok(expr);
                    }

                    let args = if self.peek_type() == &TokenType::RParen {
                        vec![]
                    } else if self.peek_type() == &TokenType::Star {
                        self.advance();
                        vec![Expr::Wildcard]
                    } else {
                        let mut a = vec![self.parse_function_arg()?];
                        while self.match_token(TokenType::Comma) {
                            a.push(self.parse_function_arg()?);
                        }
                        a
                    };

                    // Optional aggregate ORDER BY inside arg list (Postgres / Spark):
                    //   array_agg(x ORDER BY y DESC)
                    //   string_agg(x, ',' ORDER BY y)
                    let mut agg_order_by: Vec<OrderByItem> = vec![];
                    if self.peek_type() == &TokenType::Order {
                        self.advance();
                        self.expect(TokenType::By)?;
                        agg_order_by = self.parse_order_by_items()?;
                    }
                    // BigQuery / Snowflake: `ARRAY_AGG(x [ORDER BY y] LIMIT n)`.
                    // Swallow the trailing LIMIT clause inside the function call.
                    if self.peek_type() == &TokenType::Limit {
                        self.advance();
                        let _ = self.parse_expr();
                    }
                    // DuckDB aggregate-state modifier:
                    //   `count(1) EXPORT_STATE` returns the aggregate state
                    //   rather than its final value. We don't model it.
                    if self.check_keyword("EXPORT_STATE") {
                        self.advance();
                    }
                    self.expect(TokenType::RParen)?;

                    // Optional WITHIN GROUP (ORDER BY ...) — ordered-set aggregates
                    //   percentile_cont(0.5) WITHIN GROUP (ORDER BY x)
                    //   listagg(x, ',') WITHIN GROUP (ORDER BY x)
                    let mut within_group = false;
                    let mut wg_order_by: Vec<OrderByItem> = vec![];
                    if self.check_keyword("WITHIN") {
                        self.advance();
                        self.expect_keyword("GROUP")?;
                        self.expect(TokenType::LParen)?;
                        self.expect(TokenType::Order)?;
                        self.expect(TokenType::By)?;
                        wg_order_by = self.parse_order_by_items()?;
                        self.expect(TokenType::RParen)?;
                        within_group = true;
                    }

                    let final_order_by = if within_group {
                        wg_order_by
                    } else {
                        agg_order_by
                    };

                    // Try to construct a typed function variant only when there are no
                    // aggregate-specific clauses (otherwise we lose them).
                    if final_order_by.is_empty()
                        && !within_group
                        && let Some(typed) = Self::try_typed_function(&name, args.clone(), distinct)
                    {
                        return Ok(typed);
                    }

                    Ok(Expr::Function {
                        name,
                        args,
                        distinct,
                        filter: None,
                        over: None,
                        order_by: final_order_by,
                        within_group,
                    })
                }
                // Qualified column: table.column or table.*
                else if self.match_token(TokenType::Dot) {
                    if self.peek_type() == &TokenType::Star {
                        self.advance();
                        Ok(Expr::QualifiedWildcard { table: name })
                    } else {
                        // ClickHouse JSON subobject and typed access at the
                        // first dot: `json.^a`, `json.:Int64`.
                        let _ = self.match_token(TokenType::BitwiseXor);
                        let _ = self.match_token(TokenType::Colon);
                        let (mut col, mut col_qs) = if matches!(self.peek_type(), TokenType::Number)
                        {
                            // ClickHouse tuple index `x.1`.
                            let v = self.peek().value.clone();
                            self.advance();
                            (v, QuoteStyle::None)
                        } else if matches!(self.peek_type(), TokenType::Null) {
                            // ClickHouse JSON subcolumn `.null` (e.g.
                            // `arr.null`, `t.s.null`). Accept the keyword as
                            // a field name in dotted-access position.
                            let v = self.peek().value.clone();
                            self.advance();
                            (v, QuoteStyle::None)
                        } else {
                            self.expect_name_with_quote()?
                        };
                        // Handle 3+ part qualified names like `db.schema.table.column`
                        // (DuckDB, ClickHouse). We collapse everything except the
                        // final segment into the `table` field as a dotted string.
                        let mut table = name;
                        let mut table_qs = name_qs;
                        while self.match_token(TokenType::Dot) {
                            if self.peek_type() == &TokenType::Star {
                                self.advance();
                                let mut full = table;
                                full.push('.');
                                full.push_str(&col);
                                return Ok(Expr::QualifiedWildcard { table: full });
                            }
                            // ClickHouse JSON subobject (`json.^a`) and typed
                            // access (`json.a.:Int64`) — swallow the operator
                            // so the following name can be consumed normally.
                            let _ = self.match_token(TokenType::BitwiseXor);
                            let _ = self.match_token(TokenType::Colon);
                            // ClickHouse tuple index (`t.1`): treat number as
                            // a synthetic field name.
                            let (next_col, next_qs) =
                                if matches!(self.peek_type(), TokenType::Number) {
                                    let v = self.peek().value.clone();
                                    self.advance();
                                    (v, QuoteStyle::None)
                                } else if matches!(self.peek_type(), TokenType::Null) {
                                    let v = self.peek().value.clone();
                                    self.advance();
                                    (v, QuoteStyle::None)
                                } else {
                                    self.expect_name_with_quote()?
                                };
                            table.push('.');
                            table.push_str(&col);
                            table_qs = col_qs;
                            col = next_col;
                            col_qs = next_qs;
                        }
                        // Function call on dotted name: db.schema.func(args).
                        if self.peek_type() == &TokenType::LParen {
                            self.advance();
                            let mut full = table;
                            full.push('.');
                            full.push_str(&col);
                            let args = if self.peek_type() != &TokenType::RParen {
                                let mut a = vec![self.parse_function_arg()?];
                                while self.match_token(TokenType::Comma) {
                                    a.push(self.parse_function_arg()?);
                                }
                                a
                            } else {
                                vec![]
                            };
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Function {
                                name: full,
                                args,
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        Ok(Expr::Column {
                            table: Some(table),
                            name: col,
                            quote_style: col_qs,
                            table_quote_style: table_qs,
                        })
                    }
                } else {
                    Ok(Expr::Column {
                        table: None,
                        name,
                        quote_style: name_qs,
                        table_quote_style: QuoteStyle::None,
                    })
                }
            }

            _ => {
                // Fallback: any other token whose value is a valid identifier
                // and is immediately followed by `(` is treated as a function
                // call. This handles reserved keywords used as Spark/Hive
                // built-ins (IF, ALL, ANY, EXISTS, MOD, etc.) and dialect
                // functions that happen to collide with token types.
                let v = token.value.clone();
                let is_word =
                    !v.is_empty() && v.chars().all(|c| c.is_ascii_alphanumeric() || c == '_');
                if is_word
                    && matches!(
                        self.peek_offset(1).map(|t| &t.token_type),
                        Some(TokenType::LParen)
                    )
                {
                    // TRY_CAST / SAFE_CAST / TRY_TO_TIMESTAMP / … — same
                    // shape as `CAST(expr AS type)`. Lower to `Expr::Cast`
                    // (or back to a function call when the form doesn't
                    // match).
                    let upper = v.to_ascii_uppercase();
                    if matches!(upper.as_str(), "TRY_CAST" | "SAFE_CAST") {
                        self.advance();
                        self.advance(); // consume '('
                        let inner = self.parse_expr()?;
                        if self.match_token(TokenType::As) {
                            let data_type = self.parse_data_type()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Cast {
                                expr: Box::new(inner),
                                data_type,
                            });
                        }
                        // Fall back: treat as ordinary function call.
                        let mut args = vec![inner];
                        while self.match_token(TokenType::Comma) {
                            args.push(self.parse_expr()?);
                        }
                        self.expect(TokenType::RParen)?;
                        return Ok(Expr::Function {
                            name: v,
                            args,
                            distinct: false,
                            filter: None,
                            over: None,
                            order_by: Vec::new(),
                            within_group: false,
                        });
                    }
                    self.advance();
                    self.advance(); // consume '('
                    let upper = v.to_ascii_uppercase();
                    // Standard SQL `SUBSTRING(expr FROM start [FOR length])`
                    // and MySQL `SUBSTRING(expr FROM start)` / `…FOR length`.
                    if matches!(upper.as_str(), "SUBSTRING" | "SUBSTR")
                        && self.peek_type() != &TokenType::RParen
                    {
                        let saved = self.pos;
                        let first = self.parse_expr()?;
                        if self.match_token(TokenType::From) {
                            let start = self.parse_expr()?;
                            let length = if self.check_keyword("FOR") {
                                self.advance();
                                Some(self.parse_expr()?)
                            } else {
                                None
                            };
                            self.expect(TokenType::RParen)?;
                            let mut args = vec![first, start];
                            if let Some(len) = length {
                                args.push(len);
                            }
                            return Ok(Expr::Function {
                                name: v,
                                args,
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        if self.check_keyword("FOR") {
                            self.advance();
                            let length = self.parse_expr()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Function {
                                name: v,
                                args: vec![first, length],
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        // Fall back: re-parse as comma list.
                        self.pos = saved;
                    }
                    // Standard `TRIM([LEADING|TRAILING|BOTH] [chars] FROM expr)`
                    // and `TRIM(expr [, chars])` (already covered by comma).
                    if upper == "TRIM" && self.peek_type() != &TokenType::RParen {
                        let saved = self.pos;
                        if self.check_keyword("LEADING")
                            || self.check_keyword("TRAILING")
                            || self.check_keyword("BOTH")
                        {
                            self.advance();
                        }
                        if self.peek_type() == &TokenType::From {
                            self.advance();
                            let expr = self.parse_expr()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Function {
                                name: v,
                                args: vec![expr],
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        // chars FROM expr
                        let chars = self.parse_expr()?;
                        if self.match_token(TokenType::From) {
                            let expr = self.parse_expr()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Function {
                                name: v,
                                args: vec![expr, chars],
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        // Plain comma list — fall back.
                        self.pos = saved;
                    }
                    // Standard `OVERLAY(expr PLACING str FROM start [FOR len])`.
                    if upper == "OVERLAY" && self.peek_type() != &TokenType::RParen {
                        let saved = self.pos;
                        let target = self.parse_expr()?;
                        if self.check_keyword("PLACING") {
                            self.advance();
                            let placing = self.parse_expr()?;
                            self.expect(TokenType::From)?;
                            let from = self.parse_expr()?;
                            let len = if self.check_keyword("FOR") {
                                self.advance();
                                Some(self.parse_expr()?)
                            } else {
                                None
                            };
                            self.expect(TokenType::RParen)?;
                            let mut args = vec![target, placing, from];
                            if let Some(l) = len {
                                args.push(l);
                            }
                            return Ok(Expr::Function {
                                name: v,
                                args,
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        self.pos = saved;
                    }
                    // Standard `POSITION(needle IN haystack)`.
                    if upper == "POSITION" && self.peek_type() != &TokenType::RParen {
                        let saved = self.pos;
                        let needle = self.parse_expr()?;
                        if self.check_keyword("IN") {
                            self.advance();
                            let haystack = self.parse_expr()?;
                            self.expect(TokenType::RParen)?;
                            return Ok(Expr::Function {
                                name: v,
                                args: vec![needle, haystack],
                                distinct: false,
                                filter: None,
                                over: None,
                                order_by: Vec::new(),
                                within_group: false,
                            });
                        }
                        self.pos = saved;
                    }
                    let mut args = Vec::new();
                    if self.peek_type() != &TokenType::RParen {
                        args.push(self.parse_function_arg()?);
                        while self.match_token(TokenType::Comma) {
                            args.push(self.parse_function_arg()?);
                        }
                    }
                    self.expect(TokenType::RParen)?;
                    return Ok(Expr::Function {
                        name: v,
                        args,
                        distinct: false,
                        filter: None,
                        over: None,
                        order_by: Vec::new(),
                        within_group: false,
                    });
                }
                Err(SqlglotError::UnexpectedToken { token })
            }
        }
    }

    /// Parse a single function-call argument. Accepts the DuckDB / PostgreSQL
    /// named-argument syntaxes `name := value` and `name => value` and falls
    /// back to a plain expression for positional arguments. The argument
    /// name is discarded — we don't model it in the AST.
    fn parse_function_arg(&mut self) -> Result<Expr> {
        // Hive table-valued function clause: `noop(on tbl partition by p
        // order by q distribute by r cluster by s sort by t)`. The arg
        // list begins with the `ON` keyword and is followed by a series
        // of windowing-style clauses we don't model. Swallow it as an
        // opaque payload so we don't reject the call.
        if matches!(self.peek_type(), TokenType::On) {
            let mut depth = 0usize;
            while !matches!(self.peek_type(), TokenType::Eof) {
                match self.peek_type() {
                    TokenType::LParen => depth += 1,
                    TokenType::RParen => {
                        if depth == 0 {
                            break;
                        }
                        depth -= 1;
                    }
                    TokenType::Comma if depth == 0 => break,
                    _ => {}
                }
                self.advance();
            }
            return Ok(Expr::Null);
        }
        if self.is_name_token()
            || self.is_data_type_token()
            || matches!(self.peek_type(), TokenType::Recursive)
        {
            let next = self.peek_offset(1).map(|t| &t.token_type);
            if matches!(next, Some(TokenType::Colon)) {
                let after = self.peek_offset(2).map(|t| &t.token_type);
                if matches!(after, Some(TokenType::Eq)) {
                    self.advance();
                    self.advance();
                    self.advance();
                    return self.parse_expr();
                }
            }
            if matches!(next, Some(TokenType::DoubleArrow)) {
                self.advance();
                self.advance();
                return self.parse_expr();
            }
        }
        // ClickHouse table functions: `view(SELECT …)`, `cluster(…)` etc.
        // accept a full SELECT / WITH / UNION inside the arg list. Parse
        // it as a Subquery so the surrounding call closes properly.
        if matches!(self.peek_type(), TokenType::Select | TokenType::With) {
            let stmt = self.parse_statement_inner()?;
            return Ok(Expr::Subquery(Box::new(stmt)));
        }
        let mut expr = self.parse_expr()?;
        // Oracle / Snowflake / MySQL `JSON_OBJECT('k' : value, ...)` and the
        // `JSON_OBJECTAGG(k : v)` family use `:` as a key-value separator
        // inside function args. After parsing the first expression, swallow
        // a bare `:` and parse the value side; emit the value as the arg
        // (we don't model JSON key-value pairs in the AST). Only fire when
        // the next-after-colon is not another `:` (`::` cast) and not `=`
        // (`:=` named arg, already handled above).
        if matches!(self.peek_type(), TokenType::Colon)
            && !matches!(
                self.peek_offset(1).map(|t| &t.token_type),
                Some(TokenType::Colon) | Some(TokenType::Eq)
            )
        {
            self.advance(); // :
            expr = self.parse_expr()?;
            // Optional `FORMAT JSON` suffix (Oracle).
            if self.peek().value.eq_ignore_ascii_case("FORMAT")
                && self
                    .peek_offset(1)
                    .map(|t| t.value.eq_ignore_ascii_case("JSON"))
                    .unwrap_or(false)
            {
                self.advance();
                self.advance();
            }
        }
        // ClickHouse: `func(expr AS alias)` — swallow the alias.
        if self.match_token(TokenType::As) && self.is_name_token() {
            self.advance();
        }
        // Spark / DataBricks UDTF call: `UDTF(TABLE(t) [PARTITION BY cols]
        // [ORDER BY cols])`. Swallow the table-argument modifiers opaquely.
        if self.peek_type() == &TokenType::Partition
            && self
                .peek_offset(1)
                .map(|t| matches!(t.token_type, TokenType::By))
                .unwrap_or(false)
        {
            self.advance(); // PARTITION
            self.advance(); // BY
            // Comma-separated expression list (column refs / exprs).
            let _ = self.parse_expr()?;
            while self.match_token(TokenType::Comma) {
                let _ = self.parse_expr()?;
            }
        }
        if self.peek_type() == &TokenType::Order
            && self
                .peek_offset(1)
                .map(|t| matches!(t.token_type, TokenType::By))
                .unwrap_or(false)
        {
            self.advance(); // ORDER
            self.advance(); // BY
            let _ = self.parse_order_by_items()?;
        }
        // BigQuery / DuckDB / Snowflake / Oracle window-function nulls
        // modifier: `LAST_VALUE(arg IGNORE NULLS)`, `... RESPECT NULLS`.
        // Swallow opaquely; we don't model it in the AST.
        if (self.peek().value.eq_ignore_ascii_case("IGNORE")
            || self.peek().value.eq_ignore_ascii_case("RESPECT"))
            && self
                .peek_offset(1)
                .map(|t| t.token_type == TokenType::Null || t.value.eq_ignore_ascii_case("NULLS"))
                .unwrap_or(false)
        {
            self.advance();
            self.advance();
        }
        // Postgres JSON helpers: `JSON_SERIALIZE(expr RETURNING type)`,
        // `JSON_QUERY(... RETURNING jsonb FORMAT JSON)`,
        // `JSON_VALUE(... RETURNING type DEFAULT v ON EMPTY|ERROR …)`. After
        // any RETURNING clause, swallow the optional FORMAT, DEFAULT, ON
        // EMPTY/ERROR tail so the call parses cleanly.
        if self.match_token(TokenType::Returning) {
            if self.is_data_type_token() || self.is_name_token() {
                let _ = self.parse_data_type();
            }
        }
        // SQL/JSON `PASSING v AS name [, v AS name]*` clause inside
        // JSON_EXISTS / JSON_VALUE / JSON_QUERY argument lists.
        if self.check_keyword("PASSING") {
            self.advance();
            loop {
                let _ = self.parse_expr()?;
                if self.match_token(TokenType::As) && self.is_name_token() {
                    self.advance();
                }
                if !self.match_token(TokenType::Comma) {
                    break;
                }
            }
        }
        // SQL/JSON behavior clauses: `NULL|ERROR|EMPTY [ARRAY|OBJECT]|
        // DEFAULT expr ON EMPTY|ERROR`. Swallow them opaquely; the
        // surrounding call still resolves to its primary expression.
        loop {
            let is_default = self.peek_type() == &TokenType::Default;
            let is_behavior_kw = self.check_keyword("ERROR")
                || self.check_keyword("NULL")
                || self.peek_type() == &TokenType::Null
                || self.check_keyword("EMPTY")
                || self.check_keyword("TRUE")
                || self.check_keyword("FALSE")
                || self.check_keyword("UNKNOWN");
            if !is_default && !is_behavior_kw {
                break;
            }
            // Look ahead: behavior keyword must be followed (possibly via
            // optional ARRAY/OBJECT/expr) by `ON ERROR|EMPTY` to qualify.
            let saved = self.pos;
            if is_default {
                self.advance();
                let _ = self.parse_expr();
            } else {
                self.advance();
                if self.check_keyword("ARRAY") || self.check_keyword("OBJECT") {
                    self.advance();
                }
            }
            if self.peek_type() == &TokenType::On
                && self
                    .peek_offset(1)
                    .map(|t| {
                        t.value.eq_ignore_ascii_case("ERROR")
                            || t.value.eq_ignore_ascii_case("EMPTY")
                    })
                    .unwrap_or(false)
            {
                self.advance(); // ON
                self.advance(); // ERROR / EMPTY
            } else {
                // Not actually a behavior clause — rewind.
                self.pos = saved;
                break;
            }
        }
        // MySQL `CONVERT(expr USING charset)` — swallow USING + name.
        if self.match_token(TokenType::Using) {
            if self.is_name_token() {
                self.advance();
            }
        }
        // ON EMPTY / ON ERROR / DEFAULT … ON EMPTY|ERROR / FORMAT … —
        // tolerated tail clauses common to JSON_VALUE / JSON_QUERY /
        // JSON_EXISTS. Loop while one of the recognized starters appears.
        loop {
            let starts = self.peek_type() == &TokenType::Default
                || self.match_keyword_clone("FORMAT")
                || (self.peek_type() == &TokenType::On
                    && self
                        .peek_offset(1)
                        .map(|t| {
                            t.value.eq_ignore_ascii_case("EMPTY")
                                || t.value.eq_ignore_ascii_case("ERROR")
                        })
                        .unwrap_or(false));
            if !starts {
                break;
            }
            // Consume up to the next top-level `,` / `)` / EOF, tracking
            // nesting so embedded parens (e.g. `DEFAULT ('C' COLLATE "C")`)
            // don't terminate prematurely.
            let mut depth = 0i32;
            while !matches!(self.peek_type(), TokenType::Eof) {
                match self.peek_type() {
                    TokenType::LParen | TokenType::LBracket => depth += 1,
                    TokenType::RParen | TokenType::RBracket => {
                        if depth == 0 {
                            break;
                        }
                        depth -= 1;
                    }
                    TokenType::Comma if depth == 0 => break,
                    _ => {}
                }
                self.advance();
            }
        }
        Ok(expr)
    }

    /// True when the current token is a name token whose uppercase value
    /// equals `kw`. Does NOT advance the token cursor.
    fn match_keyword_clone(&self, kw: &str) -> bool {
        self.check_keyword(kw)
    }

    fn is_data_type_token(&self) -> bool {
        self.is_data_type_token_kind(self.peek_type())
    }

    fn is_data_type_token_kind(&self, tt: &TokenType) -> bool {
        matches!(
            tt,
            TokenType::Int
                | TokenType::Integer
                | TokenType::BigInt
                | TokenType::SmallInt
                | TokenType::TinyInt
                | TokenType::Float
                | TokenType::Double
                | TokenType::Decimal
                | TokenType::Numeric
                | TokenType::Real
                | TokenType::Varchar
                | TokenType::Char
                | TokenType::Text
                | TokenType::Boolean
                | TokenType::Date
                | TokenType::Timestamp
                | TokenType::TimestampTz
                | TokenType::Time
                | TokenType::Interval
                | TokenType::Blob
                | TokenType::Bytea
                | TokenType::Json
                | TokenType::Jsonb
                | TokenType::Uuid
                | TokenType::Array
                | TokenType::Map
                | TokenType::Struct
        )
    }

    fn parse_datetime_field(&mut self) -> Result<DateTimeField> {
        let token = self.peek().clone();
        let field = match &token.token_type {
            TokenType::Year => DateTimeField::Year,
            TokenType::Month => DateTimeField::Month,
            TokenType::Day => DateTimeField::Day,
            TokenType::Hour => DateTimeField::Hour,
            TokenType::Minute => DateTimeField::Minute,
            TokenType::Second => DateTimeField::Second,
            TokenType::Epoch => DateTimeField::Epoch,
            _ => {
                let name = token.value.to_uppercase();
                match name.as_str() {
                    "YEAR" => DateTimeField::Year,
                    "QUARTER" => DateTimeField::Quarter,
                    "MONTH" => DateTimeField::Month,
                    "WEEK" => DateTimeField::Week,
                    "DAY" => DateTimeField::Day,
                    "DOW" | "DAYOFWEEK" => DateTimeField::DayOfWeek,
                    "DOY" | "DAYOFYEAR" => DateTimeField::DayOfYear,
                    "HOUR" => DateTimeField::Hour,
                    "MINUTE" => DateTimeField::Minute,
                    "SECOND" => DateTimeField::Second,
                    "MILLISECOND" | "MILLISECONDS" | "MS" => DateTimeField::Millisecond,
                    "MICROSECOND" | "MICROSECONDS" | "US" => DateTimeField::Microsecond,
                    "NANOSECOND" | "NANOSECONDS" | "NS" => DateTimeField::Nanosecond,
                    "YEARS" => DateTimeField::Year,
                    "QUARTERS" => DateTimeField::Quarter,
                    "MONTHS" => DateTimeField::Month,
                    "WEEKS" => DateTimeField::Week,
                    "DAYS" => DateTimeField::Day,
                    "HOURS" => DateTimeField::Hour,
                    "MINUTES" => DateTimeField::Minute,
                    "SECONDS" => DateTimeField::Second,
                    "EPOCH" => DateTimeField::Epoch,
                    "TIMEZONE" => DateTimeField::Timezone,
                    "TIMEZONE_HOUR" => DateTimeField::TimezoneHour,
                    "TIMEZONE_MINUTE" => DateTimeField::TimezoneMinute,
                    // MySQL composite interval units. We don't model them
                    // distinctly; lower to the dominant component so the
                    // surrounding parse completes.
                    "DAY_HOUR" | "DAY_MINUTE" | "DAY_SECOND" | "DAY_MICROSECOND" => {
                        DateTimeField::Day
                    }
                    "HOUR_MINUTE" | "HOUR_SECOND" | "HOUR_MICROSECOND" => DateTimeField::Hour,
                    "MINUTE_SECOND" | "MINUTE_MICROSECOND" => DateTimeField::Minute,
                    "SECOND_MICROSECOND" => DateTimeField::Second,
                    "YEAR_MONTH" => DateTimeField::Year,
                    _ => {
                        return Err(SqlglotError::ParserError {
                            message: format!("Unknown datetime field: {name}"),
                        });
                    }
                }
            }
        };
        self.advance();
        Ok(field)
    }

    fn try_parse_datetime_field(&mut self) -> Option<DateTimeField> {
        let saved = self.pos;
        match self.parse_datetime_field() {
            Ok(field) => Some(field),
            Err(_) => {
                self.pos = saved;
                None
            }
        }
    }

    /// Parse the inside of `GROUP_CONCAT(...)` (caller has already consumed
    /// the `(` and optional `DISTINCT`). Returns a typed `GroupConcat`
    /// expression. Does NOT consume the trailing `)`.
    fn parse_group_concat_call(&mut self, distinct: bool) -> Result<Expr> {
        let mut exprs: Vec<Expr> = Vec::new();
        let mut order_by: Vec<OrderByItem> = Vec::new();
        let mut separator: Option<Box<Expr>> = None;

        if self.peek_type() != &TokenType::RParen {
            exprs.push(self.parse_expr()?);
            while self.peek_type() == &TokenType::Comma {
                // ORDER BY / SEPARATOR are alternative terminators, not args.
                // Peek one past the comma to disambiguate `f(a, b)` from
                // `f(a, b ORDER BY ...)` — but comma here always introduces
                // another positional arg, so just keep consuming.
                self.advance();
                exprs.push(self.parse_expr()?);
            }

            if self.match_token(TokenType::Order) {
                self.expect(TokenType::By)?;
                order_by = self.parse_order_by_items()?;
            }

            if self.match_keyword("SEPARATOR") {
                separator = Some(Box::new(self.parse_expr()?));
            }
        }

        Ok(Expr::TypedFunction {
            func: TypedFunction::GroupConcat {
                exprs,
                separator,
                order_by,
                distinct,
            },
            filter: None,
            over: None,
        })
    }

    /// Try to construct a typed function expression from a parsed function call.
    /// Returns `None` if the function name is not recognized, falling back to
    /// the generic `Expr::Function`.
    fn try_typed_function(name: &str, args: Vec<Expr>, distinct: bool) -> Option<Expr> {
        let upper = name.to_uppercase();
        let tf = match upper.as_str() {
            // ── Date/Time ──────────────────────────────────────────
            "DATE_ADD" | "DATEADD" | "TIMESTAMPADD" => {
                let mut it = args.into_iter();
                let first = it.next()?;
                let second = it.next()?;
                let third = it.next();
                // Handle DATEADD(unit, interval, expr) — TSQL/Snowflake arg order
                if upper == "DATEADD" {
                    if let Some(third_arg) = third {
                        // 3-arg: DATEADD(unit, interval, expr)
                        let unit = Self::expr_to_datetime_field(&first);
                        TypedFunction::DateAdd {
                            expr: Box::new(third_arg),
                            interval: Box::new(second),
                            unit,
                        }
                    } else {
                        TypedFunction::DateAdd {
                            expr: Box::new(first),
                            interval: Box::new(second),
                            unit: None,
                        }
                    }
                } else {
                    // DATE_ADD(expr, interval [, unit])
                    let unit = third.as_ref().and_then(Self::expr_to_datetime_field);
                    TypedFunction::DateAdd {
                        expr: Box::new(first),
                        interval: Box::new(second),
                        unit,
                    }
                }
            }
            "DATE_DIFF" | "DATEDIFF" | "TIMESTAMPDIFF" => {
                let mut it = args.into_iter();
                let first = it.next()?;
                let second = it.next()?;
                let third = it.next();
                if let Some(third_arg) = third {
                    if upper == "DATEDIFF" {
                        // DATEDIFF(unit, start, end) — TSQL/Snowflake
                        let unit = Self::expr_to_datetime_field(&first);
                        TypedFunction::DateDiff {
                            start: Box::new(second),
                            end: Box::new(third_arg),
                            unit,
                        }
                    } else {
                        let unit = Self::expr_to_datetime_field(&third_arg);
                        TypedFunction::DateDiff {
                            start: Box::new(first),
                            end: Box::new(second),
                            unit,
                        }
                    }
                } else {
                    TypedFunction::DateDiff {
                        start: Box::new(first),
                        end: Box::new(second),
                        unit: None,
                    }
                }
            }
            "DATE_TRUNC" | "DATETRUNC" => {
                let mut it = args.into_iter();
                let first = it.next()?;
                let second = it.next()?;
                // DATE_TRUNC('unit', expr) or DATE_TRUNC(unit, expr)
                let (unit, expr) = if let Some(u) = Self::expr_to_datetime_field(&first) {
                    (u, second)
                } else if let Some(u) = Self::expr_to_datetime_field(&second) {
                    (u, first)
                } else {
                    // Default: first = unit string, second = expr
                    return None;
                };
                TypedFunction::DateTrunc {
                    unit,
                    expr: Box::new(expr),
                }
            }
            "DATE_SUB" | "DATESUB" => {
                let mut it = args.into_iter();
                let first = it.next()?;
                let second = it.next()?;
                let third = it.next();
                let unit = third.as_ref().and_then(Self::expr_to_datetime_field);
                TypedFunction::DateSub {
                    expr: Box::new(first),
                    interval: Box::new(second),
                    unit,
                }
            }
            "CURRENT_DATE" => TypedFunction::CurrentDate,
            "CURRENT_TIME" | "CURTIME" => TypedFunction::CurrentTime,
            "CURRENT_TIMESTAMP" | "NOW" | "GETDATE" | "SYSDATE" => TypedFunction::CurrentTimestamp,
            "STR_TO_TIME" | "STR_TO_DATE" | "TO_TIMESTAMP" | "PARSE_TIMESTAMP"
            | "PARSE_DATETIME" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let format = it.next()?;
                TypedFunction::StrToTime {
                    expr: Box::new(expr),
                    format: Box::new(format),
                }
            }
            "TIME_TO_STR" | "DATE_FORMAT" | "FORMAT_TIMESTAMP" | "FORMAT_DATETIME" | "TO_CHAR" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let format = it.next()?;
                TypedFunction::TimeToStr {
                    expr: Box::new(expr),
                    format: Box::new(format),
                }
            }
            "TS_OR_DS_TO_DATE" => {
                let mut it = args.into_iter();
                TypedFunction::TsOrDsToDate {
                    expr: Box::new(it.next()?),
                }
            }
            "YEAR" => {
                let mut it = args.into_iter();
                TypedFunction::Year {
                    expr: Box::new(it.next()?),
                }
            }
            "MONTH" => {
                let mut it = args.into_iter();
                TypedFunction::Month {
                    expr: Box::new(it.next()?),
                }
            }
            "DAY" | "DAYOFMONTH" => {
                let mut it = args.into_iter();
                TypedFunction::Day {
                    expr: Box::new(it.next()?),
                }
            }

            // ── String ─────────────────────────────────────────────
            "TRIM" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                TypedFunction::Trim {
                    expr: Box::new(expr),
                    trim_type: TrimType::Both,
                    trim_chars: None,
                }
            }
            "LTRIM" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                TypedFunction::Trim {
                    expr: Box::new(expr),
                    trim_type: TrimType::Leading,
                    trim_chars: None,
                }
            }
            "RTRIM" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                TypedFunction::Trim {
                    expr: Box::new(expr),
                    trim_type: TrimType::Trailing,
                    trim_chars: None,
                }
            }
            "SUBSTRING" | "SUBSTR" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let start = it.next()?;
                let length = it.next();
                TypedFunction::Substring {
                    expr: Box::new(expr),
                    start: Box::new(start),
                    length: length.map(Box::new),
                }
            }
            "UPPER" | "UCASE" => {
                let mut it = args.into_iter();
                TypedFunction::Upper {
                    expr: Box::new(it.next()?),
                }
            }
            "LOWER" | "LCASE" => {
                let mut it = args.into_iter();
                TypedFunction::Lower {
                    expr: Box::new(it.next()?),
                }
            }
            "REGEXP_LIKE" | "RLIKE" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let pattern = it.next()?;
                let flags = it.next();
                TypedFunction::RegexpLike {
                    expr: Box::new(expr),
                    pattern: Box::new(pattern),
                    flags: flags.map(Box::new),
                }
            }
            "REGEXP_EXTRACT" | "REGEXP_SUBSTR" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let pattern = it.next()?;
                let group_index = it.next();
                TypedFunction::RegexpExtract {
                    expr: Box::new(expr),
                    pattern: Box::new(pattern),
                    group_index: group_index.map(Box::new),
                }
            }
            "REGEXP_REPLACE" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let pattern = it.next()?;
                let replacement = it.next()?;
                let flags = it.next();
                TypedFunction::RegexpReplace {
                    expr: Box::new(expr),
                    pattern: Box::new(pattern),
                    replacement: Box::new(replacement),
                    flags: flags.map(Box::new),
                }
            }
            "CONCAT_WS" => {
                let mut it = args.into_iter();
                let separator = it.next()?;
                let exprs: Vec<Expr> = it.collect();
                TypedFunction::ConcatWs {
                    separator: Box::new(separator),
                    exprs,
                }
            }
            "SPLIT" | "STRING_SPLIT" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let delimiter = it.next()?;
                TypedFunction::Split {
                    expr: Box::new(expr),
                    delimiter: Box::new(delimiter),
                }
            }
            "INITCAP" => {
                let mut it = args.into_iter();
                TypedFunction::Initcap {
                    expr: Box::new(it.next()?),
                }
            }
            "LENGTH" | "LEN" | "CHAR_LENGTH" | "CHARACTER_LENGTH" => {
                let mut it = args.into_iter();
                TypedFunction::Length {
                    expr: Box::new(it.next()?),
                }
            }
            "REPLACE" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let from = it.next()?;
                let to = it.next()?;
                TypedFunction::Replace {
                    expr: Box::new(expr),
                    from: Box::new(from),
                    to: Box::new(to),
                }
            }
            "REVERSE" => {
                let mut it = args.into_iter();
                TypedFunction::Reverse {
                    expr: Box::new(it.next()?),
                }
            }
            "LEFT" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let n = it.next()?;
                TypedFunction::Left {
                    expr: Box::new(expr),
                    n: Box::new(n),
                }
            }
            "RIGHT" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let n = it.next()?;
                TypedFunction::Right {
                    expr: Box::new(expr),
                    n: Box::new(n),
                }
            }
            "LPAD" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let length = it.next()?;
                let pad = it.next();
                TypedFunction::Lpad {
                    expr: Box::new(expr),
                    length: Box::new(length),
                    pad: pad.map(Box::new),
                }
            }
            "RPAD" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let length = it.next()?;
                let pad = it.next();
                TypedFunction::Rpad {
                    expr: Box::new(expr),
                    length: Box::new(length),
                    pad: pad.map(Box::new),
                }
            }

            // ── Aggregate ──────────────────────────────────────────
            "COUNT" => {
                let mut it = args.into_iter();
                let expr = it.next().unwrap_or(Expr::Wildcard);
                TypedFunction::Count {
                    expr: Box::new(expr),
                    distinct,
                }
            }
            "SUM" => {
                let mut it = args.into_iter();
                TypedFunction::Sum {
                    expr: Box::new(it.next()?),
                    distinct,
                }
            }
            "AVG" => {
                let mut it = args.into_iter();
                TypedFunction::Avg {
                    expr: Box::new(it.next()?),
                    distinct,
                }
            }
            "MIN" => {
                let mut it = args.into_iter();
                TypedFunction::Min {
                    expr: Box::new(it.next()?),
                }
            }
            "MAX" => {
                let mut it = args.into_iter();
                TypedFunction::Max {
                    expr: Box::new(it.next()?),
                }
            }
            "ARRAY_AGG" | "LIST" | "COLLECT_LIST" => {
                let mut it = args.into_iter();
                TypedFunction::ArrayAgg {
                    expr: Box::new(it.next()?),
                    distinct,
                }
            }
            "APPROX_DISTINCT" | "APPROX_COUNT_DISTINCT" => {
                let mut it = args.into_iter();
                TypedFunction::ApproxDistinct {
                    expr: Box::new(it.next()?),
                }
            }
            "VARIANCE" | "VAR_SAMP" | "VAR" => {
                let mut it = args.into_iter();
                TypedFunction::Variance {
                    expr: Box::new(it.next()?),
                }
            }
            "VAR_POP" => {
                let mut it = args.into_iter();
                TypedFunction::VariancePop {
                    expr: Box::new(it.next()?),
                }
            }
            "STDDEV" | "STDDEV_SAMP" => {
                let mut it = args.into_iter();
                TypedFunction::Stddev {
                    expr: Box::new(it.next()?),
                }
            }
            "STDDEV_POP" => {
                let mut it = args.into_iter();
                TypedFunction::StddevPop {
                    expr: Box::new(it.next()?),
                }
            }

            // ── Array ──────────────────────────────────────────────
            "ARRAY_CONCAT" | "ARRAY_CAT" => TypedFunction::ArrayConcat { arrays: args },
            "ARRAY_CONTAINS" => {
                let mut it = args.into_iter();
                let array = it.next()?;
                let element = it.next()?;
                TypedFunction::ArrayContains {
                    array: Box::new(array),
                    element: Box::new(element),
                }
            }
            "ARRAY_SIZE" | "ARRAY_LENGTH" | "CARDINALITY" => {
                let mut it = args.into_iter();
                TypedFunction::ArraySize {
                    expr: Box::new(it.next()?),
                }
            }
            "EXPLODE" => {
                let mut it = args.into_iter();
                TypedFunction::Explode {
                    expr: Box::new(it.next()?),
                }
            }
            "GENERATE_SERIES" | "SEQUENCE" => {
                let mut it = args.into_iter();
                let start = it.next()?;
                let stop = it.next()?;
                let step = it.next();
                TypedFunction::GenerateSeries {
                    start: Box::new(start),
                    stop: Box::new(stop),
                    step: step.map(Box::new),
                }
            }
            "FLATTEN" => {
                let mut it = args.into_iter();
                TypedFunction::Flatten {
                    expr: Box::new(it.next()?),
                }
            }

            // ── JSON ───────────────────────────────────────────────
            "JSON_EXTRACT" | "JSON_VALUE" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let path = it.next()?;
                TypedFunction::JSONExtract {
                    expr: Box::new(expr),
                    path: Box::new(path),
                }
            }
            "JSON_EXTRACT_SCALAR" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let path = it.next()?;
                TypedFunction::JSONExtractScalar {
                    expr: Box::new(expr),
                    path: Box::new(path),
                }
            }
            "PARSE_JSON" | "JSON_PARSE" => {
                let mut it = args.into_iter();
                TypedFunction::ParseJSON {
                    expr: Box::new(it.next()?),
                }
            }
            "JSON_FORMAT" | "TO_JSON" | "TO_JSON_STRING" => {
                let mut it = args.into_iter();
                TypedFunction::JSONFormat {
                    expr: Box::new(it.next()?),
                }
            }

            // ── Window ─────────────────────────────────────────────
            "ROW_NUMBER" => TypedFunction::RowNumber,
            "RANK" => TypedFunction::Rank,
            "DENSE_RANK" => TypedFunction::DenseRank,
            "NTILE" => {
                let mut it = args.into_iter();
                TypedFunction::NTile {
                    n: Box::new(it.next()?),
                }
            }
            "LEAD" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let offset = it.next();
                let default = it.next();
                TypedFunction::Lead {
                    expr: Box::new(expr),
                    offset: offset.map(Box::new),
                    default: default.map(Box::new),
                }
            }
            "LAG" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let offset = it.next();
                let default = it.next();
                TypedFunction::Lag {
                    expr: Box::new(expr),
                    offset: offset.map(Box::new),
                    default: default.map(Box::new),
                }
            }
            "FIRST_VALUE" => {
                let mut it = args.into_iter();
                TypedFunction::FirstValue {
                    expr: Box::new(it.next()?),
                }
            }
            "LAST_VALUE" => {
                let mut it = args.into_iter();
                TypedFunction::LastValue {
                    expr: Box::new(it.next()?),
                }
            }

            // ── Math ───────────────────────────────────────────────
            "ABS" => {
                let mut it = args.into_iter();
                TypedFunction::Abs {
                    expr: Box::new(it.next()?),
                }
            }
            "CEIL" | "CEILING" => {
                let mut it = args.into_iter();
                TypedFunction::Ceil {
                    expr: Box::new(it.next()?),
                }
            }
            "FLOOR" => {
                let mut it = args.into_iter();
                TypedFunction::Floor {
                    expr: Box::new(it.next()?),
                }
            }
            "ROUND" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let decimals = it.next();
                TypedFunction::Round {
                    expr: Box::new(expr),
                    decimals: decimals.map(Box::new),
                }
            }
            "LOG" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let base = it.next();
                TypedFunction::Log {
                    expr: Box::new(expr),
                    base: base.map(Box::new),
                }
            }
            "LN" => {
                let mut it = args.into_iter();
                TypedFunction::Ln {
                    expr: Box::new(it.next()?),
                }
            }
            "POW" | "POWER" => {
                let mut it = args.into_iter();
                let base = it.next()?;
                let exponent = it.next()?;
                TypedFunction::Pow {
                    base: Box::new(base),
                    exponent: Box::new(exponent),
                }
            }
            "SQRT" => {
                let mut it = args.into_iter();
                TypedFunction::Sqrt {
                    expr: Box::new(it.next()?),
                }
            }
            "GREATEST" => TypedFunction::Greatest { exprs: args },
            "LEAST" => TypedFunction::Least { exprs: args },
            "MOD" => {
                let mut it = args.into_iter();
                let left = it.next()?;
                let right = it.next()?;
                TypedFunction::Mod {
                    left: Box::new(left),
                    right: Box::new(right),
                }
            }

            // ── Conversion ─────────────────────────────────────────
            "HEX" | "TO_HEX" => {
                let mut it = args.into_iter();
                TypedFunction::Hex {
                    expr: Box::new(it.next()?),
                }
            }
            "UNHEX" | "FROM_HEX" => {
                let mut it = args.into_iter();
                TypedFunction::Unhex {
                    expr: Box::new(it.next()?),
                }
            }
            "MD5" => {
                let mut it = args.into_iter();
                TypedFunction::Md5 {
                    expr: Box::new(it.next()?),
                }
            }
            "SHA" | "SHA1" => {
                let mut it = args.into_iter();
                TypedFunction::Sha {
                    expr: Box::new(it.next()?),
                }
            }
            "SHA2" | "SHA256" | "SHA512" => {
                let mut it = args.into_iter();
                let expr = it.next()?;
                let bit_length = it.next().unwrap_or(Expr::Number("256".to_string()));
                TypedFunction::Sha2 {
                    expr: Box::new(expr),
                    bit_length: Box::new(bit_length),
                }
            }

            // Not a recognized typed function
            _ => return None,
        };

        Some(Expr::TypedFunction {
            func: tf,
            filter: None,
            over: None,
        })
    }

    /// Try to extract a DateTimeField from a column-name expression.
    fn expr_to_datetime_field(expr: &Expr) -> Option<DateTimeField> {
        match expr {
            Expr::Column {
                name, table: None, ..
            } => match name.to_uppercase().as_str() {
                "YEAR" => Some(DateTimeField::Year),
                "QUARTER" => Some(DateTimeField::Quarter),
                "MONTH" => Some(DateTimeField::Month),
                "WEEK" => Some(DateTimeField::Week),
                "DAY" => Some(DateTimeField::Day),
                "HOUR" => Some(DateTimeField::Hour),
                "MINUTE" => Some(DateTimeField::Minute),
                "SECOND" => Some(DateTimeField::Second),
                "MILLISECOND" => Some(DateTimeField::Millisecond),
                "MICROSECOND" => Some(DateTimeField::Microsecond),
                _ => None,
            },
            Expr::StringLiteral(s) | Expr::NationalStringLiteral(s) => {
                match s.to_uppercase().as_str() {
                    "YEAR" => Some(DateTimeField::Year),
                    "QUARTER" => Some(DateTimeField::Quarter),
                    "MONTH" => Some(DateTimeField::Month),
                    "WEEK" => Some(DateTimeField::Week),
                    "DAY" => Some(DateTimeField::Day),
                    "HOUR" => Some(DateTimeField::Hour),
                    "MINUTE" => Some(DateTimeField::Minute),
                    "SECOND" => Some(DateTimeField::Second),
                    "MILLISECOND" => Some(DateTimeField::Millisecond),
                    "MICROSECOND" => Some(DateTimeField::Microsecond),
                    _ => None,
                }
            }
            _ => None,
        }
    }

    fn parse_case_expr(&mut self) -> Result<Expr> {
        self.expect(TokenType::Case)?;

        let operand = if self.peek_type() != &TokenType::When {
            Some(Box::new(self.parse_expr()?))
        } else {
            None
        };

        let mut when_clauses = Vec::new();
        while self.match_token(TokenType::When) {
            let condition = self.parse_expr()?;
            self.expect(TokenType::Then)?;
            let result = self.parse_expr()?;
            when_clauses.push((condition, result));
        }

        let else_clause = if self.match_token(TokenType::Else) {
            Some(Box::new(self.parse_expr()?))
        } else {
            None
        };

        self.expect(TokenType::End)?;

        Ok(Expr::Case {
            operand,
            when_clauses,
            else_clause,
        })
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_parse_simple_select() {
        let stmt = Parser::new("SELECT a, b FROM t")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                assert_eq!(sel.columns.len(), 2);
                assert!(sel.from.is_some());
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_select_with_where() {
        let stmt = Parser::new("SELECT x FROM t WHERE x > 10")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => assert!(sel.where_clause.is_some()),
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_select_wildcard() {
        let stmt = Parser::new("SELECT * FROM users")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                assert_eq!(sel.columns.len(), 1);
                assert!(matches!(sel.columns[0], SelectItem::Wildcard));
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_insert() {
        let stmt = Parser::new("INSERT INTO t (a, b) VALUES (1, 'hello')")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Insert(ins) => {
                assert_eq!(ins.table.name, "t");
                assert_eq!(ins.columns, vec!["a", "b"]);
                match &ins.source {
                    InsertSource::Values(rows) => {
                        assert_eq!(rows.len(), 1);
                        assert_eq!(rows[0].len(), 2);
                    }
                    _ => panic!("Expected VALUES"),
                }
            }
            _ => panic!("Expected INSERT"),
        }
    }

    #[test]
    fn test_parse_delete() {
        let stmt = Parser::new("DELETE FROM users WHERE id = 1")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Delete(del) => {
                assert_eq!(del.table.name, "users");
                assert!(del.where_clause.is_some());
            }
            _ => panic!("Expected DELETE"),
        }
    }

    #[test]
    fn test_parse_join() {
        let stmt = Parser::new("SELECT a.id, b.name FROM a INNER JOIN b ON a.id = b.a_id")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                assert_eq!(sel.joins.len(), 1);
                assert_eq!(sel.joins[0].join_type, JoinType::Inner);
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_cte() {
        let stmt = Parser::new("WITH cte AS (SELECT 1 AS x) SELECT x FROM cte")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                assert_eq!(sel.ctes.len(), 1);
                assert_eq!(sel.ctes[0].name, "cte");
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_union() {
        let stmt = Parser::new("SELECT 1 UNION ALL SELECT 2")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::SetOperation(sop) => {
                assert_eq!(sop.op, SetOperationType::Union);
                assert!(sop.all);
            }
            _ => panic!("Expected SetOperation"),
        }
    }

    #[test]
    fn test_parse_cast() {
        let stmt = Parser::new("SELECT CAST(x AS INT) FROM t")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                if let SelectItem::Expr { expr, .. } = &sel.columns[0] {
                    assert!(matches!(expr, Expr::Cast { .. }));
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_subquery() {
        let stmt = Parser::new("SELECT * FROM (SELECT 1 AS x) AS sub")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                if let Some(from) = &sel.from {
                    assert!(matches!(from.source, TableSource::Subquery { .. }));
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn cr014_paren_setop_derived_table_parses() {
        // CR-014: a parenthesised set operation used as a derived table, where
        // each branch is itself parenthesised, must parse. Previously failed
        // with `Expected RParen, got Except/Union/Intersect`.
        for op in ["EXCEPT", "UNION", "UNION ALL", "INTERSECT"] {
            let sql = format!("SELECT count(*) FROM ((SELECT 1) {op} (SELECT 2)) x");
            assert!(
                Parser::new(&sql).unwrap().parse_statements().is_ok(),
                "must parse: {sql}"
            );
        }
    }

    #[test]
    fn cr014_chained_except_derived_table_parses() {
        // TPC-DS q87 shape: chained EXCEPT of parenthesised branches.
        let sql = "SELECT count(*) FROM ((SELECT 1 AS a) EXCEPT (SELECT 2 AS a) \
                   EXCEPT (SELECT 3 AS a)) cool_cust";
        let stmt = Parser::new(sql).unwrap().parse_statement().unwrap();
        match stmt {
            Statement::Select(sel) => {
                let from = sel.from.expect("FROM clause present");
                match from.source {
                    TableSource::Subquery { query, alias, .. } => {
                        assert_eq!(alias.as_deref(), Some("cool_cust"));
                        assert!(matches!(*query, Statement::SetOperation(_)));
                    }
                    _ => panic!("Expected subquery derived table"),
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn cr014_controls_still_parse() {
        // Redundant nesting and no-branch-parens set-op were already OK; keep
        // them green. The parenthesised-join derived table must also still
        // parse (no regression from removing the paren-counting heuristic).
        for sql in [
            "SELECT count(*) FROM ((SELECT 1)) x",
            "SELECT count(*) FROM (SELECT 1 EXCEPT SELECT 2) x",
            "SELECT * FROM (a JOIN b ON a.id = b.id) x",
        ] {
            assert!(
                Parser::new(sql).unwrap().parse_statements().is_ok(),
                "must parse: {sql}"
            );
        }
    }

    #[test]
    fn test_parse_exists() {
        let stmt = Parser::new("SELECT * FROM t WHERE EXISTS (SELECT 1 FROM t2)")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                assert!(sel.where_clause.is_some());
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_window_function() {
        let stmt = Parser::new(
            "SELECT ROW_NUMBER() OVER (PARTITION BY dept ORDER BY salary DESC) FROM emp",
        )
        .unwrap()
        .parse_statement()
        .unwrap();
        match stmt {
            Statement::Select(sel) => {
                if let SelectItem::Expr { expr, .. } = &sel.columns[0] {
                    match expr {
                        Expr::TypedFunction { over, .. } => {
                            assert!(over.is_some());
                        }
                        Expr::Function { over, .. } => {
                            assert!(over.is_some());
                        }
                        _ => panic!("Expected function"),
                    }
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_multiple_statements() {
        let stmts = Parser::new("SELECT 1; SELECT 2;")
            .unwrap()
            .parse_statements()
            .unwrap();
        assert_eq!(stmts.len(), 2);
    }

    #[test]
    fn test_parse_insert_select() {
        let stmt = Parser::new("INSERT INTO t SELECT * FROM s")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Insert(ins) => {
                assert!(matches!(ins.source, InsertSource::Query(_)));
            }
            _ => panic!("Expected INSERT"),
        }
    }

    #[test]
    fn test_parse_create_table_constraints() {
        let stmt =
            Parser::new("CREATE TABLE t (id INT PRIMARY KEY, name VARCHAR(100) NOT NULL UNIQUE)")
                .unwrap()
                .parse_statement()
                .unwrap();
        match stmt {
            Statement::CreateTable(ct) => {
                assert_eq!(ct.columns.len(), 2);
                assert!(ct.columns[0].primary_key);
                assert!(ct.columns[1].unique);
            }
            _ => panic!("Expected CREATE TABLE"),
        }
    }

    #[test]
    fn test_parse_extract() {
        let stmt = Parser::new("SELECT EXTRACT(YEAR FROM created_at) FROM t")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                if let SelectItem::Expr { expr, .. } = &sel.columns[0] {
                    assert!(matches!(expr, Expr::Extract { .. }));
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_postgres_cast() {
        let stmt = Parser::new("SELECT x::int FROM t")
            .unwrap()
            .parse_statement()
            .unwrap();
        match stmt {
            Statement::Select(sel) => {
                if let SelectItem::Expr { expr, .. } = &sel.columns[0] {
                    assert!(matches!(expr, Expr::Cast { .. }));
                }
            }
            _ => panic!("Expected SELECT"),
        }
    }

    #[test]
    fn test_parse_on_conflict_expression_targets() {
        let stmt = Parser::new(
            "INSERT INTO t VALUES (1, 'Crowberry') ON CONFLICT (lower(fruit) collate \"C\" text_pattern_ops, key) DO NOTHING",
        )
        .unwrap()
        .parse_statement()
        .unwrap();

        match stmt {
            Statement::Insert(ins) => {
                let on_conflict = ins.on_conflict.expect("Expected ON CONFLICT");
                assert_eq!(on_conflict.columns.len(), 2);
                assert!(on_conflict.columns[0].starts_with("lower"));
                assert!(on_conflict.columns[0].contains("text_pattern_ops"));
                assert_eq!(on_conflict.columns[1], "key");
            }
            _ => panic!("Expected INSERT"),
        }
    }

    #[test]
    fn test_parse_postgres_operator_sequences() {
        let cases = [
            "SELECT * FROM box_temp WHERE f1 <<| '(10,4.33334),(5,100)'",
            "SELECT * FROM box_temp WHERE f1 &<| '(10,4.3333334),(5,1)'",
            "SELECT count(*) FROM radix_text_tbl WHERE t ^@ 'Worth'",
        ];

        for sql in &cases {
            let stmt = Parser::new(sql).unwrap().parse_statement().unwrap();
            assert!(matches!(stmt, Statement::Select(_)));
        }
    }
}

/// Attach comments to the appropriate field on a parsed statement.
fn attach_comments_to_statement(stmt: &mut Statement, comments: Vec<String>) {
    match stmt {
        Statement::Select(s) => s.comments = comments,
        Statement::Insert(s) => s.comments = comments,
        Statement::Update(s) => s.comments = comments,
        Statement::Delete(s) => s.comments = comments,
        Statement::CreateTable(s) => s.comments = comments,
        Statement::DropTable(s) => s.comments = comments,
        Statement::SetOperation(s) => s.comments = comments,
        Statement::AlterTable(s) => s.comments = comments,
        Statement::CreateView(s) => s.comments = comments,
        Statement::DropView(s) => s.comments = comments,
        Statement::Truncate(s) => s.comments = comments,
        Statement::Explain(s) => s.comments = comments,
        Statement::Use(s) => s.comments = comments,
        Statement::Merge(s) => s.comments = comments,
        Statement::Command(s) => s.comments = comments,
        // Transaction and Expression don't have comment fields
        Statement::Transaction(_) | Statement::Expression(_) => {}
    }
}