stryke/

lexer.rs

use crate::error::{ErrorKind, StrykeError, StrykeResult};
use crate::token::{keyword_or_ident, Token};

/// Private-use character for a literal `$` inside double-quoted / `qq` strings (from `\$` in source).
/// The parser maps this to `$` without variable interpolation (CPAN `eval qq/…/` code generators).
pub const LITERAL_DOLLAR_IN_DQUOTE: char = '\u{E000}';
/// Private-use character for a literal `@` inside double-quoted / `qq` strings (from `\@` in source).
/// Mirrors `LITERAL_DOLLAR_IN_DQUOTE`; suppresses array interpolation so `"\@x"` is the literal `@x`.
pub const LITERAL_AT_IN_DQUOTE: char = '\u{E001}';

/// Resolve `\N{U+XXXX}` hex codepoints and `\N{LATIN SMALL LETTER E}` Unicode character names.
fn parse_unicode_name(name: &str) -> Option<char> {
    if let Some(hex) = name.strip_prefix("U+") {
        let val = u32::from_str_radix(hex, 16).ok()?;
        char::from_u32(val)
    } else {
        unicode_names2::character(name)
    }
}

/// Flag letters after `m//`, `qr//`, etc. (`c` = `/gc`, `o` = compile once; CPAN uses both).
const REGEX_FLAG_CHARS: &str = "gimsxecor";
/// `Lexer` — see fields for layout.
pub struct Lexer {
    /// `input` field.
    input: Vec<char>,
    /// `pos` field.
    pos: usize,
    /// `line` field.
    pub line: usize,
    /// Line where the most recently-returned token starts. Set by
    /// [`Self::next_token`] right after its leading
    /// `skip_whitespace_and_comments` call so the value reflects the
    /// **emitted** token's source position even when the call recursed
    /// through a POD / heredoc skip (which advances `self.line` many
    /// lines before producing the real token). Read by [`Self::tokenize`].
    pub token_start_line: usize,
    /// Tracks whether the last token was a term (value/variable/close-delim)
    /// to disambiguate `/` as division vs regex and `{` as hash-ref vs block.
    last_was_term: bool,
    /// Tracks whether the last token was a method-call arrow (`->`). After
    /// `->`, identifiers `s` / `tr` / `y` / `q` / `qq` / `qw` / `qr` / `m`
    /// are method names — never substitution / transliteration / quote-like
    /// operators. Without this gate, `$obj->y` followed by `,` would consume
    /// `, …, …` as a transliteration body.
    last_was_arrow: bool,
    /// Snapshot of [`Self::last_was_arrow`] taken at the start of each
    /// [`Self::next_token`] call so identifier-decoding logic can read the
    /// previous-token state without racing against its own writes.
    prev_arrow: bool,
    /// Source path for [`StrykeError`] (e.g. real script or required `.pm` path).
    error_file: String,
    /// When > 0, the lexer treats `m` followed by `/` as a plain identifier
    /// instead of `m//` regex syntax. Used in thread/pipeline stages where
    /// `/m/` should be a regex grep filter, not `m//`.
    pub suppress_m_regex: u32,
    /// Set true by [`Self::next_token`] right before returning a token that
    /// originated from a *bare* positional alias (`_`, `_0`, `_1`, …) — i.e.
    /// without a leading `$` sigil. Read by [`Self::tokenize`] to record the
    /// emitted token's index in [`Self::bare_positional_indices`]. Reset to
    /// `false` at the top of every `next_token` call.
    pub last_was_bare_positional: bool,
    /// Indices into the token vector returned by [`Self::tokenize`] for every
    /// bare-positional token. Used by the parser's `my $X = EXPR` rule to
    /// auto-wrap an RHS that contains free positional aliases into an
    /// implicit zero-arg coderef (so `my $f = _ * 2` ≡ `my $f = fn { _ * 2 }`).
    pub bare_positional_indices: std::collections::HashSet<usize>,
}

impl Lexer {
    /// `new` — see implementation.
    pub fn new(input: &str) -> Self {
        Self::new_with_file(input, "-e")
    }
    /// `new_with_file` — see implementation.
    pub fn new_with_file(input: &str, file: impl Into<String>) -> Self {
        Self {
            input: input.chars().collect(),
            pos: 0,
            line: 1,
            token_start_line: 1,
            last_was_term: false,
            last_was_arrow: false,
            prev_arrow: false,
            error_file: file.into(),
            suppress_m_regex: 0,
            last_was_bare_positional: false,
            bare_positional_indices: std::collections::HashSet::new(),
        }
    }

    fn syntax_err(&self, message: impl Into<String>, line: usize) -> StrykeError {
        StrykeError::new(ErrorKind::Syntax, message, line, self.error_file.clone())
    }

    /// Used by the `s` / `tr` / `y` lexer arms when the identifier is followed
    /// by `,`. Returns `true` only when the rest of the statement looks like a
    /// genuine `s,PAT,REPL,FLAGS` / `tr,FROM,TO,FLAGS` shape — at least 2 more
    /// commas before the next statement terminator (`;`, newline, EOF, or
    /// closing brace/paren/bracket from the enclosing context). Without this
    /// gate, `struct Pt { x, y, z }` would consume `y, z }` as a transliteration
    /// body, and `$obj->y, ...` would eat the rest of the call.
    fn lookahead_is_comma_delim_subst(&self) -> bool {
        let mut commas = 0usize;
        let mut depth_paren = 0i32;
        let mut depth_bracket = 0i32;
        let mut depth_brace = 0i32;
        let mut i = self.pos;
        while i < self.input.len() {
            let c = self.input[i];
            match c {
                '\\' => {
                    i += 2; // skip escaped char (regex backslash escapes are common in pat/repl)
                    continue;
                }
                '(' => depth_paren += 1,
                ')' => {
                    if depth_paren == 0 {
                        break;
                    }
                    depth_paren -= 1;
                }
                '[' => depth_bracket += 1,
                ']' => {
                    if depth_bracket == 0 {
                        break;
                    }
                    depth_bracket -= 1;
                }
                '{' => depth_brace += 1,
                '}' => {
                    if depth_brace == 0 {
                        break;
                    }
                    depth_brace -= 1;
                }
                ';' | '\n' => break,
                ',' if depth_paren == 0 && depth_bracket == 0 && depth_brace == 0 => {
                    commas += 1;
                    if commas >= 3 {
                        return true;
                    }
                }
                _ => {}
            }
            i += 1;
        }
        // Need 3 total commas: `s,P,R,F` / `tr,F,T,F` (FLAGS may be empty,
        // but the third comma must still be present).
        commas >= 3
    }

    fn peek(&self) -> Option<char> {
        self.input.get(self.pos).copied()
    }

    fn peek_at(&self, offset: usize) -> Option<char> {
        self.input.get(self.pos + offset).copied()
    }

    /// True when `=` at `eq_pos` is Perl POD (`=head1`, `=cut`, …): first non-whitespace on the line.
    /// Otherwise `$_=foo` would misparse `=f` as POD and swallow the rest of the file.
    fn at_line_start_for_pod(&self, eq_pos: usize) -> bool {
        let mut i = eq_pos;
        while i > 0 {
            i -= 1;
            let c = self.input[i];
            if c == '\n' {
                return true;
            }
            if !c.is_whitespace() {
                return false;
            }
        }
        true
    }

    fn advance(&mut self) -> Option<char> {
        let ch = self.input.get(self.pos).copied();
        if let Some(c) = ch {
            if c == '\n' {
                self.line += 1;
            }
            self.pos += 1;
        }
        ch
    }

    fn skip_whitespace_and_comments(&mut self) {
        while self.pos < self.input.len() {
            let ch = self.input[self.pos];
            if ch == '#' {
                // Line comment
                while self.pos < self.input.len() && self.input[self.pos] != '\n' {
                    self.pos += 1;
                }
            } else if ch == '\\' && self.peek_at(1) == Some('\n') {
                // Backslash-newline: line continuation (shell-style)
                // Don't increment line — continued line is logically part of the same line
                self.pos += 2;
            } else if ch.is_whitespace() {
                if ch == '\n' {
                    self.line += 1;
                }
                self.pos += 1;
            } else {
                break;
            }
        }
    }

    /// Whitespace only — used after `q`/`qq`/`qr`/… before the opening delimiter so `#` is not
    /// mistaken for a line comment (`qr#...#`, `qw#...#`).
    fn skip_whitespace_only(&mut self) {
        while self.pos < self.input.len() {
            let ch = self.input[self.pos];
            if ch.is_whitespace() {
                if ch == '\n' {
                    self.line += 1;
                }
                self.pos += 1;
            } else {
                break;
            }
        }
    }

    fn read_while(&mut self, pred: impl Fn(char) -> bool) -> String {
        let mut s = String::new();
        while let Some(ch) = self.peek() {
            if pred(ch) {
                s.push(ch);
                self.advance();
            } else {
                break;
            }
        }
        s
    }

    /// Peek past whitespace and check whether the next token starts a range
    /// operator: `:`, `..`, `...`, or `~`. Used by the hex-integer lexer
    /// to switch into range-friendly DoubleString mode so `0x00:0xFF:1` and
    /// `0x00~0xFF~1` iterate with hex output instead of decimal. `~` here
    /// must NOT be `~>` / `~>>` (the thread macro) — those are operators on
    /// a value, not range starters.
    fn next_is_range_separator(&self) -> bool {
        let mut i = self.pos;
        while i < self.input.len() && matches!(self.input[i], ' ' | '\t') {
            i += 1;
        }
        if i >= self.input.len() {
            return false;
        }
        match self.input[i] {
            ':' => true,
            '.' if self.input.get(i + 1) == Some(&'.') => true,
            '~' if self.input.get(i + 1) != Some(&'>') => true,
            _ => false,
        }
    }

    /// `YYYY-MM-DD` / `YYYY-MM` lookahead. Called from [`Self::read_number`]
    /// when the just-consumed integer part is exactly 4 digits followed by
    /// `-<digit>`. Tries the longer `YYYY-MM-DD` shape first (full ISO date),
    /// falling back to `YYYY-MM` (year-month). Both shapes require valid
    /// month (01..=12) and, for the date form, valid day (01..=31). On match
    /// returns the literal string and advances `self.pos` past it; on
    /// failure restores `self.pos` so the caller falls through to the
    /// existing arithmetic-as-`-` path. The 4-digit year requirement is the
    /// disambiguator vs. plain subtraction (`2022-01-01` = date,
    /// `5-2-1` = arithmetic).
    fn try_consume_iso_date_tail(&mut self, start: usize) -> Option<String> {
        let saved = self.pos;
        let year: String = self.input[start..self.pos].iter().collect();
        if year.len() != 4 || year.parse::<u16>().is_err() {
            return None;
        }
        // Match `-MM`
        if self.peek() != Some('-') {
            return None;
        }
        if !self.peek_at(1).is_some_and(|c| c.is_ascii_digit())
            || !self.peek_at(2).is_some_and(|c| c.is_ascii_digit())
        {
            return None;
        }
        // Reject when third char after `-` is also a digit (e.g. `2022-100`)
        // — that's arithmetic, not a month.
        if self.peek_at(3).is_some_and(|c| c.is_ascii_digit()) {
            return None;
        }
        let month_str: String = self.input[self.pos + 1..self.pos + 3].iter().collect();
        let month: u8 = match month_str.parse() {
            Ok(m) if (1..=12).contains(&m) => m,
            _ => return None,
        };
        // Provisionally consume `-MM`
        self.advance(); // -
        self.advance(); // M
        self.advance(); // M
                        // Try `-DD` extension
        if self.peek() == Some('-')
            && self.peek_at(1).is_some_and(|c| c.is_ascii_digit())
            && self.peek_at(2).is_some_and(|c| c.is_ascii_digit())
            && !self.peek_at(3).is_some_and(|c| c.is_ascii_digit())
        {
            let day_str: String = self.input[self.pos + 1..self.pos + 3].iter().collect();
            if let Ok(day) = day_str.parse::<u8>() {
                if (1..=31).contains(&day) {
                    self.advance(); // -
                    self.advance(); // D
                    self.advance(); // D
                    let _ = month; // already validated
                    return Some(format!("{}-{}-{:02}", year, month_str, day));
                }
            }
        }
        // Year-month form `YYYY-MM`. Reject if followed by another `-DIGIT`
        // (would be arithmetic) — caught above by the third-digit guard.
        Some(format!("{}-{}", year, month_str)).filter(|_| {
            // No risky trailing chars beyond what we've already consumed.
            let _ = saved;
            true
        })
    }

    /// IPv6 lookahead from an arbitrary starting pos. Called from
    /// [`Self::read_number`] (digit-prefix), the identifier path (hex-letter
    /// prefix `fe80::1`), and the `:` lexer arm (zero-compressed prefix
    /// `::1`). `start` is where the IPv6 candidate begins in `self.input`;
    /// `self.pos` may already be partway through but is reset here so the
    /// scanner controls consumption. Greedily consumes hex digits, `:`, and
    /// at most one `::`, then validates with Rust's [`std::net::Ipv6Addr`]
    /// parser. On success returns the literal and leaves `self.pos` past
    /// it; on failure restores `self.pos` to its pre-call value.
    /// Acts only when the candidate has at least 2 colons — single-colon
    /// `1:5` is unambiguous range syntax, and 3-segment chains of pure-digit
    /// groups (`1:5:1`) never parse as IPv6 so range-with-step is preserved.
    fn try_consume_ipv6_tail(&mut self, start: usize) -> Option<String> {
        let saved = self.pos;
        self.pos = start;
        let mut seen_double_colon = false;
        let mut prev_was_colon = false;
        let mut colon_count = 0usize;
        while self.pos < self.input.len() {
            let c = self.input[self.pos];
            if c == ':' {
                colon_count += 1;
                if prev_was_colon {
                    if seen_double_colon {
                        break;
                    }
                    seen_double_colon = true;
                }
                prev_was_colon = true;
                self.advance();
                continue;
            }
            if c.is_ascii_hexdigit() {
                prev_was_colon = false;
                self.advance();
                continue;
            }
            break;
        }
        // Strip a trailing single colon (likely a range separator the lexer
        // greedily ate); a trailing `::` is part of the address.
        if self.pos > start
            && self.input[self.pos - 1] == ':'
            && (self.pos < start + 2 || self.input[self.pos - 2] != ':')
        {
            self.pos -= 1;
            colon_count -= 1;
        }
        // Package-separator disambiguator: when the candidate ends right
        // before an ASCII letter or `_` (an identifier continuation), the
        // `::` is almost certainly a package qualifier (`B::GV::SAFENAME`)
        // rather than IPv6 zero-compression. IPv6 lookahead bails so the
        // standard package-separator path runs.
        if self.pos < self.input.len() {
            let next = self.input[self.pos];
            if next.is_ascii_alphabetic() && !next.is_ascii_hexdigit() || next == '_' {
                self.pos = saved;
                return None;
            }
            // Three-segment package path: `A::B::C` greedy-matched `A::B` as
            // IPv6 zero-compressed (legal: `0:0:0:0:0:0:A:B`). Detect the
            // `::IDENT` continuation and bail so the standard PackageSep path
            // runs. Without this, `package A::B::C` lexes as
            // `Ident, DoubleString("A::B"), PackageSep, Ident("C")`.
            if next == ':'
                && self.input.get(self.pos + 1) == Some(&':')
                && self
                    .input
                    .get(self.pos + 2)
                    .is_some_and(|c| c.is_ascii_alphabetic() || *c == '_')
            {
                self.pos = saved;
                return None;
            }
        }
        let candidate: String = self.input[start..self.pos].iter().collect();
        // Require at least one hex digit. The bare `::` form is technically
        // valid IPv6 (all-zeros) but in real code it nearly always means
        // something else — array-slice default step (`@a[::]`), package
        // separator at the start of an empty list, etc. Users who want the
        // unspecified address can write `::0`.
        if colon_count < 2
            || !candidate.chars().any(|c| c.is_ascii_hexdigit())
            || candidate.parse::<std::net::Ipv6Addr>().is_err()
        {
            self.pos = saved;
            return None;
        }
        // Package-qualifier disambiguator: a 2-segment candidate
        // `LETTERS::LETTERS` with no digits anywhere parses as IPv6
        // (e.g. `d::ab` ≡ `d:0:0:0:0:0:0:ab`, `dead::beef`) but in
        // Perl/stryke source it is overwhelmingly a package qualifier
        // (`fn d::ab ($x) { ... }`, `Foo::Bar->method`). Real IPv6
        // literals in source carry at least one decimal digit (`fe80::1`,
        // `::1`, `1::dead`), which keeps them on the IPv6 path. Bare
        // pure-letter forms are routed back through PackageSep.
        if !candidate.chars().any(|c| c.is_ascii_digit()) {
            let segments: Vec<&str> = candidate.split("::").collect();
            if segments.len() == 2
                && segments
                    .iter()
                    .all(|s| !s.is_empty() && s.chars().all(|c| c.is_ascii_alphabetic()))
            {
                self.pos = saved;
                return None;
            }
        }
        Some(candidate)
    }

    /// IPv4 dotted-quad lookahead. Called from [`Self::read_number`] when the
    /// just-consumed integer part is followed by `.<digit>`. Speculatively
    /// matches 3 more `.<digits>` segments and accepts only when every octet
    /// parses as `u8` (0..=255). On match, returns the full dotted-quad
    /// string (e.g. `"192.168.255.255"`) and advances `self.pos` past it; on
    /// failure restores `self.pos` to its pre-call value so the caller falls
    /// through to the existing float-lexing path.
    fn try_consume_ipv4_tail(&mut self, start: usize) -> Option<String> {
        let saved = self.pos;
        // We've already consumed the first octet (start..self.pos).
        let first: String = self.input[start..self.pos].iter().collect();
        if first.parse::<u8>().is_err() {
            return None;
        }
        let mut octets: Vec<String> = vec![first];
        for _ in 0..3 {
            if self.peek() != Some('.') {
                self.pos = saved;
                return None;
            }
            if !self.peek_at(1).is_some_and(|c| c.is_ascii_digit()) {
                self.pos = saved;
                return None;
            }
            self.advance(); // consume '.'
            let oct_start = self.pos;
            while self.peek().is_some_and(|c| c.is_ascii_digit()) {
                self.advance();
            }
            let octet: String = self.input[oct_start..self.pos].iter().collect();
            if octet.parse::<u8>().is_err() {
                self.pos = saved;
                return None;
            }
            octets.push(octet);
        }
        // Reject 5-segment chains like `1.2.3.4.5` — the trailing `.<digit>`
        // means the user wrote something else (e.g. version number, list of
        // floats). Falling back to float lexing is safer than half-eating it.
        if self.peek() == Some('.') && self.peek_at(1).is_some_and(|c| c.is_ascii_digit()) {
            self.pos = saved;
            return None;
        }
        Some(octets.join("."))
    }

    fn read_number(&mut self) -> StrykeResult<Token> {
        let start = self.pos;
        let mut is_float = false;
        let mut is_hex = false;
        let mut is_oct = false;
        let mut is_bin = false;

        if self.peek() == Some('0') {
            match self.peek_at(1) {
                Some('x') | Some('X') => {
                    is_hex = true;
                    self.advance();
                    self.advance();
                }
                Some('b') | Some('B') => {
                    is_bin = true;
                    self.advance();
                    self.advance();
                }
                // `0o777` — Perl 5.34+ octal prefix (alongside the bare-`0`
                // form). (BUG-082) Read the same digit pool as bare `0...`
                // octals, but skip the `0o` prefix in the conversion.
                Some('o') | Some('O') => {
                    self.advance();
                    self.advance();
                    let digits = self.read_while(|c| c.is_ascii_digit() || c == '_');
                    let clean: String = digits.chars().filter(|&c| c != '_').collect();
                    let val = i64::from_str_radix(&clean, 8)
                        .map_err(|_| self.syntax_err("Invalid octal literal", self.line))?;
                    return Ok(Token::Integer(val));
                }
                Some(c) if c.is_ascii_digit() => {
                    is_oct = true;
                }
                _ => {}
            }
        }

        if is_hex {
            let digits = self.read_while(|c| c.is_ascii_hexdigit() || c == '_');
            let clean: String = digits.chars().filter(|&c| c != '_').collect();
            let val = i64::from_str_radix(&clean, 16)
                .map_err(|_| self.syntax_err("Invalid hex literal", self.line))?;
            // Range-context lookahead: `0x00:0xFF:1` should iterate as hex
            // strings (`0x00`, `0x01`, …, `0xFF`), preserving the leading
            // `0x` and case-of-digits. When the next non-whitespace token is
            // a range separator (`:`, `..`, `...`, or `!!!`), produce a
            // string-typed literal so the runtime range op can detect the
            // hex format and emit hex output. In all other contexts the
            // hex is a normal integer (arithmetic, assignment, etc.).
            if self.next_is_range_separator() {
                let raw: String = self.input[start..self.pos].iter().collect();
                return Ok(Token::DoubleString(raw));
            }
            return Ok(Token::Integer(val));
        }
        if is_bin {
            let digits = self.read_while(|c| c == '0' || c == '1' || c == '_');
            let clean: String = digits.chars().filter(|&c| c != '_').collect();
            let val = i64::from_str_radix(&clean, 2)
                .map_err(|_| self.syntax_err("Invalid binary literal", self.line))?;
            return Ok(Token::Integer(val));
        }

        // Decimal or octal
        let _int_part = self.read_while(|c| c.is_ascii_digit() || c == '_');
        // IPv4 dotted-quad lookahead: `192.168.255.255` should lex as ONE
        // string token, not as `192.168` (float) `.` `255.255` (float). Try
        // to consume 3 more `.NUM` segments where every octet is 0..=255 AND
        // not followed by another `.NUM` (so a 5-segment chain like
        // `1.2.3.4.5` cleanly fails the ipv4 path and falls back to floats).
        // Only fires when the current `.NUM` would have been a float decimal
        // — preserves all existing float lexing.
        if self.peek() == Some('.') && self.peek_at(1).is_some_and(|c| c.is_ascii_digit()) {
            if let Some(consumed) = self.try_consume_ipv4_tail(start) {
                return Ok(Token::DoubleString(consumed));
            }
            is_float = true;
            self.advance(); // consume '.'
            let _frac = self.read_while(|c| c.is_ascii_digit() || c == '_');
        }
        // ISO-date / year-month lookahead: `2022-01-01` and `2022-01`.
        // Distinct from arithmetic `2022 - 01 - 01` only because the lexer
        // greedily consumes the dotted form here. The 4-digit-year guard
        // inside [`Self::try_consume_iso_date_tail`] keeps `5-2-1` parsing
        // as arithmetic.
        if !is_float
            && self.peek() == Some('-')
            && self.peek_at(1).is_some_and(|c| c.is_ascii_digit())
        {
            if let Some(consumed) = self.try_consume_iso_date_tail(start) {
                return Ok(Token::DoubleString(consumed));
            }
        }
        // IPv6 lookahead: a hex-digit-only integer part followed by `:` and
        // more hex / `:` could be IPv6. Try to parse and accept; on failure
        // fall through to the existing range / arithmetic paths so plain
        // numeric ranges (`1:10`) keep their meaning.
        if !is_float && self.peek() == Some(':') {
            if let Some(consumed) = self.try_consume_ipv6_tail(start) {
                return Ok(Token::DoubleString(consumed));
            }
        }
        // Scientific notation
        if let Some('e') | Some('E') = self.peek() {
            is_float = true;
            self.advance();
            if let Some('+') | Some('-') = self.peek() {
                self.advance();
            }
            let _exp = self.read_while(|c| c.is_ascii_digit() || c == '_');
        }

        let raw: String = self.input[start..self.pos].iter().collect();
        let clean: String = raw.chars().filter(|&c| c != '_').collect();

        if is_float {
            let val: f64 = clean
                .parse()
                .map_err(|_| self.syntax_err("Invalid float literal", self.line))?;
            Ok(Token::Float(val))
        } else if is_oct && clean.starts_with('0') && clean.len() > 1 {
            let val = i64::from_str_radix(&clean[1..], 8)
                .map_err(|_| self.syntax_err("Invalid octal literal", self.line))?;
            Ok(Token::Integer(val))
        } else {
            let val: i64 = clean
                .parse()
                .map_err(|_| self.syntax_err("Invalid integer literal", self.line))?;
            Ok(Token::Integer(val))
        }
    }

    fn read_single_quoted_string(&mut self) -> StrykeResult<Token> {
        self.advance(); // consume opening '
        let mut s = String::new();
        loop {
            match self.advance() {
                Some('\\') => match self.peek() {
                    Some('\\') => {
                        s.push('\\');
                        self.advance();
                    }
                    Some('\'') => {
                        s.push('\'');
                        self.advance();
                    }
                    _ => s.push('\\'),
                },
                Some('\'') => break,
                Some(c) => s.push(c),
                None => return Err(self.syntax_err("Unterminated single-quoted string", self.line)),
            }
        }
        Ok(Token::SingleString(s))
    }

    fn read_double_quoted_string(&mut self) -> StrykeResult<Token> {
        self.advance(); // consume opening "
                        // Triple-quoted form: `"""..."""` — multi-line interpolating string.
                        // The opening `"` was just consumed; if the next two chars are also
                        // `"`, we're in triple-quote mode. Read until the closing `"""`,
                        // preserving raw newlines (no indent stripping). Interpolation
                        // (`$var`, `@arr`, `#{expr}`) flows through unchanged because the
                        // resulting `Token::DoubleString` body goes through the same
                        // downstream interpolator as a normal `"..."`.
        if self.peek() == Some('"') && self.peek_at(1) == Some('"') {
            self.advance(); // consume 2nd "
            self.advance(); // consume 3rd "
            let s = self.read_triple_quoted_body(true)?;
            return Ok(Token::DoubleString(s));
        }
        let s = self.read_interpolating_until('"')?;
        Ok(Token::DoubleString(s))
    }

    /// Read the body of a triple-quoted string up to and including the
    /// closing `"""`. `interpolate=true` honors backslash escapes the same
    /// way `read_escaped_until` does (so `\n`, `\t`, `\\`, `\$`, `\@`,
    /// `\"` etc. work inside `"""..."""`). `interpolate=false` is "raw"
    /// mode: every byte is copied verbatim, including backslashes; the
    /// only way out is a literal `"""`.
    ///
    /// Newlines are preserved verbatim — no indent stripping. The user
    /// chose the indentation; we don't second-guess it.
    fn read_triple_quoted_body(&mut self, interpolate: bool) -> StrykeResult<String> {
        let mut s = String::new();
        loop {
            // Check for closing `"""` at the current position.
            if self.peek() == Some('"')
                && self.peek_at(1) == Some('"')
                && self.peek_at(2) == Some('"')
            {
                self.advance(); // 1st "
                self.advance(); // 2nd "
                self.advance(); // 3rd "
                return Ok(s);
            }
            let c = match self.advance() {
                Some(c) => c,
                None => {
                    return Err(self.syntax_err(
                        "Unterminated triple-quoted string (missing closing \"\"\")",
                        self.line,
                    ))
                }
            };
            if interpolate && c == '\\' {
                // Handle escapes the same way single-line `"..."` does.
                // Reuse the existing escape table by hand for the common
                // cases; anything we don't recognise falls through as
                // `\` followed by the next char (matching Perl's
                // permissive double-quote escape behavior).
                let next = self.advance();
                match next {
                    Some('n') => s.push('\n'),
                    Some('t') => s.push('\t'),
                    Some('r') => s.push('\r'),
                    Some('\\') => s.push('\\'),
                    Some('"') => s.push('"'),
                    Some('$') => s.push(LITERAL_DOLLAR_IN_DQUOTE),
                    Some('@') => s.push(LITERAL_AT_IN_DQUOTE),
                    Some('0') => s.push('\0'),
                    Some('a') => s.push('\x07'),
                    Some('b') => s.push('\x08'),
                    Some('f') => s.push('\x0C'),
                    Some('e') => s.push('\x1B'),
                    Some(other) => {
                        s.push('\\');
                        s.push(other);
                    }
                    None => {
                        return Err(self.syntax_err(
                            "Unterminated escape at end of triple-quoted string",
                            self.line,
                        ))
                    }
                }
                continue;
            }
            if c == '\n' {
                self.line += 1;
            }
            s.push(c);
        }
    }

    /// Same as [`Self::read_escaped_until`] but preserves `\1`..`\9` verbatim
    /// (Perl's `s/(\d+)/\1/` numbered back-reference). Multi-digit octals
    /// (`\012`, `\077`) still resolve here, matching Perl's documented
    /// "two-or-more digit escapes are octal" rule.
    fn read_substitution_replacement(&mut self, term: char) -> StrykeResult<String> {
        self.read_escaped_until_inner(term, true, false)
    }

    fn read_escaped_until(&mut self, term: char) -> StrykeResult<String> {
        self.read_escaped_until_inner(term, false, false)
    }

    /// Like [`Self::read_escaped_until`] but for interpolating bodies
    /// (`"…"`, `qq…`, backticks): `#{…}` / `${…}` / `@{…}` expression
    /// regions are tracked so a `"` (or whatever `term` is) inside them
    /// does not end the string — `"ok #{"tom"}"` lexes as one token.
    fn read_interpolating_until(&mut self, term: char) -> StrykeResult<String> {
        self.read_escaped_until_inner(term, false, true)
    }

    /// Body parser for `q{…}`, `s/.../.../`, etc. When `defer_single_digit` is
    /// `true`, `\1`..`\9` (followed by a non-octal-digit) are preserved
    /// verbatim so the substitution layer can expand them as numbered
    /// back-references (Perl's documented `s///` behavior). Multi-digit
    /// octals (`\012`) still resolve numerically in both modes.
    fn read_escaped_until_inner(
        &mut self,
        term: char,
        defer_single_digit: bool,
        interp: bool,
    ) -> StrykeResult<String> {
        // `#{expr}` / `${expr}` / `@{expr}` interpolation-region state
        // (`interp` mode only). While `depth > 0` the terminator does not
        // end the string, and `{` / `}` inside nested '…' / "…" literals
        // (with `\` escapes) don't skew the depth count. State is tracked
        // over the chars *pushed to the body* — the downstream interpolator
        // re-lexes that body, so its escaping rules are the ones that govern.
        fn interp_track(c: char, depth: &mut usize, quote: &mut Option<char>, esc: &mut bool) {
            if let Some(q) = *quote {
                if *esc {
                    *esc = false;
                } else if c == '\\' {
                    *esc = true;
                } else if c == q {
                    *quote = None;
                }
            } else {
                match c {
                    '\'' | '"' => *quote = Some(c),
                    '{' => *depth += 1,
                    '}' => *depth = depth.saturating_sub(1),
                    _ => {}
                }
            }
        }
        let mut interp_depth = 0usize;
        let mut interp_quote: Option<char> = None;
        let mut interp_esc = false;
        let mut s = String::new();
        loop {
            match self.advance() {
                Some('\\') => match self.advance() {
                    Some('n') => s.push('\n'),
                    Some('t') => s.push('\t'),
                    Some('r') => s.push('\r'),
                    Some('\\') => {
                        if interp_depth > 0 {
                            interp_track(
                                '\\',
                                &mut interp_depth,
                                &mut interp_quote,
                                &mut interp_esc,
                            );
                        }
                        s.push('\\');
                    }
                    Some(c @ '0'..='7') => {
                        // In substitution-replacement mode, defer `\1`..`\9`
                        // (with no further octal digit) so the replacement
                        // expander can read them as numbered back-refs.
                        if defer_single_digit && c != '0' && !matches!(self.peek(), Some('0'..='7'))
                        {
                            s.push('\\');
                            s.push(c);
                            continue;
                        }
                        let mut oct = String::new();
                        oct.push(c);
                        for _ in 0..2 {
                            match self.peek() {
                                Some(d) if ('0'..='7').contains(&d) => {
                                    oct.push(self.advance().unwrap());
                                }
                                _ => break,
                            }
                        }
                        let val = u32::from_str_radix(&oct, 8).unwrap();
                        let ch = char::from_u32(val)
                            .ok_or_else(|| self.syntax_err("Invalid octal escape", self.line))?;
                        s.push(ch);
                    }
                    Some('a') => s.push('\x07'),
                    Some('b') => s.push('\x08'),
                    Some('f') => s.push('\x0C'),
                    Some('e') => s.push('\x1B'),
                    Some('$') => s.push(LITERAL_DOLLAR_IN_DQUOTE),
                    Some('@') => s.push(LITERAL_AT_IN_DQUOTE),
                    Some('c') => {
                        let ch = self
                            .advance()
                            .ok_or_else(|| self.syntax_err("Unterminated \\c escape", self.line))?;
                        s.push(char::from(ch.to_ascii_uppercase() as u8 ^ 0x40));
                    }
                    Some('o') if self.peek() == Some('{') => {
                        self.advance(); // '{'
                        let oct = self.read_while(|c| c != '}');
                        if self.peek() != Some('}') {
                            return Err(
                                self.syntax_err("Unterminated \\o{...} in string", self.line)
                            );
                        }
                        self.advance(); // '}'
                        if oct.is_empty() {
                            return Err(self.syntax_err("Empty \\o{} in string", self.line));
                        }
                        let val = u32::from_str_radix(&oct, 8).map_err(|_| {
                            self.syntax_err("Invalid octal digits in \\o{...}", self.line)
                        })?;
                        let c = char::from_u32(val).ok_or_else(|| {
                            self.syntax_err("Invalid Unicode scalar value in \\o{...}", self.line)
                        })?;
                        s.push(c);
                    }
                    Some('u') if self.peek() == Some('{') => {
                        self.advance(); // '{'
                        let hex = self.read_while(|c| c != '}');
                        if self.peek() != Some('}') {
                            return Err(
                                self.syntax_err("Unterminated \\u{...} in string", self.line)
                            );
                        }
                        self.advance(); // '}'
                        if hex.is_empty() {
                            return Err(self.syntax_err("Empty \\u{} in string", self.line));
                        }
                        let val = u32::from_str_radix(&hex, 16).map_err(|_| {
                            self.syntax_err("Invalid hex digits in \\u{...}", self.line)
                        })?;
                        let c = char::from_u32(val).ok_or_else(|| {
                            self.syntax_err("Invalid Unicode scalar value in \\u{...}", self.line)
                        })?;
                        s.push(c);
                    }
                    Some('N') if self.peek() == Some('{') => {
                        self.advance(); // '{'
                        let name = self.read_while(|c| c != '}');
                        if self.peek() != Some('}') {
                            return Err(
                                self.syntax_err("Unterminated \\N{...} in string", self.line)
                            );
                        }
                        self.advance(); // '}'
                        if name.is_empty() {
                            return Err(self.syntax_err("Empty \\N{} in string", self.line));
                        }
                        let c = parse_unicode_name(&name).ok_or_else(|| {
                            self.syntax_err(
                                format!("Unknown Unicode character name: {name}"),
                                self.line,
                            )
                        })?;
                        s.push(c);
                    }
                    Some('x') => {
                        if self.peek() == Some('{') {
                            self.advance(); // '{'
                            let hex = self.read_while(|c| c != '}');
                            if self.peek() != Some('}') {
                                return Err(
                                    self.syntax_err("Unterminated \\x{...} in string", self.line)
                                );
                            }
                            self.advance(); // '}'
                            if hex.is_empty() {
                                return Err(self.syntax_err("Empty \\x{} in string", self.line));
                            }
                            let val = u32::from_str_radix(&hex, 16).map_err(|_| {
                                self.syntax_err("Invalid hex digits in \\x{...}", self.line)
                            })?;
                            let c = char::from_u32(val).ok_or_else(|| {
                                self.syntax_err(
                                    "Invalid Unicode scalar value in \\x{...}",
                                    self.line,
                                )
                            })?;
                            s.push(c);
                        } else {
                            // Unbraced: up to two hex digits (Perl: "\\x414" is "\\x41" + "4").
                            let mut hex = String::new();
                            for _ in 0..2 {
                                match self.peek() {
                                    Some(c) if c.is_ascii_hexdigit() => {
                                        hex.push(self.advance().unwrap());
                                    }
                                    _ => break,
                                }
                            }
                            if hex.is_empty() {
                                // Perl: bare "\\x" in a string yields NUL.
                                s.push('\0');
                            } else if let Ok(val) = u32::from_str_radix(&hex, 16) {
                                if let Some(c) = char::from_u32(val) {
                                    s.push(c);
                                } else {
                                    return Err(self.syntax_err(
                                        "Invalid code point in \\x escape",
                                        self.line,
                                    ));
                                }
                            }
                        }
                    }
                    Some(c) if c == term => {
                        if interp_depth > 0 {
                            interp_track(c, &mut interp_depth, &mut interp_quote, &mut interp_esc);
                        }
                        s.push(c);
                    }
                    Some(c) => {
                        if interp_depth > 0 {
                            interp_track(
                                '\\',
                                &mut interp_depth,
                                &mut interp_quote,
                                &mut interp_esc,
                            );
                            interp_track(c, &mut interp_depth, &mut interp_quote, &mut interp_esc);
                        }
                        s.push('\\');
                        s.push(c);
                    }
                    None => return Err(self.syntax_err("Unterminated string", self.line)),
                },
                Some(c) if c == term && interp_depth == 0 => break,
                Some(c) => {
                    if interp_depth > 0 {
                        interp_track(c, &mut interp_depth, &mut interp_quote, &mut interp_esc);
                    } else if interp
                        && matches!(c, '$' | '@' | '#')
                        && self.peek() == Some('{')
                        && (c != '#' || !crate::compat_mode())
                    {
                        // `${…}` / `@{…}` always interpolate; `#{…}` is a
                        // stryke extension, literal under --compat.
                        interp_depth = 1;
                        s.push(c);
                        self.advance(); // the `{`
                        s.push('{');
                        continue;
                    }
                    s.push(c);
                }
                None => return Err(self.syntax_err("Unterminated string", self.line)),
            }
        }
        Ok(s)
    }

    /// `q(...)` / `qq(...)` with pairing delimiters — Perl balances nested `()`, `[]`, `{}`, `<>`
    /// so `q(sub ($) { 1 })` does not end at the `)` in `($)` (core `Carp.pm` uses `eval(q(...))`).
    fn read_q_qq_balanced_body(
        &mut self,
        open: char,
        close: char,
        is_qq: bool,
    ) -> StrykeResult<String> {
        let mut s = String::new();
        let mut depth: usize = 1;
        loop {
            match self.peek() {
                Some('\\') => {
                    self.advance();
                    if is_qq {
                        match self.advance() {
                            Some('n') => s.push('\n'),
                            Some('t') => s.push('\t'),
                            Some('r') => s.push('\r'),
                            Some('\\') => s.push('\\'),
                            Some(c @ '0'..='7') => {
                                let mut oct = String::new();
                                oct.push(c);
                                for _ in 0..2 {
                                    match self.peek() {
                                        Some(d) if ('0'..='7').contains(&d) => {
                                            oct.push(self.advance().unwrap());
                                        }
                                        _ => break,
                                    }
                                }
                                let val = u32::from_str_radix(&oct, 8).unwrap();
                                let ch = char::from_u32(val).ok_or_else(|| {
                                    self.syntax_err("Invalid octal escape", self.line)
                                })?;
                                s.push(ch);
                            }
                            Some('a') => s.push('\x07'),
                            Some('b') => s.push('\x08'),
                            Some('f') => s.push('\x0C'),
                            Some('e') => s.push('\x1B'),
                            Some('$') => s.push(LITERAL_DOLLAR_IN_DQUOTE),
                            Some('@') => s.push(LITERAL_AT_IN_DQUOTE),
                            Some('c') => {
                                let ch = self.advance().ok_or_else(|| {
                                    self.syntax_err("Unterminated \\c escape", self.line)
                                })?;
                                s.push(char::from(ch.to_ascii_uppercase() as u8 ^ 0x40));
                            }
                            Some('o') if self.peek() == Some('{') => {
                                self.advance();
                                let oct = self.read_while(|c| c != '}');
                                if self.peek() != Some('}') {
                                    return Err(self.syntax_err(
                                        "Unterminated \\o{...} in qq string",
                                        self.line,
                                    ));
                                }
                                self.advance();
                                if oct.is_empty() {
                                    return Err(
                                        self.syntax_err("Empty \\o{} in qq string", self.line)
                                    );
                                }
                                let val = u32::from_str_radix(&oct, 8).map_err(|_| {
                                    self.syntax_err("Invalid octal digits in \\o{...}", self.line)
                                })?;
                                let c = char::from_u32(val).ok_or_else(|| {
                                    self.syntax_err(
                                        "Invalid Unicode scalar value in \\o{...}",
                                        self.line,
                                    )
                                })?;
                                s.push(c);
                            }
                            Some('u') if self.peek() == Some('{') => {
                                self.advance();
                                let hex = self.read_while(|c| c != '}');
                                if self.peek() != Some('}') {
                                    return Err(self.syntax_err(
                                        "Unterminated \\u{...} in qq string",
                                        self.line,
                                    ));
                                }
                                self.advance();
                                if hex.is_empty() {
                                    return Err(
                                        self.syntax_err("Empty \\u{} in qq string", self.line)
                                    );
                                }
                                let val = u32::from_str_radix(&hex, 16).map_err(|_| {
                                    self.syntax_err("Invalid hex digits in \\u{...}", self.line)
                                })?;
                                let c = char::from_u32(val).ok_or_else(|| {
                                    self.syntax_err(
                                        "Invalid Unicode scalar value in \\u{...}",
                                        self.line,
                                    )
                                })?;
                                s.push(c);
                            }
                            Some('N') if self.peek() == Some('{') => {
                                self.advance();
                                let name = self.read_while(|c| c != '}');
                                if self.peek() != Some('}') {
                                    return Err(self.syntax_err(
                                        "Unterminated \\N{...} in qq string",
                                        self.line,
                                    ));
                                }
                                self.advance();
                                if name.is_empty() {
                                    return Err(
                                        self.syntax_err("Empty \\N{} in qq string", self.line)
                                    );
                                }
                                let c = parse_unicode_name(&name).ok_or_else(|| {
                                    self.syntax_err(
                                        format!("Unknown Unicode character name: {name}"),
                                        self.line,
                                    )
                                })?;
                                s.push(c);
                            }
                            Some('x') => {
                                if self.peek() == Some('{') {
                                    self.advance();
                                    let hex = self.read_while(|c| c != '}');
                                    if self.peek() != Some('}') {
                                        return Err(self.syntax_err(
                                            "Unterminated \\x{...} in qq string",
                                            self.line,
                                        ));
                                    }
                                    self.advance();
                                    if hex.is_empty() {
                                        return Err(
                                            self.syntax_err("Empty \\x{} in qq string", self.line)
                                        );
                                    }
                                    let val = u32::from_str_radix(&hex, 16).map_err(|_| {
                                        self.syntax_err("Invalid hex digits in \\x{...}", self.line)
                                    })?;
                                    let c = char::from_u32(val).ok_or_else(|| {
                                        self.syntax_err(
                                            "Invalid Unicode scalar value in \\x{...}",
                                            self.line,
                                        )
                                    })?;
                                    s.push(c);
                                } else {
                                    let mut hex = String::new();
                                    for _ in 0..2 {
                                        match self.peek() {
                                            Some(c) if c.is_ascii_hexdigit() => {
                                                hex.push(self.advance().unwrap());
                                            }
                                            _ => break,
                                        }
                                    }
                                    if hex.is_empty() {
                                        s.push('\0');
                                    } else if let Ok(val) = u32::from_str_radix(&hex, 16) {
                                        if let Some(c) = char::from_u32(val) {
                                            s.push(c);
                                        } else {
                                            return Err(self.syntax_err(
                                                "Invalid code point in \\x escape",
                                                self.line,
                                            ));
                                        }
                                    }
                                }
                            }
                            Some(c) if c == close && depth == 1 => s.push(close),
                            Some(c) => {
                                s.push('\\');
                                s.push(c);
                            }
                            None => {
                                return Err(
                                    self.syntax_err("Unterminated qq(...) string", self.line)
                                );
                            }
                        }
                    } else {
                        match self.advance() {
                            Some(c) if c == close && depth == 1 => s.push(close),
                            Some(c) => {
                                s.push('\\');
                                s.push(c);
                            }
                            None => {
                                return Err(
                                    self.syntax_err("Unterminated q(...) string", self.line)
                                );
                            }
                        }
                    }
                }
                Some(c) if c == open => {
                    self.advance();
                    depth += 1;
                    s.push(open);
                }
                Some(c) if c == close => {
                    self.advance();
                    if depth == 1 {
                        break;
                    }
                    depth -= 1;
                    s.push(close);
                }
                Some(c) => {
                    self.advance();
                    s.push(c);
                }
                None => {
                    return Err(self.syntax_err("Unterminated q/qq bracketed string", self.line));
                }
            }
        }
        Ok(s)
    }

    fn read_regex(&mut self) -> StrykeResult<Token> {
        self.advance(); // consume opening /
        let mut pattern = String::new();
        loop {
            match self.advance() {
                Some('\\') => {
                    pattern.push('\\');
                    if let Some(c) = self.advance() {
                        pattern.push(c);
                    }
                }
                Some('/') => break,
                Some(c) => pattern.push(c),
                None => return Err(self.syntax_err("Unterminated regex", self.line)),
            }
        }
        let flags = self.read_while(|c| REGEX_FLAG_CHARS.contains(c));
        Ok(Token::Regex(pattern, flags, '/'))
    }

    fn read_qw(&mut self) -> StrykeResult<Token> {
        // Already consumed 'qw', now expect delimiter
        self.skip_whitespace_only();
        let open = self
            .advance()
            .ok_or_else(|| self.syntax_err("Expected delimiter after qw", self.line))?;
        let close = match open {
            '(' => ')',
            '[' => ']',
            '{' => '}',
            '<' => '>',
            c => c,
        };
        let mut words = Vec::new();
        if matches!(open, '(' | '[' | '{' | '<') {
            // Perl balances nested delimiters in `qw( ... )` / `qw[ ... ]` / … so
            // `qw( (SV*)pWARN_ALL )` is one word (core `B.pm` line 88).
            let mut depth: usize = 1;
            let mut buf = String::new();
            loop {
                match self.peek() {
                    None => {
                        return Err(self.syntax_err("Unterminated qw()", self.line));
                    }
                    Some(c) if depth == 1 && c.is_whitespace() => {
                        self.advance();
                        if !buf.is_empty() {
                            words.push(buf.clone());
                            buf.clear();
                        }
                        while self.peek().is_some_and(|c| c.is_whitespace()) {
                            self.advance();
                        }
                    }
                    Some(c) if c == close && depth == 1 => {
                        self.advance();
                        if !buf.is_empty() {
                            words.push(buf);
                        }
                        break;
                    }
                    Some(c) if c == open => {
                        depth += 1;
                        buf.push(self.advance().unwrap());
                    }
                    Some(c) if c == close => {
                        // `depth == 1 && close` is handled above (final qw delimiter).
                        debug_assert!(depth >= 2);
                        depth -= 1;
                        buf.push(self.advance().unwrap());
                    }
                    Some(_) => {
                        buf.push(self.advance().unwrap());
                    }
                }
            }
            return Ok(Token::QW(words));
        }
        loop {
            // Skip whitespace inside qw
            while let Some(ch) = self.peek() {
                if ch.is_whitespace() {
                    self.advance();
                } else {
                    break;
                }
            }
            if self.peek() == Some(close) {
                self.advance();
                break;
            }
            if self.peek().is_none() {
                return Err(self.syntax_err("Unterminated qw()", self.line));
            }
            let word = self.read_while(|c| !c.is_whitespace() && c != close);
            if !word.is_empty() {
                words.push(word);
            }
        }
        Ok(Token::QW(words))
    }

    fn read_heredoc_tag(&mut self) -> StrykeResult<(String, bool, bool)> {
        self.read_heredoc_tag_inner(false)
    }

    fn read_heredoc_tag_inner(&mut self, indented: bool) -> StrykeResult<(String, bool, bool)> {
        // We've consumed '<<'. Now figure out the tag.
        // Returns (tag, interpolate, indented).
        let quoted;
        let tag;
        match self.peek() {
            Some('\'') => {
                self.advance();
                tag = self.read_while(|c| c != '\'');
                self.advance(); // closing quote
                quoted = false; // no interpolation
            }
            Some('"') => {
                self.advance();
                tag = self.read_while(|c| c != '"');
                self.advance();
                quoted = true;
            }
            Some('~') => {
                self.advance(); // indented heredoc
                return self.read_heredoc_tag_inner(true); // recurse with indented=true
            }
            _ => {
                tag = self.read_while(|c| c.is_alphanumeric() || c == '_');
                quoted = true;
            }
        }
        Ok((tag, quoted, indented))
    }

    fn read_heredoc_body(&mut self, tag: &str, indented: bool) -> StrykeResult<String> {
        // Read until we find a line that is exactly the tag (or, for indented heredocs,
        // a line whose trimmed content equals the tag).
        let mut lines: Vec<String> = Vec::new();
        // First, skip to end of current line
        while let Some(ch) = self.peek() {
            if ch == '\n' {
                self.advance();
                break;
            }
            self.advance();
        }
        let mut terminator_indent: Option<usize> = None;
        loop {
            let _line_start = self.pos;
            let line = self.read_while(|c| c != '\n');
            if line.trim() == tag {
                // For indented heredocs, the terminator's leading whitespace determines
                // how much to strip from all body lines.
                if indented {
                    terminator_indent = Some(line.len() - line.trim_start().len());
                }
                break;
            }
            lines.push(line);
            if self.peek() == Some('\n') {
                self.advance();
            } else if self.pos >= self.input.len() {
                return Err(self.syntax_err(
                    format!("Unterminated heredoc (looking for '{tag}')"),
                    self.line,
                ));
            }
        }
        if self.peek() == Some('\n') {
            self.advance();
        }
        // For indented heredocs (<<~), strip leading whitespace from each line,
        // up to the amount of indentation on the terminator line.
        if indented {
            let strip = terminator_indent.unwrap_or(0);
            let mut body = String::new();
            for line in lines {
                let ws_count = line.len() - line.trim_start().len();
                let to_strip = ws_count.min(strip);
                body.push_str(&line[to_strip..]);
                body.push('\n');
            }
            Ok(body)
        } else {
            let mut body = String::new();
            for line in lines {
                body.push_str(&line);
                body.push('\n');
            }
            Ok(body)
        }
    }

    fn read_identifier(&mut self) -> String {
        self.read_while(|c| c.is_alphanumeric() || c == '_')
    }

    /// `Foo::Bar::Baz` after the leading sigil.
    fn read_package_qualified_identifier(&mut self) -> String {
        let mut s = self.read_identifier();
        while self.peek() == Some(':') && self.input.get(self.pos + 1) == Some(&':') {
            self.advance();
            self.advance();
            s.push_str("::");
            s.push_str(&self.read_identifier());
        }
        s
    }

    /// Body lines for `format N =` … `.` (excluding the closing `.` line).
    fn read_format_body(&mut self) -> StrykeResult<Vec<String>> {
        while self.peek().is_some_and(|c| c == ' ' || c == '\t') {
            self.advance();
        }
        if self.peek() == Some('\n') {
            self.advance();
        }
        let mut lines = Vec::new();
        loop {
            let mut line = String::new();
            while let Some(c) = self.peek() {
                if c == '\n' {
                    self.advance();
                    break;
                }
                if c == '\r' {
                    self.advance();
                    if self.peek() == Some('\n') {
                        self.advance();
                    }
                    break;
                }
                line.push(c);
                self.advance();
            }
            if line.trim() == "." {
                break;
            }
            lines.push(line);
            if self.peek().is_none() {
                return Err(self.syntax_err(
                    "Unterminated format (expected '.' on its own line before end of file)",
                    self.line,
                ));
            }
        }
        Ok(lines)
    }

    fn read_variable_name(&mut self) -> String {
        // Handle special vars like $_, $!, $0, $/, $^I, etc.
        match self.peek() {
            // Second `$` in `$$_{` — with leading `$` already consumed, we have `$` `_` `{` → `$_` then `{`.
            Some('$')
                if self.input.get(self.pos + 1) == Some(&'_')
                    && self.input.get(self.pos + 2) == Some(&'{') =>
            {
                self.advance(); // second $
                self.advance(); // `_` of `$_`
                "_".to_string()
            }
            // `$::{$key}` / `$::Foo` — stash access (`%::`) and package names rooted at `::` (Perl `$::` ≡ main stash).
            Some(':') if self.input.get(self.pos + 1) == Some(&':') => {
                self.advance();
                self.advance();
                let mut s = "::".to_string();
                if self.peek().is_some_and(|c| c.is_alphabetic() || c == '_') {
                    s.push_str(&self.read_identifier());
                }
                while self.peek() == Some(':') && self.input.get(self.pos + 1) == Some(&':') {
                    self.advance();
                    self.advance();
                    s.push_str("::");
                    s.push_str(&self.read_identifier());
                }
                s
            }
            Some(c) if c.is_alphabetic() || c == '_' => {
                let mut ident = self.read_package_qualified_identifier();
                // `$main::!`, `$main::?`, `$main::0` — stryke
                // implements the Perl docs faithfully ("All
                // punctuation variables like $_ reside in main"). The
                // qualified identifier ends with the trailing `::`
                // (because the punctuation leaf isn't alphanumeric);
                // pick up a single punctuation / digit char or a
                // caret-prefixed letter (`$main::^O`) as the leaf.
                // The scope getter canonicalizes `main::PUNCT` →
                // `PUNCT` via `strip_main_prefix` so storage stays
                // unified.
                //
                // Disabled in `--compat`: Perl 5.42's parser rejects
                // `$main::!` (treats `$main::` as the empty-name var
                // and leaves `!` for the next token). To stay
                // byte-identical to perl(1) we don't consume the
                // punct leaf when compat mode is on. Stryke-strict
                // default mode still gets the docs-faithful behavior.
                if ident.ends_with("::") && !crate::compat_mode() {
                    match self.peek() {
                        Some('^')
                            if self
                                .input
                                .get(self.pos + 1)
                                .is_some_and(|c| c.is_alphabetic()) =>
                        {
                            self.advance();
                            let c2 = self.advance().unwrap();
                            ident.push('^');
                            ident.push(c2);
                        }
                        Some(c) if "!@$&*+;',\"\\|?/<>.0123456789~%-=()[]{}".contains(c) => {
                            self.advance();
                            ident.push(c);
                        }
                        _ => {}
                    }
                }
                // `$_<`, `$_<<`, … — outer topic chain (stryke extension). Also
                // applies to positional slots: `$_0<<<<<`, `$_1<<<<<`, etc. The
                // canonical scope key is `_<<<<<` (slot 0) or `_N<<<<<` (slot N).
                //
                // Indexed-ascent shortcut: `$_<N` ≡ `$_<<<...<` (N chevrons),
                // with N a positive integer. `$_<3` is much more readable than
                // `$_<<<` past depth 2. The lexer synthesizes the chevron form
                // so the rest of the system (scope keys, parse) is unchanged.
                let is_topic_slot = ident == "_"
                    || (ident.len() > 1
                        && ident.starts_with('_')
                        && ident[1..].bytes().all(|b| b.is_ascii_digit()));
                if is_topic_slot {
                    let mut lts = String::new();
                    while self.peek() == Some('<') {
                        self.advance();
                        lts.push('<');
                    }
                    // Indexed-ascent: after a single `<`, if the next chars are
                    // digits NOT followed by `>` or `:` (which would make it a
                    // string slice like `$_<1:5>`), expand `<N` to N chevrons.
                    if lts.len() == 1 && self.peek().is_some_and(|c| c.is_ascii_digit()) {
                        let mut peek_off = 0usize;
                        while self.peek_at(peek_off).is_some_and(|c| c.is_ascii_digit()) {
                            peek_off += 1;
                        }
                        let trailing = self.peek_at(peek_off);
                        let is_slice = matches!(trailing, Some(':') | Some('>'));
                        if !is_slice {
                            let mut digits = String::new();
                            for _ in 0..peek_off {
                                if let Some(c) = self.advance() {
                                    digits.push(c);
                                }
                            }
                            if let Ok(n) = digits.parse::<usize>() {
                                if n >= 1 {
                                    // Replace the single `<` already collected
                                    // with N chevrons (we already consumed 1).
                                    for _ in 1..n {
                                        lts.push('<');
                                    }
                                }
                            }
                        }
                    }
                    if !lts.is_empty() {
                        return format!("{}{}", ident, lts);
                    }
                }
                ident
            }
            Some('^') => {
                self.advance();
                // Perl `$^I`, `$^O`, … — caret plus one letter (or `^` alone).
                if self.peek().is_some_and(|c| c.is_alphabetic()) {
                    let c2 = self.advance().unwrap();
                    format!("^{}", c2)
                } else {
                    "^".to_string()
                }
            }
            // `${name}` — must run before the punctuation branch (`{` is also listed there).
            Some('{') => {
                self.advance(); // {
                let name = self.read_while(|c| c != '}');
                if self.peek() == Some('}') {
                    self.advance();
                }
                name
            }
            // Perl `$#name` — last index of `@name` (scalar name stored as `#name`).
            Some('#') => {
                self.advance();
                if self.peek().is_some_and(|c| c.is_alphabetic() || c == '_') {
                    let mut name = String::from("#");
                    name.push_str(&self.read_package_qualified_identifier());
                    name
                } else {
                    "#".to_string()
                }
            }
            Some(c) if "!@$&*+;',\"\\|?/<>.0123456789~%-=()[]{}".contains(c) => {
                self.advance();
                c.to_string()
            }
            _ => String::new(),
        }
    }

    /// `${$name}` / `${$Foo::bar}` — when the braced body is a plain scalar `$identifier`, Perl treats it
    /// like `$$name` (scalar deref). The naive lexer otherwise yields a bogus [`Token::ScalarVar`] name
    /// containing a leading `$` (e.g. Try::Tiny's `${$code_ref}`).
    fn braced_body_symbolic_scalar_deref_name(body: &str) -> Option<&str> {
        let body = body.trim();
        let rest = body.strip_prefix('$')?;
        if rest.is_empty() {
            return None;
        }
        let mut chars = rest.chars();
        let c0 = chars.next()?;
        if !(c0.is_alphabetic() || c0 == '_') {
            return None;
        }
        for c in chars {
            if !(c.is_alphanumeric() || c == '_' || c == ':') {
                return None;
            }
        }
        Some(rest)
    }
    /// `next_token` — see implementation.
    pub fn next_token(&mut self) -> StrykeResult<Token> {
        self.skip_whitespace_and_comments();
        // Stamp the start line for [`Self::tokenize`]. Recursive calls
        // through POD / heredoc skip will overwrite this with the post-
        // skip line, so the emitted token always reports the source line
        // it actually lives on.
        self.token_start_line = self.line;

        if self.pos >= self.input.len() {
            return Ok(Token::Eof);
        }

        // `last_was_arrow` is consumed at most once per token: the s/tr/y/q/qq
        // /qw/qr/m guards check whether the IMMEDIATELY previous token was
        // `->`. Reset here; the Arrow / ArrowBrace return paths re-arm it
        // for the next `next_token` call. We snapshot before the reset so
        // identifier-decoding logic below can read the previous-token state
        // via `self.prev_arrow` (set up via a one-shot field swap).
        self.prev_arrow = self.last_was_arrow;
        self.last_was_arrow = false;
        self.last_was_bare_positional = false;

        let ch = self.input[self.pos];
        match ch {
            // Variables
            '$' => {
                self.advance();
                // `$$foo` — symbolic scalar deref (Perl `${$foo}`-style lookup)
                if self.peek() == Some('$') {
                    // `$$_{` — Perl parses as `$_->{...}` (implicit arrow on `$_`), not `$$` PID + `_`.
                    let is_dollar_under_brace = self.input.get(self.pos + 1) == Some(&'_')
                        && self.input.get(self.pos + 2) == Some(&'{');
                    if !is_dollar_under_brace {
                        self.advance();
                        if self.peek().is_some_and(|c| c.is_alphabetic() || c == '_') {
                            let name = self.read_identifier();
                            self.last_was_term = true;
                            return Ok(Token::DerefScalarVar(name));
                        }
                        // `$$` — process id (Perl `$$`)
                        self.last_was_term = true;
                        return Ok(Token::ScalarVar("$$".to_string()));
                    }
                }
                let name = self.read_variable_name();
                if name.is_empty() {
                    return Err(self.syntax_err("Expected variable name after $", self.line));
                }
                // `--no-interop`: reject `$a` / `$b` (Perl's reduce/sort/pair*
                // comparator-bind globals). Stryke's runtime also binds `$_0`
                // / `$_1` for the same positions; in idiomatic-only mode users
                // must use those instead.
                if crate::no_interop_mode() && (name == "a" || name == "b") {
                    return Err(self.syntax_err(
                        format!(
                            "stryke uses `_` / `_1` (bareword in code) or `$_` / `$_1` \
                             (sigil inside string interpolation / when whitespace \
                             would change parsing) instead of `${}` (--no-interop is active)",
                            name
                        ),
                        self.line,
                    ));
                }
                self.last_was_term = true;
                if let Some(tail) = Self::braced_body_symbolic_scalar_deref_name(&name) {
                    return Ok(Token::DerefScalarVar(tail.to_string()));
                }
                Ok(Token::ScalarVar(name))
            }
            '@' => {
                self.advance();
                if self.peek() == Some('-') {
                    self.advance();
                    self.last_was_term = true;
                    return Ok(Token::ArrayVar("-".to_string()));
                }
                if self.peek() == Some('+') {
                    self.advance();
                    self.last_was_term = true;
                    return Ok(Token::ArrayVar("+".to_string()));
                }
                if self.peek() == Some('^')
                    && self
                        .input
                        .get(self.pos + 1)
                        .is_some_and(|c| c.is_alphabetic() || *c == '_')
                {
                    self.advance();
                    let name = format!("^{}", self.read_package_qualified_identifier());
                    self.last_was_term = true;
                    return Ok(Token::ArrayVar(name));
                }
                if self.peek() == Some('_') || self.peek().is_some_and(|c| c.is_alphabetic()) {
                    let name = self.read_package_qualified_identifier();
                    self.last_was_term = true;
                    return Ok(Token::ArrayVar(name));
                }
                self.last_was_term = false;
                Ok(Token::ArrayAt)
            }
            '%' if !self.last_was_term => {
                self.advance();
                // `%+` — named regex captures (Perl special hash)
                if self.peek() == Some('+') {
                    self.advance();
                    self.last_was_term = true;
                    return Ok(Token::HashVar("+".to_string()));
                }
                if self.peek() == Some('^')
                    && self
                        .input
                        .get(self.pos + 1)
                        .is_some_and(|c| c.is_alphabetic() || *c == '_')
                {
                    self.advance();
                    let name = format!("^{}", self.read_package_qualified_identifier());
                    self.last_was_term = true;
                    return Ok(Token::HashVar(name));
                }
                if self.peek().is_some_and(|c| c.is_alphabetic() || c == '_') {
                    let name = self.read_package_qualified_identifier();
                    self.last_was_term = true;
                    return Ok(Token::HashVar(name));
                }
                self.last_was_term = false;
                Ok(Token::HashPercent)
            }

            // Numbers
            '0'..='9' => {
                let tok = self.read_number()?;
                self.last_was_term = true;
                Ok(tok)
            }

            // Strings
            '\'' => {
                let tok = self.read_single_quoted_string()?;
                self.last_was_term = true;
                Ok(tok)
            }
            '"' => {
                let tok = self.read_double_quoted_string()?;
                self.last_was_term = true;
                Ok(tok)
            }

            // Backtick — Perl `` `cmd` `` (qx), not a plain double-quoted string
            '`' => {
                self.advance();
                let cmd = self.read_interpolating_until('`')?;
                self.last_was_term = true;
                Ok(Token::BacktickString(cmd))
            }

            // Regex or division
            '/' => {
                if !self.last_was_term {
                    let tok = self.read_regex()?;
                    self.last_was_term = true;
                    return Ok(tok);
                }
                self.advance();
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::DivAssign);
                }
                if self.peek() == Some('/') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::DefinedOrAssign);
                    }
                    self.last_was_term = false;
                    return Ok(Token::DefinedOr);
                }
                self.last_was_term = false;
                Ok(Token::Slash)
            }

            // Operators and punctuation
            '+' => {
                self.advance();
                if self.peek() == Some('+') {
                    self.advance();
                    // Whether it was term depends on context
                    return Ok(Token::Increment);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::PlusAssign);
                }
                self.last_was_term = false;
                Ok(Token::Plus)
            }
            '-' => {
                self.advance();
                // File test operators: -e, -f, -d, etc.
                if !self.last_was_term {
                    if let Some(c) = self.peek() {
                        if "efdlpSszrwxoRWXOBCTMAgut".contains(c)
                            && self.peek_at(1).is_none_or(|n| {
                                n.is_whitespace()
                                    || n == '$'
                                    || n == '\''
                                    || n == '"'
                                    || n == '('
                                    || n == ')'
                                    || n == '}'
                                    || n == ';'
                                    || n == ','
                            })
                        {
                            self.advance();
                            self.last_was_term = false;
                            return Ok(Token::FileTest(c));
                        }
                    }
                }
                if self.peek() == Some('-') {
                    self.advance();
                    return Ok(Token::Decrement);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::MinusAssign);
                }
                if self.peek() == Some('>') {
                    self.advance();
                    if self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::ThreadArrowLast);
                    }
                    self.last_was_term = false;
                    // Arm the arrow flag so the next identifier (e.g. `y`,
                    // `s`, `tr`, `m`, `q…`) decodes as a method name, not
                    // a substitution / transliteration / quote-like body.
                    self.last_was_arrow = true;
                    return Ok(Token::Arrow);
                }
                self.last_was_term = false;
                Ok(Token::Minus)
            }
            '*' => {
                self.advance();
                if self.peek() == Some('*') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::PowAssign);
                    }
                    self.last_was_term = false;
                    return Ok(Token::Power);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::MulAssign);
                }
                self.last_was_term = false;
                Ok(Token::Star)
            }
            '%' => {
                // Only reached when last_was_term is true (hash sigil handled above)
                self.advance();
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::ModAssign);
                }
                self.last_was_term = false;
                Ok(Token::Percent)
            }
            '.' => {
                self.advance();
                if self.peek() == Some('.') {
                    self.advance();
                    if self.peek() == Some('.') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::RangeExclusive);
                    }
                    self.last_was_term = false;
                    return Ok(Token::Range);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::DotAssign);
                }
                self.last_was_term = false;
                Ok(Token::Dot)
            }
            '=' => {
                let eq_pos = self.pos;
                self.advance();
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::NumEq);
                }
                if self.peek() == Some('~') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::BindMatch);
                }
                if self.peek() == Some('>') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::FatArrow);
                }
                // POD: =head1 etc — only when `=` begins the line (after optional whitespace).
                if self.peek().is_some_and(|c| c.is_alphabetic())
                    && self.at_line_start_for_pod(eq_pos)
                {
                    // Skip POD
                    loop {
                        let line = self.read_while(|c| c != '\n');
                        if self.peek() == Some('\n') {
                            self.advance();
                        }
                        if line.starts_with("=cut") || self.pos >= self.input.len() {
                            break;
                        }
                    }
                    return self.next_token();
                }
                self.last_was_term = false;
                Ok(Token::Assign)
            }
            '!' => {
                self.advance();
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::NumNe);
                }
                if self.peek() == Some('~') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::BindNotMatch);
                }
                self.last_was_term = false;
                Ok(Token::LogNot)
            }
            '<' => {
                self.advance();
                let after_lt = self.pos;
                // Readline `<$fh>` (scalar handle) — must come before `<IDENT>` / numeric `<`.
                if self.peek() == Some('$') {
                    self.advance();
                    let name = self.read_variable_name();
                    if !name.is_empty() && self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = true;
                        return Ok(Token::ReadLine(name));
                    }
                    self.pos = after_lt;
                }
                // Diamond operator <> or <STDIN>
                if self.peek() == Some('>') {
                    self.advance();
                    self.last_was_term = true;
                    return Ok(Token::Diamond);
                }
                if self.peek().is_some_and(|c| c.is_uppercase()) {
                    let name = self.read_identifier();
                    if self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = true;
                        return Ok(Token::ReadLine(name));
                    }
                    // Not a readline, put back — this is tricky, we'll handle as less-than
                    // followed by ident. For simplicity, return the ident separately.
                    self.last_was_term = false;
                    return Ok(Token::NumLt);
                }
                // `<main::STDIN>` — `main` is the default package for
                // special filehandles. Accept the qualified form and
                // strip the prefix so the runtime sees `STDIN`.
                if self.peek() == Some('m')
                    && self.input.get(after_lt + 1) == Some(&'a')
                    && self.input.get(after_lt + 2) == Some(&'i')
                    && self.input.get(after_lt + 3) == Some(&'n')
                    && self.input.get(after_lt + 4) == Some(&':')
                    && self.input.get(after_lt + 5) == Some(&':')
                    && self
                        .input
                        .get(after_lt + 6)
                        .is_some_and(|c| c.is_uppercase())
                {
                    // Consume `main::`.
                    for _ in 0..6 {
                        self.advance();
                    }
                    let name = self.read_identifier();
                    if self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = true;
                        return Ok(Token::ReadLine(name));
                    }
                    // Restore and fall through if it wasn't actually a readline.
                    self.pos = after_lt;
                }
                if self.peek() == Some('=') {
                    self.advance();
                    if self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::Spaceship);
                    }
                    self.last_was_term = false;
                    return Ok(Token::NumLe);
                }
                if self.peek() == Some('<') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::ShiftLeftAssign);
                    }
                    // `<<` — binary shift after a complete term (`1 << 4`, `"x" << 2`); heredoc when a
                    // term is expected (`print <<EOF`, `my $x = <<EOF`, after `.` / `,` / `(` …).
                    //
                    // `}` always sets `last_was_term=true`, but a `}` ending
                    // a block / fn body followed by a newline and `<<TAG` on
                    // the next line is unambiguously heredoc, not shift —
                    // `block << bareword` is meaningless. Disambiguate by
                    // peeking after `<<`: if the next char looks like the
                    // start of a heredoc tag (uppercase, `_`, `~`, `"`,
                    // `'`), prefer heredoc even when last_was_term is set.
                    // Numeric / sigil / lowercase still falls through to
                    // ShiftLeft so `1 << 4` and `$x << $shift` still work.
                    let looks_like_heredoc_tag =
                        matches!(self.peek(), Some('~') | Some('"') | Some('\'') | Some('_'),)
                            || self.peek().is_some_and(|c| c.is_ascii_uppercase());
                    if self.last_was_term && !looks_like_heredoc_tag {
                        self.last_was_term = false;
                        return Ok(Token::ShiftLeft);
                    }
                    let (tag, interpolate, indented) = self.read_heredoc_tag()?;
                    let body = self.read_heredoc_body(&tag, indented)?;
                    self.last_was_term = true;
                    return Ok(Token::HereDoc(tag, body, interpolate));
                }
                self.last_was_term = false;
                Ok(Token::NumLt)
            }
            '>' => {
                self.advance();
                if self.peek() == Some('{') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::ArrowBrace);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::NumGe);
                }
                if self.peek() == Some('>') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::ShiftRightAssign);
                    }
                    self.last_was_term = false;
                    return Ok(Token::ShiftRight);
                }
                self.last_was_term = false;
                Ok(Token::NumGt)
            }
            '&' => {
                self.advance();
                if self.peek() == Some('&') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::AndAssign);
                    }
                    self.last_was_term = false;
                    return Ok(Token::LogAnd);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::BitAndAssign);
                }
                self.last_was_term = false;
                Ok(Token::BitAnd)
            }
            '|' => {
                self.advance();
                if self.peek() == Some('|') {
                    self.advance();
                    if self.peek() == Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::OrAssign);
                    }
                    self.last_was_term = false;
                    return Ok(Token::LogOr);
                }
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::BitOrAssign);
                }
                if self.peek() == Some('>') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::PipeForward);
                }
                self.last_was_term = false;
                Ok(Token::BitOr)
            }
            '^' => {
                self.advance();
                if self.peek() == Some('=') {
                    self.advance();
                    self.last_was_term = false;
                    return Ok(Token::XorAssign);
                }
                self.last_was_term = false;
                Ok(Token::BitXor)
            }
            '~' => {
                self.advance();
                if self.peek() == Some('>') {
                    self.advance();
                    if self.peek() == Some('>') {
                        self.advance();
                        self.last_was_term = false;
                        return Ok(Token::ThreadArrowLast);
                    }
                    self.last_was_term = false;
                    return Ok(Token::ThreadArrow);
                }
                // `~s>` (streaming thread-first) / `~s>>` (streaming thread-last)
                // — per-item streaming thread-macros that lower to
                // `par_pipeline_streaming`: each stage runs in its own worker
                // connected by bounded channels, items flow one-at-a-time.
                if self.peek() == Some('s') && self.peek_at(1) == Some('>') {
                    self.advance(); // consume 's'
                    self.advance(); // consume first '>'
                    if self.peek() == Some('>') {
                        self.advance(); // consume second '>'
                        self.last_was_term = false;
                        return Ok(Token::ThreadArrowStreamLast);
                    }
                    self.last_was_term = false;
                    return Ok(Token::ThreadArrowStream);
                }
                // `~p>` (parallel-chunk thread-first) / `~p>>` (thread-last)
                // — sugar for `par_reduce { stage1 |> stage2 |> ... } SOURCE`.
                // `||>` or `|then|` mid-pipeline switches back to a normal
                // `~>` continuation operating on the merged result.
                if self.peek() == Some('p') && self.peek_at(1) == Some('>') {
                    self.advance(); // consume 'p'
                    self.advance(); // consume first '>'
                    if self.peek() == Some('>') {
                        self.advance(); // consume second '>'
                        self.last_was_term = false;
                        return Ok(Token::ThreadArrowParLast);
                    }
                    self.last_was_term = false;
                    return Ok(Token::ThreadArrowPar);
                }
                // `~d>` (distributed thread-first) / `~d>>` (thread-last) —
                // mirrors `~p>` but each chunk is shipped to a remote worker
                // on a cluster. Syntax: `~d> on $cluster SOURCE stages...`.
                if self.peek() == Some('d') && self.peek_at(1) == Some('>') {
                    self.advance(); // consume 'd'
                    self.advance(); // consume first '>'
                    if self.peek() == Some('>') {
                        self.advance(); // consume second '>'
                        self.last_was_term = false;
                        return Ok(Token::ThreadArrowDistLast);
                    }
                    self.last_was_term = false;
                    return Ok(Token::ThreadArrowDist);
                }
                self.last_was_term = false;
                Ok(Token::BitNot)
            }
            '?' => {
                self.advance();
                // `??` / `??=` — stryke aliases for `//` / `//=` (C#/Swift
                // null-coalescing spelling). Literal under --compat. No
                // clash with ternary: two adjacent `?` are never valid there.
                if self.peek() == Some('?') && !crate::compat_mode() {
                    self.advance();
                    self.last_was_term = false;
                    if self.peek() == Some('=') {
                        self.advance();
                        return Ok(Token::DefinedOrAssign);
                    }
                    return Ok(Token::DefinedOr);
                }
                self.last_was_term = false;
                Ok(Token::Question)
            }
            ':' => {
                self.advance();
                if self.peek() == Some(':') {
                    self.advance();
                    // IPv6 zero-compressed prefix: `::1`, `::ffff:c000:280`.
                    // Only fires in term position (where `Pkg::ident` is
                    // impossible) and only when the chars after `::` form a
                    // valid IPv6 by Rust's parser. Skip when the `::` lives
                    // inside `[…]` — that's array-slice step syntax
                    // (`@a[::2]`, `@a[::-1]`), not an address.
                    let in_bracket_subscript =
                        self.input.get(self.pos.saturating_sub(3)).copied() == Some('[');
                    if !self.last_was_term && !in_bracket_subscript {
                        let saved = self.pos - 2;
                        if let Some(consumed) = self.try_consume_ipv6_tail(saved) {
                            self.last_was_term = true;
                            return Ok(Token::DoubleString(consumed));
                        }
                    }
                    self.last_was_term = false;
                    return Ok(Token::PackageSep);
                }
                self.last_was_term = false;
                Ok(Token::Colon)
            }
            '\\' => {
                self.advance();
                // Backslash-newline: line continuation (shell-style)
                // Don't increment line — continued line is logically part of the same line
                if self.peek() == Some('\n') {
                    self.pos += 1; // skip newline without incrementing self.line
                    return self.next_token();
                }
                self.last_was_term = false;
                Ok(Token::Backslash)
            }
            ',' => {
                self.advance();
                self.last_was_term = false;
                Ok(Token::Comma)
            }
            ';' => {
                self.advance();
                self.last_was_term = false;
                Ok(Token::Semicolon)
            }
            '(' => {
                self.advance();
                self.last_was_term = false;
                Ok(Token::LParen)
            }
            ')' => {
                self.advance();
                self.last_was_term = true;
                Ok(Token::RParen)
            }
            '[' => {
                self.advance();
                self.last_was_term = false;
                Ok(Token::LBracket)
            }
            ']' => {
                self.advance();
                self.last_was_term = true;
                Ok(Token::RBracket)
            }
            '{' => {
                self.advance();
                self.last_was_term = false;
                Ok(Token::LBrace)
            }
            '}' => {
                self.advance();
                self.last_was_term = true;
                Ok(Token::RBrace)
            }

            // Identifiers and keywords
            c if c.is_alphabetic() || c == '_' => {
                let ident_start = self.pos;
                let mut ident = self.read_identifier();

                // IPv6 lookahead for hex-letter prefixes: `fe80::1`, `abcd::ff`,
                // `dead:beef::1`, etc. Only fires when the just-consumed
                // identifier is a valid 1..=4 hex-digit group (i.e. could be
                // an IPv6 segment) AND the next char is `:`. Speculatively
                // greedily consumes hex / `:` / `::` and asks Rust's
                // `Ipv6Addr` parser to validate; on failure restores `pos`
                // so the identifier-as-bareword path runs unchanged.
                //
                // Skip when we're already in the middle of a package-qualified
                // path: `package A::B::C` lexes "A" then "::", and at that
                // point the next ident "B" must NOT be IPv6-trapped — `B::C`
                // is the rest of the package name, not the address
                // `0:0:0:0:0:0:B:C`. Same rule for `Foo::Bar::baz` mid-stream
                // with hex letters that happen to look like an address.
                let after_package_sep = ident_start >= 2
                    && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                    && self.input.get(ident_start.saturating_sub(1)) == Some(&':');
                if !after_package_sep
                    && self.peek() == Some(':')
                    && ident.len() <= 4
                    && ident.chars().all(|ch| ch.is_ascii_hexdigit())
                {
                    if let Some(consumed) = self.try_consume_ipv6_tail(ident_start) {
                        self.last_was_term = true;
                        return Ok(Token::DoubleString(consumed));
                    }
                }

                // Outer-topic chain in bare form: `_<<<<` (slot 0) and
                // `_N<<<<` (slot N). Greedy consume `<` chevrons immediately
                // following `_` or `_<digits>`. This is what makes
                // `_<` ≡ `$_<` ≡ `_0<` ≡ `$_0<` work without a sigil.
                // Stryke power-user note: `_ < 5` (with whitespace) still
                // tokenizes as topic-then-less-than; only `_<` with no
                // intervening space becomes a topic-slot identifier.
                let is_topic_slot = ident == "_"
                    || (ident.len() > 1
                        && ident.starts_with('_')
                        && ident[1..].bytes().all(|b| b.is_ascii_digit()));
                if is_topic_slot {
                    // Greedy `<` chevrons for the outer-topic chain, BUT only
                    // when the chevron run isn't followed by a slice index.
                    // `_<1:5>` is a string slice; `_<<<<<` is the 5-deep
                    // outer-topic. Disambiguate by peeking past the run: if
                    // the first non-`<` char is a digit, `-`, `:`, or `>`,
                    // we're in a slice — bail out and let the parser handle
                    // `<...>` as postfix subscript.
                    //
                    // Indexed-ascent shortcut: `_<N` ≡ `_<<<...<` (N chevrons)
                    // when N is digits NOT followed by `>` or `:`. So `_<3` is
                    // a depth-3 reference (more readable than `_<<<`), while
                    // `_<3>` and `_<3:5>` remain string slices.
                    let mut peek_off = 0usize;
                    while self.peek_at(peek_off) == Some('<') {
                        peek_off += 1;
                    }
                    let trailing = self.peek_at(peek_off);
                    // Single `<` followed by digits: try indexed-ascent first.
                    // Only triggers for one-chevron runs because `_<<3` would
                    // mean "depth 2 of position 3" (which is not how the
                    // grammar works) — we only allow `_<digits` at depth 1.
                    let mut indexed_ascent: Option<usize> = None;
                    if peek_off == 1 && trailing.is_some_and(|c: char| c.is_ascii_digit()) {
                        let mut off = 1usize;
                        while self.peek_at(off).is_some_and(|c| c.is_ascii_digit()) {
                            off += 1;
                        }
                        let after_digits = self.peek_at(off);
                        let still_a_slice = matches!(after_digits, Some(':') | Some('>'));
                        if !still_a_slice {
                            // Parse the digit run.
                            let mut digits = String::new();
                            for k in 1..off {
                                if let Some(c) = self.peek_at(k) {
                                    digits.push(c);
                                }
                            }
                            if let Ok(n) = digits.parse::<usize>() {
                                if n >= 1 {
                                    indexed_ascent = Some(n);
                                    // Consume `<` + the digits.
                                    self.advance();
                                    for _ in 1..off {
                                        self.advance();
                                    }
                                }
                            }
                        }
                    }
                    if let Some(n) = indexed_ascent {
                        for _ in 0..n {
                            ident.push('<');
                        }
                    } else {
                        let is_slice = peek_off > 0
                            && matches!(trailing, Some(c) if c.is_ascii_digit() || c == '-' || c == ':' || c == '>');
                        if !is_slice {
                            for _ in 0..peek_off {
                                self.advance();
                                ident.push('<');
                            }
                        }
                    }
                    // `_N` (underscore + digits, ≥ 1 digit) is a reserved
                    // positional-alias name — never a function name. Emit
                    // ScalarVar directly so bareword `_1`, `_2`, ... in
                    // expression position resolves to the scalar slot
                    // instead of being looked up as a sub call. Bare `_`
                    // alone (without digits) keeps Ident shape so the
                    // existing topic/bareword machinery still runs.
                    if ident.len() > 1
                        && ident.starts_with('_')
                        && ident.chars().nth(1).is_some_and(|c| c.is_ascii_digit())
                    {
                        self.last_was_term = true;
                        self.last_was_bare_positional = true;
                        return Ok(Token::ScalarVar(ident));
                    }
                    // Also reserve bare `_<+` (chevron-only topic ascent on
                    // slot 0) — these are never sub names.
                    if ident.starts_with('_') && ident.contains('<') {
                        self.last_was_term = true;
                        self.last_was_bare_positional = true;
                        return Ok(Token::ScalarVar(ident));
                    }
                }

                // Special multi-char constructs
                match ident.as_str() {
                    "format" => {
                        // `$obj->format` — method call, not format declaration.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `Foo::format` — namespaced identifier tail.
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // Hash-key bareword contexts: `$h{format}`, `{format => ...}`,
                        // `$h{format,...}`. The char immediately before `format`
                        // (modulo whitespace) being `{` means we're inside a
                        // hash-subscript or hash-literal — `format` is the key,
                        // not the FORMAT keyword.
                        {
                            let mut p = ident_start;
                            while p > 0 {
                                match self.input.get(p - 1) {
                                    Some(&' ') | Some(&'\t') => p -= 1,
                                    _ => break,
                                }
                            }
                            if p > 0 && self.input.get(p - 1) == Some(&'{') {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        // Lookahead-disambiguation: shapes that prove
                        // `format` is a bareword, not a declaration keyword.
                        // `}` / `,` / `;` / `)` / `]` — list/hash/expr context.
                        // `=>` — autoquoted hash key.
                        // `(` — function call (format() / format($x)).
                        // EOF — bare ident at end of input.
                        // A real `format NAME = ... .` declaration always has
                        // an identifier between `format` and `=`; if no
                        // identifier is found at the expected slot, it's
                        // a bareword.
                        {
                            let saved_pos = self.pos;
                            let saved_line = self.line;
                            self.skip_whitespace_only();
                            let bare = match self.peek() {
                                None => true,
                                Some(',' | ';' | ')' | ']' | '}' | '(') => true,
                                Some('=') if self.peek_at(1) == Some('>') => true,
                                Some(c) if !c.is_alphabetic() && c != '_' => true,
                                _ => false,
                            };
                            self.pos = saved_pos;
                            self.line = saved_line;
                            if bare {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        self.skip_whitespace_and_comments();
                        let fname = self.read_package_qualified_identifier();
                        self.skip_whitespace_and_comments();
                        if self.peek() != Some('=') {
                            return Err(
                                self.syntax_err("Expected '=' after format name", self.line)
                            );
                        }
                        self.advance();
                        let lines = self.read_format_body()?;
                        self.last_was_term = false;
                        return Ok(Token::FormatDecl { name: fname, lines });
                    }
                    "r" if self.peek() == Some('"')
                        && self.peek_at(1) == Some('"')
                        && self.peek_at(2) == Some('"') =>
                    {
                        // `r"""..."""` — raw triple-quoted string. No
                        // interpolation, no backslash escapes; every byte
                        // is copied verbatim until the closing `"""`.
                        // Only triggers when `r` is followed IMMEDIATELY
                        // by three quotes — `r 5`, `r->foo`, `r(...)` etc.
                        // still hit the generic identifier path below.
                        self.advance(); // 1st "
                        self.advance(); // 2nd "
                        self.advance(); // 3rd "
                        let s = self.read_triple_quoted_body(false)?;
                        self.last_was_term = true;
                        return Ok(Token::SingleString(s));
                    }
                    "qw" => {
                        // After `->`, `qw` is a method name, not a quote-word list.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `qw` followed by `=>` is an autoquoted hash key, not qw().
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(c) = self.peek() {
                            if c == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if matches!(c, ';' | ',' | ')' | ']' | '}' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        self.pos = start_pos; // restore for read_qw
                        self.line = start_line;
                        let tok = self.read_qw()?;
                        self.last_was_term = true;
                        return Ok(tok);
                    }
                    "qq" | "q" => {
                        // After `->`, `q` / `qq` are method names, not quote operators.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // After `::`, treat as namespaced identifier (`Foo::q`, `Foo::qq`).
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `q` / `qq` followed by `=>` is an autoquoted hash key, not a quote operator.
                        // Also treat as identifier if followed by terminators like `;`, `,`, `)`, etc.
                        // Must check AFTER skipping whitespace to handle `q => 5`.
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(c) = self.peek() {
                            // `=` followed by `>` is fat comma — `q` is a bareword key
                            if c == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos; // restore position
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            // Other terminators: `q` is an identifier
                            if matches!(c, ';' | ',' | ')' | ']' | '}' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        let delim = self.advance().ok_or_else(|| {
                            self.syntax_err("Expected delimiter after q/qq", self.line)
                        })?;
                        let close = match delim {
                            '(' => ')',
                            '[' => ']',
                            '{' => '}',
                            '<' => '>',
                            c => c,
                        };
                        let s = if matches!(delim, '(' | '[' | '{' | '<') {
                            self.read_q_qq_balanced_body(delim, close, ident == "qq")?
                        } else if ident == "qq" {
                            self.read_interpolating_until(close)?
                        } else {
                            self.read_escaped_until(close)?
                        };
                        self.last_was_term = true;
                        if ident == "qq" {
                            return Ok(Token::DoubleString(s));
                        }
                        return Ok(Token::SingleString(s));
                    }
                    "qx" => {
                        // After `->`, `qx` is a method name, not a backtick command.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // After `::`, treat as namespaced identifier (`Foo::qx`).
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `qx` followed by `=>` is an autoquoted hash key.
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(c) = self.peek() {
                            if c == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if matches!(c, ';' | ',' | ')' | ']' | '}' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        let delim = self.advance().ok_or_else(|| {
                            self.syntax_err("Expected delimiter after qx", self.line)
                        })?;
                        let close = match delim {
                            '(' => ')',
                            '[' => ']',
                            '{' => '}',
                            '<' => '>',
                            c => c,
                        };
                        let s = self.read_interpolating_until(close)?;
                        self.last_was_term = true;
                        return Ok(Token::BacktickString(s));
                    }
                    "qr" => {
                        // After `->`, `qr` is a method name, not a quoted regex.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // After `::`, treat as namespaced identifier (`Foo::qr`).
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `qr` followed by `=>` is an autoquoted hash key.
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(c) = self.peek() {
                            if c == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if matches!(c, ';' | ',' | ')' | ']' | '}' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        let delim = self.advance().ok_or_else(|| {
                            self.syntax_err("Expected delimiter after qr", self.line)
                        })?;
                        let close = match delim {
                            '(' => ')',
                            '[' => ']',
                            '{' => '}',
                            '<' => '>',
                            c => c,
                        };
                        // Regex pattern: preserve backslash escapes raw so the
                        // regex engine sees `\$`, `\@`, `\d`, etc. as written.
                        // Do NOT route through `read_escaped_until` — that's
                        // for double-quoted strings and rewrites `\$` to a
                        // private-use sentinel that the regex compiler can't
                        // decode (would silently strip the `$`).
                        let mut pattern = String::new();
                        loop {
                            match self.advance() {
                                Some('\\') => {
                                    pattern.push('\\');
                                    if let Some(c) = self.advance() {
                                        pattern.push(c);
                                    }
                                }
                                Some(c) if c == close => break,
                                Some(c) => pattern.push(c),
                                None => {
                                    return Err(self.syntax_err("Unterminated qr regex", self.line))
                                }
                            }
                        }
                        let flags = self.read_while(|c| REGEX_FLAG_CHARS.contains(c));
                        self.last_was_term = true;
                        return Ok(Token::Regex(pattern, flags, delim));
                    }
                    "m" => {
                        // After `->`, `m` is a method name, not a regex match.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `Foo::m` — after `::`, `m` is the tail of a namespaced
                        // identifier, never a regex-match operator.
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `m` followed by terminators is a bareword, not match operator.
                        // Must check AFTER skipping whitespace to handle `m => "val"`.
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(d) = self.peek() {
                            if d == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if matches!(d, ';' | ',' | ')' | ']' | '}' | '>' | ':' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        self.pos = start_pos;
                        self.line = start_line;
                        // m/pattern/flags — try parsing as regex, but backtrack if
                        // unterminated (handles thread stages where `/m/` is a grep filter)
                        if self.suppress_m_regex == 0 {
                            if let Some(delim) = self.peek() {
                                if !delim.is_alphanumeric() && delim != '_' {
                                    // Save state for backtracking
                                    let saved_pos = self.pos;
                                    let saved_line = self.line;
                                    self.advance(); // consume delimiter
                                    let close = match delim {
                                        '(' => ')',
                                        '[' => ']',
                                        '{' => '}',
                                        '<' => '>',
                                        c => c,
                                    };
                                    let mut pattern = String::new();
                                    let mut terminated = true;
                                    loop {
                                        match self.advance() {
                                            Some('\\') => {
                                                pattern.push('\\');
                                                if let Some(c) = self.advance() {
                                                    pattern.push(c);
                                                }
                                            }
                                            Some(c) if c == close => break,
                                            Some(c) if c == '\n' && close == '/' => {
                                                // Newline before closing / — not a valid m//
                                                terminated = false;
                                                break;
                                            }
                                            Some(c) => pattern.push(c),
                                            None => {
                                                return Err(self.syntax_err(
                                                    "Search pattern not terminated",
                                                    saved_line,
                                                ));
                                            }
                                        }
                                    }
                                    if terminated {
                                        let flags =
                                            self.read_while(|c| REGEX_FLAG_CHARS.contains(c));
                                        self.last_was_term = true;
                                        return Ok(Token::Regex(pattern, flags, delim));
                                    }
                                    // Newline before closing / — backtrack and treat `m` as identifier
                                    self.pos = saved_pos;
                                    self.line = saved_line;
                                }
                            }
                        }
                        // Just the identifier 'm'
                        self.last_was_term = true;
                        return Ok(Token::Ident(ident));
                    }
                    "s" => {
                        // `$obj->s` / `$obj->s(...)` — after `->`, `s` is a method name.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `Foo::s` / `Foo::Bar::s` — after `::`, `s` is the tail
                        // segment of a namespaced identifier, never substitution.
                        // Same check shape as the IPv6 / `after_package_sep` guard
                        // up at line ~2158: previous two chars are `::`.
                        if ident_start >= 2
                            && self.input.get(ident_start.saturating_sub(2)) == Some(&':')
                            && self.input.get(ident_start.saturating_sub(1)) == Some(&':')
                        {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // `s` followed by terminators is a bareword, not substitution.
                        // Must check AFTER skipping whitespace to handle `s => "val"`.
                        // `,` is treated as a terminator UNLESS the lookahead shows the
                        // full `s,PAT,REPL,FLAGS` shape (≥ 2 more commas before the
                        // statement ends) — that gates the comma-delim case to genuine
                        // substitutions like `perl -pe 's,\bt\b,b,g'` while leaving
                        // bareword `s` alone in struct fields, list literals, and
                        // function args.
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(d) = self.peek() {
                            if d == '=' && self.peek_at(1) == Some('>') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if matches!(d, ';' | ')' | ']' | '}' | '>' | ':' | '\n') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if d == ',' && !self.lookahead_is_comma_delim_subst() {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        self.pos = start_pos;
                        self.line = start_line;
                        // s/pattern/replacement/flags
                        if let Some(delim) = self.peek() {
                            if !delim.is_alphanumeric() && delim != '_' && delim != ' ' {
                                self.advance();
                                let close = match delim {
                                    '(' => ')',
                                    '[' => ']',
                                    '{' => '}',
                                    '<' => '>',
                                    c => c,
                                };
                                let mut pattern = String::new();
                                loop {
                                    match self.advance() {
                                        Some('\\') => {
                                            pattern.push('\\');
                                            if let Some(c) = self.advance() {
                                                pattern.push(c);
                                            }
                                        }
                                        Some(c) if c == close => break,
                                        Some(c) => pattern.push(c),
                                        None => {
                                            return Err(self.syntax_err(
                                                "Unterminated s/// pattern",
                                                self.line,
                                            ))
                                        }
                                    }
                                }
                                // For paired delimiters, read the opening of the replacement part
                                if "([{<".contains(delim) {
                                    self.skip_whitespace_only();
                                    let open2 = self.advance().unwrap_or(delim);
                                    let close = match open2 {
                                        '(' => ')',
                                        '[' => ']',
                                        '{' => '}',
                                        '<' => '>',
                                        c => c,
                                    };
                                    let replacement = self.read_substitution_replacement(close)?;
                                    let flags = self.read_while(|c| REGEX_FLAG_CHARS.contains(c));
                                    self.last_was_term = true;
                                    // Encode as special token — parser will decode
                                    // Format: \x00s\x00pattern\x00replacement\x00flags\x00delim
                                    return Ok(Token::Ident(format!(
                                        "\x00s\x00{}\x00{}\x00{}\x00{}",
                                        pattern, replacement, flags, delim
                                    )));
                                }
                                let replacement = self.read_substitution_replacement(close)?;
                                let flags = self.read_while(|c| REGEX_FLAG_CHARS.contains(c));
                                self.last_was_term = true;
                                return Ok(Token::Ident(format!(
                                    "\x00s\x00{}\x00{}\x00{}\x00{}",
                                    pattern, replacement, flags, delim
                                )));
                            }
                        }
                        self.last_was_term = true;
                        return Ok(Token::Ident(ident));
                    }
                    "tr" | "y" => {
                        // `$obj->tr` / `$obj->y` — after `->`, this is a method name,
                        // not transliteration.
                        if self.prev_arrow {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // After `::`, treat as package-qualified identifier, not transliteration.
                        // e.g. `Foo::y(...)` is a function call, not `y///`.
                        if self.pos >= ident.len() + 2 {
                            let prev_start = self.pos - ident.len() - 2;
                            if self.input.get(prev_start) == Some(&':')
                                && self.input.get(prev_start + 1) == Some(&':')
                            {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        // `tr` / `y` followed by terminators is a bareword, not transliteration.
                        // Check BEFORE skipping whitespace to catch newlines (implicit semicolon).
                        // `,` is treated as a terminator UNLESS the lookahead shows the
                        // full `tr,FROM,TO,FLAGS` shape — same gating as `s` above so
                        // `y` / `tr` can still appear as struct field names, list elements,
                        // and arg names without being eaten as transliteration bodies.
                        if let Some(d) = self.peek() {
                            if matches!(d, ';' | ')' | ']' | '}' | '>' | ':' | '\n') {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                            if d == ',' && !self.lookahead_is_comma_delim_subst() {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        } else {
                            self.last_was_term = true;
                            return Ok(Token::Ident(ident));
                        }
                        // Now skip whitespace to check for `=>` or `=`
                        let start_pos = self.pos;
                        let start_line = self.line;
                        self.skip_whitespace_only();
                        if let Some(d) = self.peek() {
                            // `=` alone (not `==` comparison) means assignment — y is an identifier
                            if d == '=' && self.peek_at(1) != Some('=') {
                                self.pos = start_pos;
                                self.line = start_line;
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        self.pos = start_pos;
                        self.line = start_line;
                        // Check for function signature pattern: y(...) { — this is `fn y`, not tr
                        if self.peek() == Some('(') {
                            // Scan ahead to see if there's ) followed by {
                            let scan_pos = self.pos;
                            let scan_line = self.line;
                            self.advance(); // skip (
                            let mut depth = 1;
                            while depth > 0 {
                                match self.peek() {
                                    Some('(') => {
                                        self.advance();
                                        depth += 1;
                                    }
                                    Some(')') => {
                                        self.advance();
                                        depth -= 1;
                                    }
                                    Some(_) => {
                                        self.advance();
                                    }
                                    None => break,
                                }
                            }
                            self.skip_whitespace_only();
                            let is_func_def = self.peek() == Some('{');
                            self.pos = scan_pos;
                            self.line = scan_line;
                            if is_func_def {
                                self.last_was_term = true;
                                return Ok(Token::Ident(ident));
                            }
                        }
                        // tr/from/to/flags
                        if let Some(delim) = self.peek() {
                            if !delim.is_alphanumeric() && delim != '_' && delim != ' ' {
                                self.advance();
                                let close = match delim {
                                    '(' => ')',
                                    '[' => ']',
                                    '{' => '}',
                                    '<' => '>',
                                    c => c,
                                };
                                let from = self.read_escaped_until(close)?;
                                // For paired delimiters
                                if "([{<".contains(delim) {
                                    self.skip_whitespace_only();
                                    self.advance(); // open second pair
                                }
                                let to = self.read_escaped_until(close)?;
                                let flags = self.read_while(|c| "cdsr".contains(c));
                                self.last_was_term = true;
                                return Ok(Token::Ident(format!(
                                    "\x00tr\x00{}\x00{}\x00{}\x00{}",
                                    from, to, flags, delim
                                )));
                            }
                        }
                        self.last_was_term = true;
                        return Ok(Token::Ident(ident));
                    }
                    _ => {}
                }

                // Fat arrow lookahead: ident followed by => is a string.
                // CRITICAL: save AND restore `self.line` too. `skip_whitespace_and_comments`
                // increments `self.line` for each `\n` it skips, and restoring only
                // `self.pos` leaves the line counter drifted by N. The drift bleeds into
                // every subsequent token (every `class { field\n field\n }` line shift
                // appearing on `my` / `p` / etc. after the class, and bytecode `lines[]`
                // metadata that no longer matches the source — breaking the DAP debugger
                // since BPs key off original line numbers).
                let saved_pos2 = self.pos;
                let saved_line2 = self.line;
                self.skip_whitespace_and_comments();
                if self.peek() == Some('=') && self.peek_at(1) == Some('>') {
                    self.pos = saved_pos2;
                    self.line = saved_line2;
                    self.last_was_term = true;
                    return Ok(Token::Ident(ident));
                }
                self.pos = saved_pos2;
                self.line = saved_line2;

                // Perl: `x` is the string-repetition infix operator only after a complete term.
                // After `sub`, `package`, `(`, etc. a term is expected — bare `x` must be an
                // identifier (`sub x {`, `x::Foo`, leading `x` in `(x)`).
                //
                // After `::` (package separator) or `->` (method arrow), the
                // identifier is a path leaf or method name — never an infix
                // operator. Without these gates, `keyword_or_ident("eq")` returns
                // `Token::StrEq` so `Mat::eq` fails to parse and `$obj->eq(...)`
                // silently degrades to `$obj eq …`.
                let tok = if after_package_sep || self.prev_arrow {
                    Token::Ident(ident.clone())
                } else if ident == "x" && !self.last_was_term {
                    Token::Ident("x".to_string())
                } else {
                    keyword_or_ident(&ident)
                };
                // `x=` is the string-repetition compound assignment. The
                // identifier `x` has already been consumed; peek for the
                // trailing `=` and merge into `XAssign`. Skip whitespace
                // between `x` and `=` to mirror how Perl's lexer treats
                // `$s x= 3` and `$s x =3` identically.
                let tok = if matches!(tok, Token::X) {
                    let saved_pos = self.pos;
                    let saved_line = self.line;
                    while matches!(self.peek(), Some(' ') | Some('\t')) {
                        self.advance();
                    }
                    if self.peek() == Some('=') && self.peek_at(1) != Some('=') {
                        self.advance();
                        self.last_was_term = false;
                        Token::XAssign
                    } else {
                        self.pos = saved_pos;
                        self.line = saved_line;
                        tok
                    }
                } else {
                    tok
                };
                if matches!(tok, Token::Ident(ref s) if s == "_") {
                    self.last_was_bare_positional = true;
                }
                // Keywords that expect a variable next should not set last_was_term
                // so that % is parsed as hash sigil, not modulo
                self.last_was_term = match ident.as_str() {
                    // Keywords/builtins that always expect arguments — never a term,
                    // so the next `/` is always a regex start.
                    "my"
                    | "var"
                    | "val"
                    | "mysync"
                    | "varsync"
                    | "frozen"
                    | "const"
                    | "typed"
                    | "our"
                    | "oursync"
                    | "local"
                    | "state"
                    | "return"
                    | "print"
                    | "pr"
                    | "say"
                    | "p"
                    | "die"
                    | "warn"
                    | "push"
                    | "pop"
                    | "shift"
                    | "shuffle"
                    | "chunked"
                    | "windowed"
                    | "unshift"
                    | "splice"
                    | "delete"
                    | "exists"
                    | "chomp"
                    | "chop"
                    | "defined"
                    | "keys"
                    | "values"
                    | "each"
                    | "sub"
                    | "struct"
                    | "if"
                    | "unless"
                    | "while"
                    | "until"
                    // `loop { ... }` — Rust-style infinite loop, desugars to
                    // `while (1) { ... }`. Listed here so the lexer's
                    // last_was_term tracking treats `loop` like other
                    // statement-start keywords (next char is `{`, never a
                    // continuation of a term).
                    | "loop"
                    // `ploop [N] { ... }` / `pwhile [N] (COND) { ... }` —
                    // parallel `loop` / `while`; same statement-start
                    // treatment.
                    | "ploop"
                    | "pwhile"
                    | "for"
                    | "foreach"
                    | "elsif"
                    | "use"
                    | "import"
                    | "no"
                    | "require"
                    | "eval"
                    | "do"
                    | "map"
                    | "maps"
                    | "flat_maps"
                    | "grep"
                    | "greps"
                    | "sort"
                    | "all"
                    | "any"
                    | "none"
                    | "take_while"
                    | "drop_while"
                    | "skip_while"
                    | "skip"
                    | "first_or"
                    | "tap"
                    | "peek"
                    | "with_index"
                    | "pmap"
                    | "pflat_map"
                    | "puniq"
                    | "pfirst"
                    | "pany"
                    | "pmap_chunked"
                    | "pipeline"
                    | "pgrep"
                    | "pfor"
                    | "pforeach"
                    | "par_lines"
                    | "par_walk"
                    | "pwatch"
                    | "watch"
                    | "psort"
                    | "reduce"
                    | "fold"
                    | "inject"
                    | "first"
                    | "detect"
                    | "find"
                    | "find_all"
                    | "preduce"
                    | "preduce_init"
                    | "pmap_reduce"
                    | "pcache"
                    | "fan"
                    | "fan_cap"
                    | "pchannel"
                    | "pselect"
                    | "uniq"
                    | "distinct"
                    | "flatten"
                    | "set"
                    | "list_count"
                    | "list_size"
                    | "count"
                    | "len"
                    | "size"
                    | "cnt"
                    | "zip"
                    | "async"
                    | "trace"
                    | "timer"
                    | "await"
                    | "slurp"
                    | "swallow"
                    | "ingest"
                    | "burp"
                    | "god"
                    | "capture"
                    | "fetch_url"
                    | "fetch"
                    | "fetch_json"
                    | "fetch_async"
                    | "fetch_async_json"
                    | "par_fetch"
                    | "par_csv_read"
                    | "par_pipeline"
                    | "par_pipeline_stream"
                    | "par_sed"
                    | "join"
                    | "json_encode"
                    | "json_decode"
                    | "json_jq"
                    | "jwt_encode"
                    | "jwt_decode"
                    | "jwt_decode_unsafe"
                    | "log_info"
                    | "log_warn"
                    | "log_error"
                    | "log_debug"
                    | "log_trace"
                    | "log_json"
                    | "log_level"
                    | "sha256"
                    | "sha1"
                    | "md5"
                    | "hmac_sha256"
                    | "hmac"
                    | "uuid"
                    | "base64_encode"
                    | "base64_decode"
                    | "hex_encode"
                    | "hex_decode"
                    | "gzip"
                    | "gunzip"
                    | "zstd"
                    | "zstd_decode"
                    | "datetime_utc"
                    | "datetime_from_epoch"
                    | "datetime_parse_rfc3339"
                    | "datetime_strftime"
                    | "toml_decode"
                    | "toml_encode"
                    | "yaml_decode"
                    | "yaml_encode"
                    | "url_encode"
                    | "url_decode"
                    | "uri_escape"
                    | "uri_unescape"
                    | "split"
                    | "reverse"
                    | "reversed"
                    | "not"
                    | "ref"
                    | "scalar"
                    | "try"
                    | "catch"
                    | "finally"
                    | "given"
                    | "when"
                    | "default"
                    | "eval_timeout"
                    | "tie"
                    | "retry"
                    | "rate_limit"
                    | "every"
                    | "gen"
                    | "yield"
                    | "match"
                    | "filter"
                    | "f"
                    | "reject"
                    | "grepv"
                    | "collect"
                    | "compact"
                    | "concat"
                    | "chain"
                    | "min_by"
                    | "max_by"
                    | "sort_by"
                    | "tally"
                    | "find_index"
                    | "each_with_index"
                    | "fore"
                    | "e"
                    | "ep"
                    | "flat_map"
                    | "group_by"
                    | "chunk_by"
                    | "bench" => false,
                    // `thread`/`t` are ambiguous: at statement start they're the
                    // thread keyword (expect args → false), but after an operator
                    // they could be variable names (e.g., `$x / t / 2` → true).
                    "thread" | "t" => !self.last_was_term,
                    _ => matches!(tok, Token::Ident(_)),
                };
                Ok(tok)
            }

            c => Err(self.syntax_err(format!("Unexpected character '{c}'"), self.line)),
        }
    }

    /// Tokenize entire input.
    pub fn tokenize(&mut self) -> StrykeResult<Vec<(Token, usize)>> {
        let mut tokens = Vec::new();
        loop {
            // `next_token` internally skips whitespace + POD / heredoc and
            // stamps `self.token_start_line` to the line where the emitted
            // token actually starts. Use that, not a pre-call snapshot of
            // `self.line` — POD blocks can advance the line counter by
            // many lines before the real token is produced (see
            // `token_start_line` doc).
            let tok = self.next_token()?;
            let line = self.token_start_line;
            if self.last_was_bare_positional {
                self.bare_positional_indices.insert(tokens.len());
            }
            if tok == Token::Eof {
                tokens.push((Token::Eof, line));
                break;
            }
            tokens.push((tok, line));
        }
        Ok(tokens)
    }
}

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

    #[test]
    fn tokenize_empty_yields_eof() {
        let mut l = Lexer::new("");
        let t = l.tokenize().expect("tokenize");
        assert_eq!(t.len(), 1);
        assert!(matches!(t[0].0, Token::Eof));
    }

    #[test]
    fn tokenize_integer_literal() {
        let mut l = Lexer::new("42");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(42)));
    }

    #[test]
    fn tokenize_keyword_my_and_semicolon() {
        let mut l = Lexer::new("my;");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "my"));
        assert!(matches!(t[1].0, Token::Semicolon));
    }

    #[test]
    fn tokenize_skips_hash_line_comment() {
        let mut l = Lexer::new("1#comment\n2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(1)));
        assert!(matches!(t[1].0, Token::Integer(2)));
        assert!(matches!(t[2].0, Token::Eof));
    }

    #[test]
    fn tokenize_double_quoted_string_literal() {
        let mut l = Lexer::new(r#""hi""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "hi"));
    }

    #[test]
    fn tokenize_triple_quoted_interpolating_multiline() {
        // `"""..."""` — interpolating triple-quote. Newlines preserved
        // raw; interpolation flag stays on (Token::DoubleString flows
        // through the regular string-interp pipeline downstream).
        let mut l = Lexer::new("\"\"\"hello\nworld\nline\"\"\"");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::DoubleString(ref s) if s == "hello\nworld\nline"),
            "got: {:?}",
            t[0].0
        );
    }

    #[test]
    fn tokenize_triple_quoted_empty() {
        let mut l = Lexer::new("\"\"\"\"\"\"");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s.is_empty()));
    }

    #[test]
    fn tokenize_triple_quoted_with_embedded_quotes() {
        // Two consecutive `""` inside a `"""..."""` body should not
        // close — only three `"""` in a row terminates.
        let mut l = Lexer::new("\"\"\"a \"\" b\"\"\"");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::DoubleString(ref s) if s == "a \"\" b"),
            "got: {:?}",
            t[0].0
        );
    }

    #[test]
    fn tokenize_raw_triple_quoted() {
        // `r"""..."""` — non-interpolating raw triple-quote. No escape
        // processing: `\n` stays as the two literal chars `\` and `n`.
        let mut l = Lexer::new("r\"\"\"raw \\n and $no_interp\"\"\"");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::SingleString(ref s) if s == "raw \\n and $no_interp"),
            "got: {:?}",
            t[0].0
        );
    }

    #[test]
    fn tokenize_r_bareword_not_triple_quote() {
        // Lone `r` (not followed by `"""`) is still a plain identifier.
        let mut l = Lexer::new("r => 5");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "r"));
    }

    #[test]
    fn tokenize_unterminated_triple_quote_errors() {
        let mut l = Lexer::new("\"\"\"never closes");
        assert!(l.tokenize().is_err());
    }

    #[test]
    fn tokenize_double_string_escaped_sigils_are_literal() {
        // `\$` in source becomes a sentinel + parser emits literal `$` (not outer interpolation).
        let mut l = Lexer::new(r#""my \$x""#);
        let t = l.tokenize().expect("tokenize");
        let want = format!("my {}x", LITERAL_DOLLAR_IN_DQUOTE);
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if *s == want));
    }

    #[test]
    fn tokenize_double_string_braced_hex_unicode_escape() {
        let mut l = Lexer::new(r#""\x{1215}""#);
        let t = l.tokenize().expect("tokenize");
        let want: String = ['\u{1215}'].into_iter().collect();
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if *s == want));
    }

    #[test]
    fn tokenize_double_string_braced_unicode_u_escape() {
        let mut l = Lexer::new(r#""\u{0301}""#);
        let t = l.tokenize().expect("tokenize");
        let want: String = ['\u{0301}'].into_iter().collect();
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if *s == want));
    }

    #[test]
    fn tokenize_double_string_braced_unicode_u_escape_multi() {
        // \u{0041} = 'A', \u{00E9} = 'é', \u{1F600} = '😀'
        let mut l = Lexer::new(r#""\u{0041}\u{00E9}\u{1F600}""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "Aé😀"));
    }

    #[test]
    fn tokenize_double_string_octal_escape() {
        let mut l = Lexer::new(r#""\101""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "A"));
    }

    #[test]
    fn tokenize_double_string_braced_octal_escape() {
        let mut l = Lexer::new(r#""\o{101}""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "A"));
    }

    #[test]
    fn tokenize_double_string_control_char_escape() {
        let mut l = Lexer::new(r#""\cA""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "\x01"));
    }

    #[test]
    fn tokenize_double_string_named_unicode_escape() {
        let mut l = Lexer::new(r#""\N{SNOWMAN}""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "☃"));
    }

    #[test]
    fn tokenize_double_string_named_unicode_u_plus() {
        let mut l = Lexer::new(r#""\N{U+2603}""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "☃"));
    }

    #[test]
    fn tokenize_double_string_unbraced_hex_two_digits() {
        let mut l = Lexer::new(r#""\x41""#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "A"));
    }

    #[test]
    fn tokenize_single_quoted_string_literal() {
        let mut l = Lexer::new("'x'");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::SingleString(ref s) if s == "x"));
    }

    #[test]
    fn tokenize_spaceship_operator() {
        let mut l = Lexer::new("1 <=> 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(1)));
        assert!(matches!(t[1].0, Token::Spaceship));
        assert!(matches!(t[2].0, Token::Integer(2)));
    }

    #[test]
    fn tokenize_m_regex_literal() {
        let mut l = Lexer::new("m/abc/");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Regex(ref p, ref f, _) if p == "abc" && f.is_empty()));
    }

    #[test]
    fn tokenize_q_brace_constructor() {
        let mut l = Lexer::new("q{lit}");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::SingleString(ref s) if s == "lit"));
    }

    /// `q(sub ($) { 1 })` — nested `()` must not end at the `)` in `($)` (core `Carp.pm`).
    #[test]
    fn tokenize_q_paren_balances_nested_parens_in_prototype() {
        let mut l = Lexer::new("q(fn ($) { 1 })");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::SingleString(ref s) if s == "fn ($) { 1 }"));
    }

    /// `qw( (SV*)x )` — nested `()` inside `qw(...)` (core `B.pm`).
    #[test]
    fn tokenize_qw_paren_balances_nested_parens() {
        let mut l = Lexer::new("qw( (SV*)pWARN_ALL )");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::QW(ref w) if w.len() == 1 && w[0] == "(SV*)pWARN_ALL"));
    }

    #[test]
    fn tokenize_float_literal() {
        let mut l = Lexer::new("3.25");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Float(f) if (f - 3.25).abs() < f64::EPSILON));
    }

    #[test]
    fn tokenize_scientific_float() {
        let mut l = Lexer::new("1e2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Float(f) if (f - 100.0).abs() < 1e-9));
    }

    #[test]
    fn tokenize_hex_with_underscore_separators() {
        let mut l = Lexer::new("0x_FF");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(255)));
    }

    #[test]
    fn tokenize_qr_regex_with_flags() {
        let mut l = Lexer::new("qr/pat/i");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Regex(ref p, ref f, _) if p == "pat" && f == "i"));
    }

    #[test]
    fn tokenize_m_slash_includes_gc_flags() {
        let mut l = Lexer::new("m/./gc");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(&t[0].0, Token::Regex(p, f, _) if p == "." && f == "gc"));
    }

    #[test]
    fn tokenize_m_hash_delimiter_includes_gc_flags() {
        let mut l = Lexer::new("m#\\w#gc");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(&t[0].0, Token::Regex(p, f, _) if p == r"\w" && f == "gc"));
    }

    #[test]
    fn tokenize_qr_slash_includes_gco_flags() {
        let mut l = Lexer::new("qr/x/gco");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(&t[0].0, Token::Regex(p, f, _) if p == "x" && f == "gco"));
    }

    #[test]
    fn tokenize_qw_hash_delimiter_not_line_comment() {
        // `#` after `qw` must be the opener, not `skip_whitespace_and_comments` eating the line.
        let mut l = Lexer::new("qw# a b #;");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(&t[0].0, Token::QW(w) if w == &["a", "b"]),
            "first={:?}",
            t.first()
        );
    }

    #[test]
    fn tokenize_qq_hash_delimiter_single_line() {
        let mut l = Lexer::new("qq#x#;");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(&t[0].0, Token::DoubleString(s) if s == "x"));
    }

    #[test]
    fn tokenize_qr_hash_delimiter_text_balanced_preamble() {
        let src = "qr#(\n    [!=]~\n    | split|grep|map\n    | not|and|or|xor\n)#x";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        let Token::Regex(p, f, _) = &t[0].0 else {
            panic!("expected Regex, got {:?}", t[0].0);
        };
        let rest: Vec<_> = t.iter().skip(1).take(8).map(|x| &x.0).collect();
        assert!(f.contains('x'), "flags={f:?} pattern={p:?} rest={rest:?}");
        assert!(p.contains("[!=]~"), "{p:?}");
        assert!(p.contains("split|grep|map"), "{p:?}");
    }

    #[test]
    fn tokenize_octal_integer_literal() {
        let mut l = Lexer::new("010");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(8)));
    }

    #[test]
    fn tokenize_binary_integer_literal() {
        let mut l = Lexer::new("0b1010");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(10)));
    }

    #[test]
    fn tokenize_filetest_exists() {
        let mut l = Lexer::new("-e '.'");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::FileTest('e')));
        assert!(matches!(t[1].0, Token::SingleString(ref s) if s == "."));
    }

    #[test]
    fn tokenize_filetest_tty() {
        let mut l = Lexer::new("-t 'STDIN'");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::FileTest('t')));
        assert!(matches!(t[1].0, Token::SingleString(ref s) if s == "STDIN"));
    }

    #[test]
    fn tokenize_power_and_range_operators() {
        let mut l = Lexer::new("2 ** 3");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(2)));
        assert!(matches!(t[1].0, Token::Power));
        assert!(matches!(t[2].0, Token::Integer(3)));

        let mut l = Lexer::new("1..4");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(1)));
        assert!(matches!(t[1].0, Token::Range));
        assert!(matches!(t[2].0, Token::Integer(4)));
    }

    #[test]
    fn tokenize_numeric_equality_operators() {
        let mut l = Lexer::new("1 == 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(1)));
        assert!(matches!(t[1].0, Token::NumEq));
        assert!(matches!(t[2].0, Token::Integer(2)));

        let mut l = Lexer::new("3 != 4");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(3)));
        assert!(matches!(t[1].0, Token::NumNe));
        assert!(matches!(t[2].0, Token::Integer(4)));
    }

    #[test]
    fn tokenize_logical_and_or_plus_assign() {
        let mut l = Lexer::new("1 && 0");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(1)));
        assert!(matches!(t[1].0, Token::LogAnd));
        assert!(matches!(t[2].0, Token::Integer(0)));

        let mut l = Lexer::new("0 || 9");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(0)));
        assert!(matches!(t[1].0, Token::LogOr));
        assert!(matches!(t[2].0, Token::Integer(9)));

        let mut l = Lexer::new("n += 1");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "n"));
        assert!(matches!(t[1].0, Token::PlusAssign));
        assert!(matches!(t[2].0, Token::Integer(1)));
    }

    #[test]
    fn tokenize_bitwise_and_operator() {
        let mut l = Lexer::new("3 & 5");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(3)));
        assert!(matches!(t[1].0, Token::BitAnd));
        assert!(matches!(t[2].0, Token::Integer(5)));
    }

    #[test]
    fn tokenize_braced_caret_scalar_global_phase() {
        let mut l = Lexer::new(r#"print ${^GLOBAL_PHASE}, "\n";"#);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "print"));
        assert!(matches!(t[1].0, Token::ScalarVar(ref s) if s == "^GLOBAL_PHASE"));
        assert!(matches!(t[2].0, Token::Comma));
        assert!(matches!(t[3].0, Token::DoubleString(ref s) if s == "\n"));
        assert!(matches!(t[4].0, Token::Semicolon));
    }

    #[test]
    fn tokenize_bitwise_or_and_assign() {
        let mut l = Lexer::new("$a |= $b");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "a"));
        assert!(matches!(t[1].0, Token::BitOrAssign));
        assert!(matches!(t[2].0, Token::ScalarVar(ref s) if s == "b"));

        let mut l = Lexer::new("$a &= $b");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::BitAndAssign));
    }

    #[test]
    fn tokenize_division_and_modulo() {
        let mut l = Lexer::new("7 / 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::Slash));

        let mut l = Lexer::new("7 % 3");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::Percent));
    }

    #[test]
    fn tokenize_comma_fat_arrow_and_semicolon() {
        let mut l = Lexer::new("a => 1;");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "a"));
        assert!(matches!(t[1].0, Token::FatArrow));
        assert!(matches!(t[2].0, Token::Integer(1)));
        assert!(matches!(t[3].0, Token::Semicolon));
    }

    #[test]
    fn tokenize_minus_unary_vs_binary() {
        let mut l = Lexer::new("- 5");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Minus));
        assert!(matches!(t[1].0, Token::Integer(5)));
    }

    #[test]
    fn tokenize_dollar_scalar_sigil() {
        let mut l = Lexer::new("$foo");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "foo"));
    }

    /// `=` + letter is assignment unless `=` starts the line (POD). `$_=foo` must not skip POD.
    #[test]
    fn tokenize_assign_not_pod_when_eq_not_line_start() {
        let mut l = Lexer::new("$_=foo;");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "_"));
        assert!(matches!(t[1].0, Token::Assign));
        assert!(matches!(t[2].0, Token::Ident(ref s) if s == "foo"));
        assert!(matches!(t[3].0, Token::Semicolon));
    }

    #[test]
    fn tokenize_pod_equals_still_skipped_at_line_start() {
        let mut l = Lexer::new("=head1 NAME\ncode\n=cut\n$x;");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "x"));
        assert!(matches!(t[1].0, Token::Semicolon));
    }

    #[test]
    fn tokenize_at_array_sigil() {
        let mut l = Lexer::new("@arr");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ArrayVar(ref s) if s == "arr"));
    }

    #[test]
    fn tokenize_at_caret_capture_array() {
        let mut l = Lexer::new("@^CAPTURE");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ArrayVar(ref s) if s == "^CAPTURE"));
    }

    #[test]
    fn tokenize_percent_caret_hook_hash() {
        let mut l = Lexer::new("%^HOOK");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::HashVar(ref s) if s == "^HOOK"));
    }

    #[test]
    fn tokenize_caret_letter_and_at_minus_plus() {
        let mut l = Lexer::new("$^I@-@+");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "^I"));
        assert!(matches!(t[1].0, Token::ArrayVar(ref s) if s == "-"));
        assert!(matches!(t[2].0, Token::ArrayVar(ref s) if s == "+"));
    }

    #[test]
    fn tokenize_percent_hash_sigil() {
        let mut l = Lexer::new("%h");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::HashVar(ref s) if s == "h"));
    }

    #[test]
    fn tokenize_percent_plus_named_capture_hash() {
        let mut l = Lexer::new("%+");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::HashVar(ref s) if s == "+"));
    }

    #[test]
    fn tokenize_dollar_dollar_under_brace_is_not_pid() {
        // `$$_{$k}` — second `$$` is not PID; tokenizes as `$_` then `{` (Perl `$_->{$k}`).
        let mut l = Lexer::new("$$_{$k}");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "_"));
        assert!(matches!(t[1].0, Token::LBrace));
    }

    #[test]
    fn tokenize_braced_scalar_deref_try_tiny() {
        // `${$code_ref}` ≡ `$$code_ref` (Try::Tiny blesses scalar refs to coderefs).
        let mut l = Lexer::new("${$code_ref}");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DerefScalarVar(ref s) if s == "code_ref"));
    }

    #[test]
    fn tokenize_braced_scalar_deref_package_qualified() {
        let mut l = Lexer::new("${$Foo::bar}");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DerefScalarVar(ref s) if s == "Foo::bar"));
    }

    #[test]
    fn tokenize_dollar_colon_stash_brace() {
        // `$::{$k}` — `%::` main stash (core Carp.pm line 32).
        let mut l = Lexer::new("$::{$pack}");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ScalarVar(ref s) if s == "::"));
        assert!(matches!(t[1].0, Token::LBrace));
    }

    #[test]
    fn tokenize_ampersand_then_ident_is_bitand_not_coderef() {
        // Subroutine coderef `&name` is not a distinct token; lexer emits `&` then ident.
        let mut l = Lexer::new("&f");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::BitAnd));
        assert!(matches!(t[1].0, Token::Ident(ref s) if s == "f"));
    }

    #[test]
    fn tokenize_qq_paren_constructor() {
        let mut l = Lexer::new("qq(x y)");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if s == "x y"));
    }

    #[test]
    fn tokenize_qq_slash_escaped_dollar_is_literal() {
        let mut l = Lexer::new(r#"qq/my \$y/"#);
        let t = l.tokenize().expect("tokenize");
        let want = format!("my {}y", LITERAL_DOLLAR_IN_DQUOTE);
        assert!(matches!(t[0].0, Token::DoubleString(ref s) if *s == want));
    }

    #[test]
    fn tokenize_s_substitution_alternate_delimiter() {
        let mut l = Lexer::new("s#a#b#");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s.starts_with("\x00s\x00")));
    }

    #[test]
    fn tokenize_tr_slash_delimiter() {
        let mut l = Lexer::new("tr/a/b/");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s.starts_with("\x00tr\x00")));
    }

    #[test]
    fn tokenize_y_synonym_for_tr() {
        let mut l = Lexer::new("y/x/y/");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s.starts_with("\x00tr\x00")));
    }

    #[test]
    fn tokenize_less_equal_greater_relops() {
        let mut l = Lexer::new("1 <= 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::NumLe));

        let mut l = Lexer::new("3 >= 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::NumGe));

        let mut l = Lexer::new("1 < 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::NumLt));

        let mut l = Lexer::new("3 > 2");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::NumGt));
    }

    #[test]
    fn tokenize_readline_scalar_handle() {
        let mut l = Lexer::new("<$fh>");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::ReadLine(ref s) if s == "fh"));
    }

    #[test]
    fn tokenize_shift_right_and_shift_left_assign() {
        let mut l = Lexer::new("8 >> 1");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::ShiftRight));

        let mut l = Lexer::new("8 << 1");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::ShiftLeft));

        let mut l = Lexer::new("x <<= 3");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::ShiftLeftAssign));
    }

    #[test]
    fn tokenize_heredoc_after_print_not_shift() {
        let src = "print <<EOT\nhi\nEOT\n";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "print"));
        assert!(
            matches!(&t[1].0, Token::HereDoc(tag, body, interpolate) if tag == "EOT" && body == "hi\n" && *interpolate),
            "got {:?}",
            t[1].0
        );
    }

    #[test]
    fn tokenize_bitwise_or_xor() {
        let mut l = Lexer::new("3 | 1");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::BitOr));

        let mut l = Lexer::new("3 ^ 1");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::BitXor));
    }

    #[test]
    fn tokenize_pipe_forward_vs_bitor_vs_logor() {
        // `|>` must lex as a distinct token (not `|` followed by `>`).
        let mut l = Lexer::new("1 |> f");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::PipeForward), "got {:?}", t[1].0);

        // Make sure `|` and `||` still work alongside `|>`.
        let mut l = Lexer::new("a | b || c |> d");
        let t = l.tokenize().expect("tokenize");
        let kinds: Vec<_> = t.iter().map(|(k, _)| k.clone()).collect();
        assert!(kinds.iter().any(|k| matches!(k, Token::BitOr)));
        assert!(kinds.iter().any(|k| matches!(k, Token::LogOr)));
        assert!(kinds.iter().any(|k| matches!(k, Token::PipeForward)));
    }

    #[test]
    fn tokenize_compare_and_three_way_string_ops() {
        let mut l = Lexer::new("\"a\" cmp \"b\"");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::StrCmp));
    }

    #[test]
    fn tokenize_package_double_colon_splits_qualified_name() {
        let mut l = Lexer::new("Foo::Bar::baz");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "Foo"));
        assert!(matches!(t[1].0, Token::PackageSep));
        assert!(matches!(t[2].0, Token::Ident(ref s) if s == "Bar"));
        assert!(matches!(t[3].0, Token::PackageSep));
        assert!(matches!(t[4].0, Token::Ident(ref s) if s == "baz"));
    }

    #[test]
    fn tokenize_pod_line_skipped_like_comment_prefix() {
        // `=head1` at line start starts POD; lexer should skip until =cut
        let mut l = Lexer::new("=pod\n=cut\n42");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Integer(42)));
    }

    #[test]
    fn tokenize_underscore_in_identifier() {
        let mut l = Lexer::new("__PACKAGE__");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "__PACKAGE__"));
    }

    /// `x` is the repetition operator only in infix position; after `sub` it is a sub name (Perl).
    #[test]
    fn tokenize_x_repeat_vs_sub_name() {
        let mut l = Lexer::new("3 x 4");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[1].0, Token::X));

        let mut l = Lexer::new("sub x { 1 }");
        let t = l.tokenize().expect("tokenize");
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "sub"));
        assert!(matches!(t[1].0, Token::Ident(ref s) if s == "x"));
    }

    /// Regression for the fat-arrow lookahead drift. Every identifier is
    /// followed by a speculative `skip_whitespace_and_comments` peek for
    /// `=>`, which advances `self.line` over any intervening newlines.
    /// Restoring only `self.pos` (the original bug) leaves the line counter
    /// drifted +1 per newline crossed, so every subsequent statement
    /// reports a wrong source line and breakpoints on those lines silently
    /// drop. With the save/restore fix, the line numbers match the source.
    #[test]
    fn fat_arrow_lookahead_does_not_drift_line() {
        let src = "x\ny\nz\n";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        // Token lines should match source lines.
        assert!(matches!(t[0].0, Token::Ident(ref s) if s == "x"));
        assert_eq!(t[0].1, 1, "x is on line 1");
        assert!(matches!(t[1].0, Token::Ident(ref s) if s == "y"));
        assert_eq!(t[1].1, 2, "y is on line 2");
        assert!(matches!(t[2].0, Token::Ident(ref s) if s == "z"));
        assert_eq!(t[2].1, 3, "z is on line 3");
    }

    /// Specific case from the bug report: a class-body with a typed field
    /// followed by code. The bug had `=>` lookahead consume the `\n` after
    /// `Int`, then the recursive next_token / outer tokenize re-process the
    /// same `\n`, double-incrementing `self.line`. With the fix, statements
    /// after the class body report their real source lines.
    #[test]
    fn class_field_does_not_drift_subsequent_statement_line() {
        let src = "class Foo {\n    x: Int\n}\nmy $y = 1\np $y\n";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        // Find the `my` token after the class body.
        let my_line = t
            .iter()
            .find_map(|(tok, line)| match tok {
                Token::Ident(s) if s == "my" => Some(*line),
                _ => None,
            })
            .expect("expected `my` token");
        assert_eq!(my_line, 4, "my $y = 1 lives on source line 4");
    }

    /// Same shape as `fat_arrow_lookahead_does_not_drift_line` but for the
    /// `qw` bareword-vs-quote-operator lookahead. When `qw` is followed by
    /// terminators (here a newline + closing paren) it's lexed as an Ident,
    /// not a `qw(...)` list. The `skip_whitespace_only` call inside the
    /// lookahead advances `self.line` for the `\n` it consumes; without
    /// restoring `self.line` the next token's line drifted +1.
    #[test]
    fn qw_bareword_lookahead_does_not_drift_line() {
        // `qw` followed by `)` (treated as bareword) on its own line, then
        // a new statement. The post-class-field pattern: a newline between
        // `qw` and the next statement must not bump the line counter.
        let src = "(qw\n)\nmy $x = 1\n";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        let my_line = t
            .iter()
            .find_map(|(tok, line)| match tok {
                Token::Ident(s) if s == "my" => Some(*line),
                _ => None,
            })
            .expect("expected `my` token");
        assert_eq!(my_line, 3, "my $x = 1 lives on source line 3");
    }

    /// `m` as a bareword (the parser handles `$obj->m`, sort `m`-prefix,
    /// etc.) goes through the same skip-whitespace-only lookahead as `qw`.
    /// Newlines between `m` and the next non-terminator must not drift
    /// the lexer's `self.line` past restoration.
    #[test]
    fn m_bareword_lookahead_does_not_drift_line() {
        // `$obj->m` — after `->`, `m` is a method name; the followup token
        // tracking must keep its source line.
        let src = "$obj->m\nmy $y = 2\n";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        let my_line = t
            .iter()
            .find_map(|(tok, line)| match tok {
                Token::Ident(s) if s == "my" => Some(*line),
                _ => None,
            })
            .expect("expected `my` token");
        assert_eq!(my_line, 2, "my $y = 2 lives on source line 2");
    }

    /// Bareword `_N` (underscore + ≥1 digit) lexes to `ScalarVar(_N)` —
    /// the implicit-closure-positional name that works without a sigil.
    /// The lexer carves these out at line 2241 so the parser sees them
    /// as variables, not as undefined sub names.
    #[test]
    fn bareword_underscore_n_lexes_as_scalar_var() {
        let mut l = Lexer::new("_1");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::ScalarVar(ref s) if s == "_1"),
            "_1 → ScalarVar(_1): got {:?}",
            t[0].0
        );
    }

    #[test]
    fn bareword_underscore_n_two_digits_lexes_as_scalar_var() {
        let mut l = Lexer::new("_42");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::ScalarVar(ref s) if s == "_42"),
            "_42 → ScalarVar(_42): got {:?}",
            t[0].0
        );
    }

    /// Bare `_` (no digit) stays an Ident so the existing topic /
    /// bareword-call machinery keeps working — `_` can be a sub call,
    /// a bareword filename, etc., context-dependent.
    #[test]
    fn bare_underscore_alone_lexes_as_ident() {
        let mut l = Lexer::new("_");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::Ident(ref s) if s == "_"),
            "_ → Ident(_): got {:?}",
            t[0].0
        );
    }

    /// `_foo` is a user-defined identifier — NOT a topic alias.
    /// Only `_<digits>` is reserved for the positional slot family.
    #[test]
    fn underscore_prefix_word_is_plain_ident() {
        let mut l = Lexer::new("_foo");
        let t = l.tokenize().expect("tokenize");
        assert!(
            matches!(t[0].0, Token::Ident(ref s) if s == "_foo"),
            "_foo → Ident(_foo): got {:?}",
            t[0].0
        );
    }

    /// POD blocks (`=pod ... =cut`) advance `self.line` many lines during
    /// the skip but the original tokenize loop captured `line` *before*
    /// calling next_token, then pushed (token, captured_line) — so the
    /// first token after POD got the pre-POD line instead of its real
    /// post-skip line. Now next_token stamps `self.token_start_line` after
    /// the skip and tokenize reads from there.
    #[test]
    fn token_after_pod_block_uses_post_skip_line() {
        let src = "\
=pod
big POD
block here
=cut
my $x = 1
";
        let mut l = Lexer::new(src);
        let t = l.tokenize().expect("tokenize");
        // First non-Eof token is `my`, on source line 5.
        let (first_tok, first_line) = t
            .iter()
            .find(|(tok, _)| !matches!(tok, Token::Eof))
            .expect("non-empty tokens");
        assert!(matches!(first_tok, Token::Ident(ref s) if s == "my"));
        assert_eq!(
            *first_line, 5,
            "my $x lives on source line 5 (after POD ends)"
        );
    }
}