rich/

json.rs

//! JSON pretty-printing.
//!
//! Port of upstream `rich/json.py`. [`Json`] parses a JSON string and renders it
//! with 2-space indentation (matching Python's `json.dumps(indent=2)`) and the
//! default JSON highlight colors.
//!
//! Non-ASCII strings render as UTF-8, matching upstream (`rich.json.JSON`
//! defaults to `ensure_ascii=False`); object keys keep input order, and a
//! repeated key keeps its first position but its last value — what both
//! `dict` and serde_json's `preserve_order` do. Number formatting follows
//! `json.dumps` too: floats are written
//! with Python's `float.__repr__` ([`crate::pyformat::float_repr`]).
//! [`JsonOptions`] carries upstream's `indent`, `sort_keys`, `ensure_ascii`,
//! `allow_nan` and `highlight`; highlighting is upstream's `JSONHighlighter`.
//!
//! ## Why the parser is hand-written
//!
//! Upstream's parser is Python's `json`, which differs from `serde_json` in two
//! important ways this module reproduces:
//!
//! * `json.loads` accepts (and `json.dumps(allow_nan=True)` emits) the
//!   non-finite literals `NaN`, `Infinity` and `-Infinity`. `serde_json` has no
//!   `Value` that can hold them and rejects the documents outright.
//! * `serde_json` caps nesting at 128 levels, so a 200-deep document — which
//!   CPython parses without complaint — came back as "invalid JSON".
//!
//! Raising a recursion limit only moves the failure to a stack overflow, so
//! parsing, rendering and *dropping* the tree here are all iterative: nesting
//! depth costs heap, never stack. String decoding uses `serde_json`; integer
//! tokens retain their exact digits, finite floats use Python's `repr`, and
//! overflowing exponents become signed Infinity as in Python.
//!
//! Nesting is still bounded, at [`MAX_DEPTH`] levels, where CPython raises
//! `RecursionError`: `json.loads` recurses in C, and CPython 3.12+ stops just
//! short of 10 000 levels (3.11 at its 1 000-frame recursion limit). Rendering
//! indents every line by its depth, so an unbounded document costs quadratic
//! memory: a 100 000-deep array asked for tens of gigabytes and was killed
//! rather than reporting an error. The limit accepts every document CPython
//! 3.12+ accepts; the few levels between its exact (stack-dependent) cutoff
//! and `MAX_DEPTH` are the only difference.

use std::collections::HashMap;

/// The deepest nesting [`Json::new`] accepts (DIVERGENCES #25). Deeper
/// documents fail with the error CPython's `json.loads` raises
/// (`RecursionError`); see the module docs.
pub const MAX_DEPTH: usize = 10_000;

use crate::console::{Console, ConsoleOptions};
use crate::errors::{Result, RichError};
use crate::measure::Measurement;
use crate::protocol::Renderable;
use crate::segment::Segment;
use crate::style::StyleType;
use crate::text::{Span, Text};

/// A parsed JSON document, rendered with syntax highlighting. Mirrors `rich.json.JSON`.
///
/// Like upstream, the document is re-serialized (`json.dumps`) with the
/// [`JsonOptions`] and highlighted with `JSONHighlighter` when it is built;
/// [`text`](Json::text) is that highlighted `Text`.
pub struct Json {
    value: Node,
    options: JsonOptions,
    /// The highlighted document, one `Text` per physical line (upstream
    /// wraps a `Text` line by line, so this is how it renders). Built on
    /// first use: a document near [`MAX_DEPTH`] formats to hundreds of
    /// megabytes of indentation, and one that is only parsed and dropped
    /// should not pay for that.
    lines: std::sync::OnceLock<Vec<Text>>,
    /// See [`Json::no_wrap`].
    no_wrap: bool,
    #[cfg(feature = "json-escape-safe")]
    escape_safe: bool,
}

/// The keyword arguments of `rich.json.JSON`, for [`Json::with_options`].
///
/// `skip_keys` and `check_circular` are accepted for parity but cannot change
/// the output of a document parsed from a string: its keys are all strings
/// and it has no cycles. Upstream's `default=` callable only applies to
/// `JSON.from_data`, which serializes arbitrary (non-JSON) objects.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct JsonOptions {
    /// `indent=`: the indent string, or `None` for a single line. An integer
    /// indent `n` is `n` spaces (see [`JsonOptions::indent_spaces`]).
    pub indent: Option<String>,
    /// `highlight=`: apply `JSONHighlighter` (default on).
    pub highlight: bool,
    /// `skip_keys=` (no effect on a parsed document).
    pub skip_keys: bool,
    /// `ensure_ascii=`: escape every non-ASCII character (default off).
    pub ensure_ascii: bool,
    /// `check_circular=` (no effect on a parsed document).
    pub check_circular: bool,
    /// `allow_nan=`: with it off, `NaN` and the infinities are an error
    /// (default on).
    pub allow_nan: bool,
    /// `sort_keys=`: sort object keys (default off).
    pub sort_keys: bool,
}

impl Default for JsonOptions {
    fn default() -> Self {
        JsonOptions {
            indent: Some("  ".to_string()),
            highlight: true,
            skip_keys: false,
            ensure_ascii: false,
            check_circular: true,
            allow_nan: true,
            sort_keys: false,
        }
    }
}

impl JsonOptions {
    /// An integer `indent=`: `n` spaces, as `json.dumps` turns an int into
    /// `" " * n` (so zero or a negative number puts each item on its own line
    /// with no indentation).
    pub fn indent_spaces(indent: i64) -> Option<String> {
        Some(" ".repeat(usize::try_from(indent).unwrap_or(0)))
    }
}

/// A Python `str` as generalized UTF-8 (WTF-8): UTF-8, except that a lone
/// surrogate from a `\uD800`-style escape (which Python's `json` accepts)
/// is kept as its three-byte encoding, as `surrogatepass` writes it. Byte
/// order is still code point order, so `sort_keys` sorts as Python does.
type PyStr = Vec<u8>;

/// A parsed JSON value.
///
/// Scalars keep the form they will be printed in: numbers are stored already
/// normalized where needed, strings already decoded (upstream re-encodes them
/// through `json.dumps`, so `"A"` prints as `"A"`).
#[derive(Debug)]
enum Node {
    Null,
    Bool(bool),
    Number(String),
    /// `NaN`, `Infinity` or `-Infinity`. Python's `json` round-trips these;
    /// rich's `JSONHighlighter` has no rule that matches them, so they print
    /// unstyled.
    NonFinite(&'static str),
    Str(PyStr),
    Array(Vec<Node>),
    Object(Vec<(PyStr, Node)>),
}

impl Drop for Node {
    /// Dismantle the tree with an explicit stack.
    ///
    /// The compiler's drop glue recurses once per nesting level, so a document
    /// deep enough to parse (parsing has no depth limit here) would overflow
    /// the stack on the way out — a crash with no error message at all, which
    /// is worse than the rejection this module used to hand out.
    fn drop(&mut self) {
        let mut pending: Vec<Node> = Vec::new();
        take_children(self, &mut pending);
        while let Some(mut node) = pending.pop() {
            take_children(&mut node, &mut pending);
            // `node` drops here with its children already moved out, so this
            // same `drop` runs against an empty container and stops.
        }
    }
}

/// Move a node's children into `out`, leaving the node childless.
fn take_children(node: &mut Node, out: &mut Vec<Node>) {
    match node {
        Node::Array(items) => out.append(items),
        Node::Object(entries) => out.extend(entries.drain(..).map(|(_, value)| value)),
        _ => {}
    }
}

impl Json {
    /// Parse `text` as JSON with upstream's default options. Returns an error
    /// if it is not valid JSON.
    pub fn new(text: &str) -> Result<Self> {
        Json::with_options(text, &JsonOptions::default())
    }

    /// Parse `text` as JSON and format it with `options`. Port of
    /// `rich.json.JSON(json, indent=…, highlight=…, …)`.
    pub fn with_options(text: &str, options: &JsonOptions) -> Result<Self> {
        let value = Parser::new(text).parse_document()?;
        // `json.dumps(allow_nan=False)` raises here, at construction.
        if !options.allow_nan {
            if let Some(literal) = first_non_finite(&value) {
                return Err(RichError::Json(format!(
                    "Out of range float values are not JSON compliant: {}",
                    match literal {
                        "NaN" => "nan",
                        "Infinity" => "inf",
                        _ => "-inf",
                    }
                )));
            }
        }
        Ok(Json {
            value,
            options: options.clone(),
            lines: std::sync::OnceLock::new(),
            no_wrap: false,
            #[cfg(feature = "json-escape-safe")]
            escape_safe: false,
        })
    }

    /// The formatted document, one highlighted `Text` per line.
    fn lines(&self) -> &[Text] {
        self.lines.get_or_init(|| {
            let dumped = dumps(&self.value, &self.options);
            let text = if !self.options.highlight {
                Text::new(dumped.json)
            } else if dumped.exact {
                let mut text = Text::new(dumped.json);
                let mut spans = dumped.spans;
                spans.extend(dumped.keys);
                text.set_spans(spans);
                text
            } else {
                json_highlight(dumped.json)
            };
            if text.plain().contains('\n') {
                text.split("\n", false, true)
            } else {
                vec![text]
            }
        })
    }

    /// The formatted, highlighted document. Port of the `JSON.text`
    /// attribute (upstream's `__rich__` returns it).
    pub fn text(&self) -> Text {
        match self.lines() {
            [line] => line.clone(),
            lines => Text::new("\n").join(lines),
        }
    }

    /// Opt in to escape-aware display boundaries (requires `json-escape-safe`).
    /// Cropping omits partial escapes. Folding preserves escapes that fit the
    /// width; narrower widths split oversized escapes to avoid losing content.
    /// The default remains Python rich's ordinary word folding/cropping.
    #[cfg(feature = "json-escape-safe")]
    pub fn escape_safe(mut self, enabled: bool) -> Self {
        self.escape_safe = enabled;
        self
    }

    /// Keep `rich.json.JSON`'s `self.text.no_wrap = True`, which **crops** each
    /// line at the available width instead of wrapping it.
    ///
    /// Defaults to `false`, because that is what a *top-level*
    /// `Console.print(JSON(...))` produces and that is how this renderable is
    /// normally reached. Upstream's flag really is set, but `Console.print`
    /// never renders the `Text` it is set on: `_collect_renderables` sees a
    /// `Text`, buffers it, and `check_text` hands back
    /// `Text(sep, justify=…, end=…).join(buffered)` — and `Text.join` starts
    /// from `self.blank_copy()`, i.e. from the *separator's* metadata. The
    /// separator has `no_wrap=None`, so the copy that actually renders wraps
    /// with the default `fold` overflow — or with the `no_wrap` and
    /// `overflow` that `print` was given, which reach it as
    /// [`ConsoleOptions`] and which this renderable honours while the flag is
    /// off.
    ///
    /// Turn it on whenever the document is **nested** inside another
    /// renderable — a `Panel`, `Padding`, `Styled`, `Constrain`, or rich-cli's
    /// `ForceWidth`. Those are `ConsoleRenderable`s, so `_collect_renderables`
    /// appends them untouched and the inner `Text` reaches
    /// `Text.__rich_console__` with the flag intact; `Text.wrap` then skips
    /// `divide_line` and only `truncate`s to the width.
    ///
    /// The two are not interchangeable. Wrapping keeps every character;
    /// cropping discards what does not fit, which for JSON means the printed
    /// document no longer parses — so the choice has to follow upstream's,
    /// not taste.
    #[must_use]
    pub fn no_wrap(mut self, no_wrap: bool) -> Self {
        self.no_wrap = no_wrap;
        self
    }

    /// The document's lines rendered without wrapping, joined by newlines.
    #[cfg(feature = "json-escape-safe")]
    fn render_unwrapped(&self, console: &Console) -> Vec<Segment> {
        let mut out = Vec::new();
        for (index, line) in self.lines().iter().enumerate() {
            if index > 0 {
                out.push(Segment::line());
            }
            out.extend(line.render(console.theme(), console.base_style()));
        }
        out
    }
}

/// A serialized document, with the spans `JSONHighlighter` would give it.
struct Dumped {
    json: String,
    /// Token spans (`json.brace`, `json.str`, …) in document order.
    spans: Vec<Span>,
    /// `json.key` spans, which upstream appends after every token span.
    keys: Vec<Span>,
    /// Whether `spans` + `keys` are exactly what the highlighter's regexes
    /// produce. They are for a whitespace indent and strings whose closing
    /// quote does not follow a backslash; otherwise the regexes run.
    exact: bool,
}

impl Dumped {
    fn styled(&mut self, token: &str, style: &'static str) {
        let start = self.json.len();
        self.json.push_str(token);
        self.spans.push(Span {
            start,
            end: self.json.len(),
            style: StyleType::Name(style.to_string()),
        });
    }

    /// A quoted string. `JSON_STR`'s lazy `".*?(?<!\\)"` ends at the first
    /// quote not preceded by a backslash, which is this string's own closing
    /// quote unless the string ends in an escaped backslash.
    fn string(&mut self, quoted: &str, key: bool) {
        if quoted.as_bytes().get(quoted.len().wrapping_sub(2)) == Some(&b'\\') {
            self.exact = false;
        }
        let start = self.json.len();
        self.styled(quoted, "json.str");
        if key {
            self.keys.push(Span {
                start,
                end: self.json.len(),
                style: StyleType::Name("json.key".to_string()),
            });
        }
    }
}

/// Serialize `value` as Python's `json.dumps` does with `options`, iteratively
/// so nesting depth costs heap rather than stack.
fn dumps(value: &Node, options: &JsonOptions) -> Dumped {
    use std::borrow::Cow;
    let indent = options.indent.as_deref();
    // `json.dumps` separators: `(', ', ': ')` on one line, `(',', ': ')` with
    // an indent.
    let item_separator = if indent.is_some() { "," } else { ", " };
    let newline_indent = |level: usize| -> Cow<'static, str> {
        match indent {
            Some(indent) => Cow::Owned(format!("\n{}", indent.repeat(level))),
            None => Cow::Borrowed(""),
        }
    };
    let quote_string = |string: &[u8]| -> String { quote_py(string, options.ensure_ascii) };

    /// One entry of the serializer's work list, consumed newest-first.
    enum Task<'a> {
        /// Serialize this value, `usize` levels deep.
        Value(&'a Node, usize),
        /// Write this unstyled text.
        Plain(Cow<'a, str>),
        /// Write a closing brace.
        Close(&'static str),
        /// Write an object key and its `": "`.
        Key(&'a [u8]),
    }

    let mut dumped = Dumped {
        json: String::new(),
        spans: Vec::new(),
        keys: Vec::new(),
        // The highlighter's patterns could match inside any other indent.
        exact: indent.is_none_or(|indent| indent.chars().all(|c| c == ' ' || c == '\t')),
    };
    let mut stack = vec![Task::Value(value, 0)];
    while let Some(task) = stack.pop() {
        let (node, level) = match task {
            Task::Plain(text) => {
                dumped.json.push_str(&text);
                continue;
            }
            Task::Close(brace) => {
                dumped.styled(brace, "json.brace");
                continue;
            }
            Task::Key(key) => {
                dumped.string(&quote_string(key), true);
                dumped.json.push_str(": ");
                continue;
            }
            Task::Value(node, level) => (node, level),
        };
        match node {
            Node::Null => dumped.styled("null", "json.null"),
            Node::Bool(true) => dumped.styled("true", "json.bool_true"),
            Node::Bool(false) => dumped.styled("false", "json.bool_false"),
            Node::Number(number) => dumped.styled(number, "json.number"),
            // No `JSONHighlighter` pattern matches `NaN` or the infinities.
            Node::NonFinite(literal) => dumped.json.push_str(literal),
            Node::Str(string) => dumped.string(&quote_string(string), false),
            Node::Array(items) => {
                dumped.styled("[", "json.brace");
                if items.is_empty() {
                    dumped.styled("]", "json.brace");
                    continue;
                }
                // Pushed back-to-front, so they pop in document order.
                stack.push(Task::Close("]"));
                stack.push(Task::Plain(newline_indent(level)));
                let last = items.len() - 1;
                for (index, item) in items.iter().enumerate().rev() {
                    if index != last {
                        stack.push(Task::Plain(Cow::Borrowed(item_separator)));
                    }
                    stack.push(Task::Value(item, level + 1));
                    stack.push(Task::Plain(newline_indent(level + 1)));
                }
            }
            Node::Object(entries) => {
                dumped.styled("{", "json.brace");
                if entries.is_empty() {
                    dumped.styled("}", "json.brace");
                    continue;
                }
                let mut order: Vec<&(PyStr, Node)> = entries.iter().collect();
                if options.sort_keys {
                    // `sorted(dct.items())`: keys are unique, so this is by key,
                    // in code point order (which UTF-8 byte order preserves).
                    order.sort_by(|a, b| a.0.cmp(&b.0));
                }
                stack.push(Task::Close("}"));
                stack.push(Task::Plain(newline_indent(level)));
                let last = order.len() - 1;
                for (index, (key, item)) in order.into_iter().enumerate().rev() {
                    if index != last {
                        stack.push(Task::Plain(Cow::Borrowed(item_separator)));
                    }
                    stack.push(Task::Value(item, level + 1));
                    stack.push(Task::Key(key));
                    stack.push(Task::Plain(newline_indent(level + 1)));
                }
            }
        }
    }
    dumped
}

/// The first non-finite literal in the document, if any, found iteratively.
fn first_non_finite(value: &Node) -> Option<&'static str> {
    let mut pending = vec![value];
    while let Some(node) = pending.pop() {
        match node {
            Node::NonFinite(literal) => return Some(literal),
            Node::Array(items) => pending.extend(items.iter()),
            Node::Object(entries) => pending.extend(entries.iter().map(|(_, value)| value)),
            _ => {}
        }
    }
    None
}

/// `ensure_ascii=True`: re-escape every character outside `' '..='~'` in an
/// already-quoted JSON string as `\uXXXX` (a surrogate pair above the BMP),
/// as `py_encode_basestring_ascii` does. The quoted form's own escapes are
/// ASCII already.
fn ascii_escape(quoted: &str) -> String {
    if quoted.bytes().all(|b| (b' '..=b'~').contains(&b)) {
        return quoted.to_string();
    }
    let mut out = String::with_capacity(quoted.len() + 16);
    for c in quoted.chars() {
        if (' '..='~').contains(&c) {
            out.push(c);
            continue;
        }
        let mut units = [0u16; 2];
        for unit in c.encode_utf16(&mut units) {
            out.push_str(&format!("\\u{unit:04x}"));
        }
    }
    out
}

/// Upstream's `JSONHighlighter` regexes: the combined token pattern, and the
/// string pattern its key pass re-scans with.
fn json_patterns() -> &'static (fancy_regex::Regex, fancy_regex::Regex) {
    static PATTERNS: std::sync::OnceLock<(fancy_regex::Regex, fancy_regex::Regex)> =
        std::sync::OnceLock::new();
    PATTERNS.get_or_init(|| {
        const JSON_STR: &str = r#"(?<![\\\w])(?P<str>b?".*?(?<!\\)")"#;
        let combined = [
            r"(?P<brace>[\{\[\(\)\]\}])",
            r"\b(?P<bool_true>true)\b|\b(?P<bool_false>false)\b|\b(?P<null>null)\b",
            r"(?P<number>(?<!\w)\-?[0-9]+\.?[0-9]*(e[\-\+]?\d+?)?\b|0x[0-9a-fA-F]*)",
            JSON_STR,
        ]
        .join("|");
        (
            fancy_regex::Regex::new(&combined).expect("valid JSON highlighter pattern"),
            fancy_regex::Regex::new(JSON_STR).expect("valid JSON string pattern"),
        )
    })
}

/// Highlight a formatted document. Port of `JSONHighlighter.highlight`: the
/// `json.`-prefixed token spans, then a `json.key` span on every string that
/// is followed (past whitespace) by a colon.
fn json_highlight(json: String) -> Text {
    let (combined, string) = json_patterns();
    let mut text = Text::new(json);
    text.highlight_with_regex(combined, None, "json.");
    let plain = text.plain();
    let mut keys = Vec::new();
    for found in string.find_iter(plain) {
        let Ok(found) = found else { break };
        for &byte in &plain.as_bytes()[found.end()..] {
            match byte {
                b':' => keys.push((found.start(), found.end())),
                b' ' | b'\n' | b'\r' | b'\t' => continue,
                _ => {}
            }
            break;
        }
    }
    for (start, end) in keys {
        text.push_span(Span {
            start,
            end,
            style: StyleType::Name("json.key".to_string()),
        });
    }
    text
}

impl Renderable for Json {
    /// Upstream `JSON.__rich__` returns its highlighted `Text`, so measuring a
    /// `JSON` is `Text.__rich_measure__` over the formatted document.
    fn measure(&self, _console: &Console, _options: &ConsoleOptions) -> Measurement {
        let mut minimum = 0;
        let mut maximum = 0;
        for line in self.lines() {
            let (line_minimum, line_maximum) = line.measurement();
            minimum = minimum.max(line_minimum);
            maximum = maximum.max(line_maximum);
        }
        Measurement::new(minimum, maximum)
    }

    /// `Text.__rich_console__` of the document: each line wraps (or, with
    /// `no_wrap`, is truncated) to the width with the options' `justify` and
    /// `overflow` — the document's own `overflow` is `None`.
    fn rich_render(&self, console: &Console, options: &ConsoleOptions) -> Vec<Segment> {
        #[cfg(feature = "json-escape-safe")]
        if self.escape_safe {
            let segments = self.render_unwrapped(console);
            return escape_safe_lines(&segments, options.max_width, self.no_wrap);
        }
        let no_wrap = self.no_wrap || options.no_wrap.unwrap_or(false);
        let overflow = options.overflow.unwrap_or_default();
        let tab_size = match console.tab_size() {
            0 => crate::text::DEFAULT_TAB_SIZE,
            tab_size => tab_size,
        };
        let mut out = Vec::new();
        let mut first = true;
        for line in self.lines() {
            for rendered in line.render_lines_wrapped_tabs(
                console.theme(),
                console.base_style(),
                Some(options.max_width),
                options.justify,
                overflow,
                no_wrap,
                tab_size,
            ) {
                if !first {
                    out.push(Segment::line());
                }
                first = false;
                out.extend(rendered);
            }
        }
        out
    }
}

/// Tokenize each physical line into JSON escapes and ordinary graphemes, then
/// choose every boundary from the space actually remaining. No stale absolute
/// wrap points survive an adjusted escape boundary (#98).
#[cfg(feature = "json-escape-safe")]
fn escape_safe_lines(segments: &[Segment], width: usize, crop: bool) -> Vec<Segment> {
    if width == 0 {
        return Vec::new();
    }
    let lines = Segment::split_lines(segments);
    let last = lines.len().saturating_sub(1);
    let mut out = Vec::new();
    for (line_index, line) in lines.into_iter().enumerate() {
        let plain: String = line
            .iter()
            .filter(|s| !s.control)
            .map(|s| s.text.as_str())
            .collect();
        // Keep upstream's exact behavior when there is no escape to protect.
        if !plain.contains('\\') {
            out.extend(if crop {
                Segment::crop_lines(&line, width)
            } else {
                Segment::fold_lines_words(&line, width)
            });
        } else {
            let (spans, _) = crate::cells::split_graphemes(&plain);
            let mut atoms = Vec::new();
            let mut index = 0;
            while index < spans.len() {
                let (start, mut end, mut cells) = spans[index];
                if plain.as_bytes()[start] == b'\\' {
                    let escape_end = start
                        + if plain.as_bytes().get(start + 1) == Some(&b'u') {
                            6
                        } else {
                            2
                        };
                    while end < escape_end && index + 1 < spans.len() {
                        index += 1;
                        end = spans[index].1;
                        cells += spans[index].2;
                    }
                    if !crop && cells > width {
                        // An atom wider than the whole console cannot both fit
                        // and stay atomic. Split its ASCII spelling rather than
                        // overrun into Console's final crop and lose characters.
                        for offset in start..escape_end {
                            atoms.push((offset, offset + 1, 1));
                        }
                        if end > escape_end {
                            atoms.push((escape_end, end, 0));
                        }
                        index += 1;
                        continue;
                    }
                }
                atoms.push((start, end, cells));
                index += 1;
            }
            let mut breaks = Vec::new();
            let mut cells = 0;
            let mut stop = plain.len();
            for (start, _, atom_width) in atoms {
                if cells + atom_width > width {
                    if crop {
                        stop = start;
                        break;
                    }
                    if cells > 0 {
                        breaks.push(start);
                        cells = 0;
                    }
                }
                cells += atom_width;
            }
            let mut position = 0;
            let mut next = 0;
            for segment in line {
                if segment.control {
                    out.push(segment);
                    continue;
                }
                let mut buffer = String::new();
                for ch in segment.text.chars() {
                    if position >= stop {
                        break;
                    }
                    if breaks.get(next) == Some(&position) {
                        if !buffer.is_empty() {
                            out.push(Segment::new(
                                std::mem::take(&mut buffer),
                                segment.style.clone(),
                            ));
                        }
                        out.push(Segment::line());
                        next += 1;
                    }
                    buffer.push(ch);
                    position += ch.len_utf8();
                }
                if !buffer.is_empty() {
                    out.push(Segment::new(buffer, segment.style));
                }
            }
        }
        if line_index != last {
            out.push(Segment::line());
        }
    }
    out
}

/// Serialize a string as a JSON string literal (quoted + escaped).
/// `json.dumps` of a [`PyStr`]. A lone surrogate is `\udXXX` under
/// `ensure_ascii`; otherwise Python emits the surrogate itself, which a Rust
/// string cannot hold, so it becomes U+FFFD (see docs/DIVERGENCES.md).
fn quote_py(string: &[u8], ensure_ascii: bool) -> String {
    let escape = |run: &str| -> String {
        let quoted = quote(run);
        if ensure_ascii {
            ascii_escape(&quoted)
        } else {
            quoted
        }
    };
    if let Ok(run) = std::str::from_utf8(string) {
        return escape(run);
    }
    let mut out = String::from("\"");
    let mut rest = string;
    while !rest.is_empty() {
        let valid = match std::str::from_utf8(rest) {
            Ok(_) => rest.len(),
            Err(error) => error.valid_up_to(),
        };
        let (run, tail) = rest.split_at(valid);
        let quoted = escape(std::str::from_utf8(run).unwrap_or_default());
        out.push_str(&quoted[1..quoted.len() - 1]);
        rest = tail;
        if let [0xED, b1 @ 0xA0..=0xBF, b2, tail @ ..] = rest {
            let unit = 0xD000 | (u32::from(b1 & 0x3F) << 6) | u32::from(b2 & 0x3F);
            if ensure_ascii {
                out.push_str(&format!("\\u{unit:04x}"));
            } else {
                out.push(char::REPLACEMENT_CHARACTER);
            }
            rest = tail;
        } else if !rest.is_empty() {
            // Unreachable for parser output; never loop forever.
            out.push(char::REPLACEMENT_CHARACTER);
            rest = &rest[1..];
        }
    }
    out.push('"');
    out
}

/// Decode a JSON string token as Python's `json` does, keeping unpaired
/// `\uD800`-`\uDFFF` escapes as surrogates in the result. `None` when the
/// token is invalid for another reason (the caller reports `serde_json`'s
/// error for it).
fn decode_with_surrogates(token: &str) -> Option<PyStr> {
    let inner = token.strip_prefix('"')?.strip_suffix('"')?;
    let mut out = Vec::with_capacity(inner.len());
    let mut chars = inner.chars().peekable();
    let hex4 = |chars: &mut std::iter::Peekable<std::str::Chars<'_>>| -> Option<u32> {
        let mut value = 0;
        for _ in 0..4 {
            value = value * 16 + chars.next()?.to_digit(16)?;
        }
        Some(value)
    };
    while let Some(c) = chars.next() {
        if c < ' ' {
            return None;
        }
        if c != '\\' {
            let mut buf = [0u8; 4];
            out.extend_from_slice(c.encode_utf8(&mut buf).as_bytes());
            continue;
        }
        let decoded = match chars.next()? {
            '"' => '"',
            '\\' => '\\',
            '/' => '/',
            'b' => '\u{8}',
            'f' => '\u{c}',
            'n' => '\n',
            'r' => '\r',
            't' => '\t',
            'u' => {
                let mut unit = hex4(&mut chars)?;
                if (0xD800..0xDC00).contains(&unit) {
                    let mut lookahead = chars.clone();
                    if lookahead.next() == Some('\\') && lookahead.next() == Some('u') {
                        if let Some(low @ 0xDC00..=0xDFFF) = hex4(&mut lookahead) {
                            unit = 0x10000 + ((unit - 0xD800) << 10) + (low - 0xDC00);
                            chars = lookahead;
                        }
                    }
                }
                match char::from_u32(unit) {
                    Some(c) => c,
                    None => {
                        // A lone surrogate: its generalized UTF-8 encoding.
                        out.push(0xE0 | (unit >> 12) as u8);
                        out.push(0x80 | ((unit >> 6) & 0x3F) as u8);
                        out.push(0x80 | (unit & 0x3F) as u8);
                        continue;
                    }
                }
            }
            _ => return None,
        };
        let mut buf = [0u8; 4];
        out.extend_from_slice(decoded.encode_utf8(&mut buf).as_bytes());
    }
    Some(out)
}

fn quote(string: &str) -> String {
    serde_json::to_string(string).unwrap_or_else(|_| format!("{string:?}"))
}

/// A container being filled in, held on the parser's explicit stack.
enum Frame {
    Array(Vec<Node>),
    Object {
        entries: Vec<(PyStr, Node)>,
        /// Key -> position in `entries`, so a repeated key overwrites in place
        /// (`{"a": 1, "a": 2}` is one entry) without an O(n^2) rescan. Dropped
        /// with the frame, so only the objects on the current path pay for it.
        seen: HashMap<PyStr, usize>,
        /// The key whose value is currently being parsed.
        key: PyStr,
    },
}

/// A non-recursive JSON reader. Structure is walked with an explicit stack;
/// scalar tokens are handed to `serde_json` for decoding so escapes, number
/// formats and their rejections stay identical to the rest of the workspace.
struct Parser<'a> {
    src: &'a str,
    bytes: &'a [u8],
    pos: usize,
}

impl<'a> Parser<'a> {
    fn new(src: &'a str) -> Self {
        Parser {
            src,
            bytes: src.as_bytes(),
            pos: 0,
        }
    }

    fn parse_document(&mut self) -> Result<Node> {
        let value = self.parse_value()?;
        self.skip_whitespace();
        if self.pos != self.bytes.len() {
            return Err(self.error("trailing characters"));
        }
        Ok(value)
    }

    /// Parse one value, descending into containers with an explicit stack.
    fn parse_value(&mut self) -> Result<Node> {
        let mut stack: Vec<Frame> = Vec::new();
        let mut node: Node;

        'descend: loop {
            self.skip_whitespace();
            // CPython's scanner enters a recursive call per container, empty
            // ones included, and raises once the recursion limit is reached.
            if let Some(kind @ (b'[' | b'{')) = self.peek() {
                if stack.len() >= MAX_DEPTH {
                    let what = if kind == b'[' { "array" } else { "object" };
                    return Err(self.error(&format!(
                        "maximum recursion depth exceeded while decoding a JSON {what} \
                         from a unicode string"
                    )));
                }
            }
            match self.peek() {
                Some(b'[') => {
                    self.pos += 1;
                    self.skip_whitespace();
                    if self.peek() == Some(b']') {
                        self.pos += 1;
                        node = Node::Array(Vec::new());
                    } else {
                        stack.push(Frame::Array(Vec::new()));
                        continue 'descend;
                    }
                }
                Some(b'{') => {
                    self.pos += 1;
                    self.skip_whitespace();
                    if self.peek() == Some(b'}') {
                        self.pos += 1;
                        node = Node::Object(Vec::new());
                    } else {
                        let key = self.parse_key()?;
                        stack.push(Frame::Object {
                            entries: Vec::new(),
                            seen: HashMap::new(),
                            key,
                        });
                        continue 'descend;
                    }
                }
                _ => node = self.parse_scalar()?,
            }

            // `node` is finished: hand it to its parent, then close as many
            // containers as end here.
            loop {
                let Some(frame) = stack.last_mut() else {
                    return Ok(node);
                };
                let closing = match frame {
                    Frame::Array(items) => {
                        items.push(node);
                        b']'
                    }
                    Frame::Object { entries, seen, key } => {
                        let key = std::mem::take(key);
                        match seen.get(&key) {
                            Some(&at) => entries[at].1 = node,
                            None => {
                                seen.insert(key.clone(), entries.len());
                                entries.push((key, node));
                            }
                        }
                        b'}'
                    }
                };
                self.skip_whitespace();
                match self.peek() {
                    Some(b',') => {
                        self.pos += 1;
                        if closing == b'}' {
                            let next_key = self.parse_key()?;
                            if let Some(Frame::Object { key, .. }) = stack.last_mut() {
                                *key = next_key;
                            }
                        }
                        continue 'descend;
                    }
                    Some(byte) if byte == closing => {
                        self.pos += 1;
                        node = match stack.pop() {
                            Some(Frame::Array(items)) => Node::Array(items),
                            Some(Frame::Object { entries, .. }) => Node::Object(entries),
                            None => unreachable!("the frame was just borrowed"),
                        };
                    }
                    _ if closing == b']' => return Err(self.error("expected `,` or `]`")),
                    _ => return Err(self.error("expected `,` or `}`")),
                }
            }
        }
    }

    /// Parse `"key" :`, leaving the parser on the value.
    fn parse_key(&mut self) -> Result<PyStr> {
        self.skip_whitespace();
        if self.peek() != Some(b'"') {
            return Err(self.error("key must be a string"));
        }
        let key = self.parse_string()?;
        self.skip_whitespace();
        if self.peek() != Some(b':') {
            return Err(self.error("expected `:`"));
        }
        self.pos += 1;
        Ok(key)
    }

    fn parse_scalar(&mut self) -> Result<Node> {
        match self.peek() {
            Some(b'"') => Ok(Node::Str(self.parse_string()?)),
            Some(b't') => {
                self.expect_literal("true")?;
                Ok(Node::Bool(true))
            }
            Some(b'f') => {
                self.expect_literal("false")?;
                Ok(Node::Bool(false))
            }
            Some(b'n') => {
                self.expect_literal("null")?;
                Ok(Node::Null)
            }
            // Python's json emits and accepts these three (`allow_nan=True` is
            // the default both ways), so upstream renders documents containing
            // them instead of rejecting the file.
            Some(b'N') => {
                self.expect_literal("NaN")?;
                Ok(Node::NonFinite("NaN"))
            }
            Some(b'I') => {
                self.expect_literal("Infinity")?;
                Ok(Node::NonFinite("Infinity"))
            }
            Some(b'-') if self.src[self.pos..].starts_with("-Infinity") => {
                self.pos += "-Infinity".len();
                Ok(Node::NonFinite("-Infinity"))
            }
            Some(b'-' | b'0'..=b'9') => self.parse_number(),
            Some(_) => Err(self.error("expected value")),
            None => Err(self.error("EOF while parsing a value")),
        }
    }

    /// Read a string token and decode it with `serde_json`, so escapes, lone
    /// surrogates and raw control characters behave exactly as before.
    fn parse_string(&mut self) -> Result<PyStr> {
        let start = self.pos;
        let mut end = self.pos + 1;
        loop {
            match self.bytes.get(end) {
                None => return Err(self.error_at(self.bytes.len(), "EOF while parsing a string")),
                Some(b'"') => {
                    end += 1;
                    break;
                }
                Some(b'\\') => {
                    end += 1;
                    // Step over the escaped character whole. A multi-byte
                    // character after a backslash is invalid JSON, but `end`
                    // must still land on a UTF-8 boundary or slicing panics
                    // before `serde_json` gets to reject it.
                    match self.src[end..].chars().next() {
                        Some(ch) => end += ch.len_utf8(),
                        None => {
                            return Err(
                                self.error_at(self.bytes.len(), "EOF while parsing a string")
                            )
                        }
                    }
                }
                // Continuation bytes are never `"` or `\`, so scanning byte by
                // byte cannot mistake one for a delimiter.
                Some(_) => end += 1,
            }
        }
        let token = &self.src[start..end];
        let decoded = match serde_json::from_str::<String>(token) {
            Ok(decoded) => decoded.into_bytes(),
            // `serde_json` refuses lone surrogates, which Python's `json`
            // accepts; anything else it refuses stays an error.
            Err(error) => decode_with_surrogates(token)
                .ok_or_else(|| self.error_at(start, &describe(&error)))?,
        };
        self.pos = end;
        Ok(decoded)
    }

    /// Validate JSON's number grammar before decoding, preserving arbitrary-size
    /// integers and Python's float overflow to Infinity (#74).
    fn parse_number(&mut self) -> Result<Node> {
        let start = self.pos;
        let mut end = start;
        while matches!(
            self.bytes.get(end),
            Some(b'-' | b'+' | b'.' | b'e' | b'E' | b'0'..=b'9')
        ) {
            end += 1;
        }
        let token = &self.src[start..end];
        let digits = token.as_bytes();
        let mut i = usize::from(digits.first() == Some(&b'-'));
        if digits.get(i) == Some(&b'0') {
            i += 1;
        } else {
            let first = i;
            while digits.get(i).is_some_and(u8::is_ascii_digit) {
                i += 1;
            }
            if i == first {
                return Err(self.error_at(start, "invalid number"));
            }
        }
        let mut floating = false;
        if digits.get(i) == Some(&b'.') {
            floating = true;
            i += 1;
            let first = i;
            while digits.get(i).is_some_and(u8::is_ascii_digit) {
                i += 1;
            }
            if i == first {
                return Err(self.error_at(start, "invalid number"));
            }
        }
        if matches!(digits.get(i), Some(b'e' | b'E')) {
            floating = true;
            i += 1;
            if matches!(digits.get(i), Some(b'+' | b'-')) {
                i += 1;
            }
            let first = i;
            while digits.get(i).is_some_and(u8::is_ascii_digit) {
                i += 1;
            }
            if i == first {
                return Err(self.error_at(start, "invalid number"));
            }
        }
        if i != digits.len() {
            return Err(self.error_at(start, "invalid number"));
        }
        self.pos = end;
        if !floating {
            return Ok(Node::Number(
                if token == "-0" { "0" } else { token }.to_string(),
            ));
        }
        let value: f64 = token
            .parse()
            .map_err(|_| self.error_at(start, "invalid number"))?;
        if value.is_infinite() {
            return Ok(Node::NonFinite(if value.is_sign_negative() {
                "-Infinity"
            } else {
                "Infinity"
            }));
        }
        // `json.dumps` writes a float with `float.__repr__`.
        Ok(Node::Number(crate::pyformat::float_repr(value)))
    }

    fn expect_literal(&mut self, literal: &str) -> Result<()> {
        if self.src[self.pos..].starts_with(literal) {
            self.pos += literal.len();
            Ok(())
        } else {
            Err(self.error("expected value"))
        }
    }

    fn peek(&self) -> Option<u8> {
        self.bytes.get(self.pos).copied()
    }

    fn skip_whitespace(&mut self) {
        while matches!(self.peek(), Some(b' ' | b'\t' | b'\n' | b'\r')) {
            self.pos += 1;
        }
    }

    fn error(&self, message: &str) -> RichError {
        self.error_at(self.pos, message)
    }

    fn error_at(&self, pos: usize, message: &str) -> RichError {
        let (line, column) = self.line_column(pos);
        RichError::Json(format!("{message} at line {line} column {column}"))
    }

    fn line_column(&self, pos: usize) -> (usize, usize) {
        let mut pos = pos.min(self.src.len());
        while !self.src.is_char_boundary(pos) {
            pos -= 1;
        }
        let before = &self.src[..pos];
        let line = 1 + before.matches('\n').count();
        let column = before
            .rsplit('\n')
            .next()
            .map_or(0, |tail| tail.chars().count())
            + 1;
        (line, column)
    }
}

/// `serde_json`'s message without its own `at line … column …` suffix, which
/// counts from the start of the token slice rather than the document.
fn describe(error: &serde_json::Error) -> String {
    let text = error.to_string();
    match text.find(" at line ") {
        Some(at) => text[..at].to_string(),
        None => text,
    }
}

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

    /// The spans recorded while serializing are exactly the ones upstream's
    /// `JSONHighlighter` regexes find, wherever `dumps` claims they are.
    #[test]
    fn recorded_spans_match_the_highlighter_regexes() {
        let documents = [
            r#"{"name": "Alice", "age": 30, "admin": true, "tags": ["a", "b"], "meta": null}"#,
            r#"[1e20, -0.0, 1.5e-7, 12345678901234567890, false, [], {}, [[]], {"a": {}}]"#,
            r#"{"true": "false", "x:y": "a\"b", "q\"": ": null", "n": [NaN, Infinity, -Infinity]}"#,
            r#"{"caf\u00e9": "\u2764 \ud83d\ude00", "ctl": "\u0001\t\n", "": ""}"#,
            r#"["[{(1)}]", "0x1F", "b\"x", "\\\"", "tail\\x"]"#,
        ];
        let indents = [None, Some("  "), Some("\t"), Some("")];
        for document in documents {
            for indent in indents {
                for ensure_ascii in [false, true] {
                    let options = JsonOptions {
                        indent: indent.map(str::to_string),
                        ensure_ascii,
                        sort_keys: true,
                        ..JsonOptions::default()
                    };
                    let value = Parser::new(document).parse_document().unwrap();
                    let dumped = dumps(&value, &options);
                    assert!(dumped.exact, "{document} / {indent:?}");
                    let mut recorded = dumped.spans.clone();
                    recorded.extend(dumped.keys.clone());
                    let expected = json_highlight(dumped.json.clone());
                    assert_eq!(recorded, expected.spans(), "{document} / {indent:?}");
                }
            }
        }
        // A string ending in a backslash defeats `JSON_STR`, so the regexes run.
        let value = Parser::new(r#"{"k": "a\\"}"#).parse_document().unwrap();
        assert!(!dumps(&value, &JsonOptions::default()).exact);
        let value = Parser::new("[1]").parse_document().unwrap();
        let options = JsonOptions {
            indent: Some("1".to_string()),
            ..JsonOptions::default()
        };
        assert!(!dumps(&value, &options).exact);
    }

    fn render(text: &str) -> String {
        let console = Console::builder()
            .force_terminal(true)
            .color_system(Some(ColorSystem::Truecolor))
            .width(40)
            .build();
        console.render_to_string(&Json::new(text).unwrap())
    }

    fn render_plain(text: &str, width: usize) -> String {
        let console = Console::builder().width(width).color_system(None).build();
        console.render_to_string(&Json::new(text).expect("valid json"))
    }

    #[test]
    fn empty_collections_stay_inline() {
        assert_eq!(render("{}"), "\x1b[1m{\x1b[0m\x1b[1m}\x1b[0m");
        assert_eq!(render("[]"), "\x1b[1m[\x1b[0m\x1b[1m]\x1b[0m");
    }

    #[test]
    fn object_with_scalars() {
        assert_eq!(
            render(r#"{"ok": false}"#),
            "\x1b[1m{\x1b[0m\n  \x1b[1;34m\"ok\"\x1b[0m: \x1b[3;91mfalse\x1b[0m\n\x1b[1m}\x1b[0m"
        );
    }

    #[test]
    fn invalid_json_errors() {
        assert!(Json::new("{not json}").is_err());
    }

    #[test]
    fn non_ascii_stays_utf8_in_input_order() {
        // Upstream's JSON defaults to ensure_ascii=False, so accented/symbol
        // characters render as UTF-8 (not \uXXXX), and keys keep input order.
        // (Byte-parity is guaranteed by the `json_unicode` golden.)
        let out = render("{\"name\": \"caf\u{e9}\", \"emoji\": \"\u{2764}\"}");
        assert!(out.contains("caf\u{e9}"), "café stays UTF-8: {out:?}");
        assert!(out.contains('\u{2764}'), "heart stays UTF-8");
        let name_at = out.find("name").expect("name key present");
        let emoji_at = out.find("emoji").expect("emoji key present");
        assert!(name_at < emoji_at, "keys keep input order");
    }

    #[test]
    fn a_long_value_is_wrapped_rather_than_cropped() {
        // A long string value used to be cut mid-token, so the printed document
        // was missing data -- and, for JSON, no longer parseable -- at exit 0.
        let payload = format!("{{\"k\": \"{}\"}}", "y".repeat(120));
        let out = render_plain(&payload, 40);
        assert_eq!(
            out.matches('y').count(),
            120,
            "characters were dropped:
{out}"
        );
    }

    /// Upstream wraps the JSON at word boundaries: `JSON.text` asks for
    /// `no_wrap`, but `Console.print` re-joins it through `Text(sep).join(...)`
    /// and the joined copy carries neither the flag nor an overflow, so the
    /// default `fold` word wrap applies. Character folding split words
    /// (`over t` / `he lazy`), which no upstream output ever shows.
    ///
    /// Captured from rich 15.0.0:
    /// `Console(width=40).print(JSON(...))`.
    #[test]
    fn wrapping_breaks_at_word_boundaries() {
        let payload = r#"{"k": "the quick brown fox jumps over the lazy dog and keeps running for a very long time indeed"}"#;
        assert_eq!(
            render_plain(payload, 40),
            "{\n  \"k\": \"the quick brown fox jumps over \n\
             the lazy dog and keeps running for a \n\
             very long time indeed\"\n}"
        );
    }

    #[cfg(feature = "json-escape-safe")]
    #[test]
    fn adjusted_escape_boundaries_preserve_the_remaining_payload() {
        let input = format!(r#"{{"v":"aa\u0001{}"}}"#, "b".repeat(40));
        for width in 6..=20 {
            let output = Console::builder()
                .width(width)
                .force_terminal(false)
                .build()
                .render_to_string(&Json::new(&input).unwrap().escape_safe(true));
            assert_eq!(output.matches('b').count(), 40, "width {width}: {output:?}");
        }
    }

    #[cfg(feature = "json-escape-safe")]
    #[test]
    fn wrapping_keeps_json_escapes_atomic_at_narrow_widths() {
        let payload = r#"{"v":"a\"b\\c\nd\u0001e"}"#;
        for width in 8..=14 {
            let output = Console::builder()
                .width(width)
                .force_terminal(false)
                .build()
                .render_to_string(&Json::new(payload).unwrap().escape_safe(true));
            for line in output.lines() {
                let bytes = line.as_bytes();
                let mut index = 0;
                while index < bytes.len() {
                    if bytes[index] != b'\\' {
                        index += 1;
                        continue;
                    }
                    assert!(
                        index + 1 < bytes.len(),
                        "split escape at width {width}: {output:?}"
                    );
                    if bytes[index + 1] == b'u' {
                        assert!(
                            index + 6 <= bytes.len(),
                            "split unicode escape at width {width}: {output:?}"
                        );
                        index += 6;
                    } else {
                        index += 2;
                    }
                }
            }
        }
    }

    #[cfg(feature = "json-escape-safe")]
    #[test]
    fn escape_folding_preserves_bytes_even_below_the_escape_width() {
        let payload = r#"{"v":"a\"b\\c\nd\u0001eeeeeeeeeeee"}"#;
        let wide = Console::builder()
            .width(100)
            .force_terminal(false)
            .build()
            .render_to_string(&Json::new(payload).unwrap().escape_safe(true))
            .replace('\n', "");
        for width in 1..=20 {
            let output = Console::builder()
                .width(width)
                .force_terminal(false)
                .build()
                .render_to_string(&Json::new(payload).unwrap().escape_safe(true));
            assert_eq!(output.replace('\n', ""), wide, "width {width}");
            assert!(output
                .lines()
                .all(|line| crate::cells::cell_len(line) <= width));
        }
    }

    #[cfg(feature = "json-escape-safe")]
    #[test]
    fn escape_cropping_never_emits_a_partial_escape() {
        let payload = r#""a\"b\\c\nd\u0001eeee""#;
        for width in 1..=24 {
            let output = Console::builder()
                .width(width)
                .force_terminal(false)
                .build()
                .render_to_string(&Json::new(payload).unwrap().no_wrap(true).escape_safe(true));
            let mut chars = output.chars();
            while let Some(c) = chars.next() {
                if c == '\\' {
                    let next = chars.next().expect("complete short escape");
                    if next == 'u' {
                        for _ in 0..4 {
                            assert!(chars.next().is_some_and(|c| c.is_ascii_hexdigit()));
                        }
                    }
                }
            }
        }
    }

    /// Nested inside another renderable, `JSON.text.no_wrap` survives and each
    /// line is **cropped** at the width rather than wrapped — see
    /// [`Json::no_wrap`]. Wrapping here instead was silent content loss: with
    /// `rich -j doc.json -w 120` on an 80-column console the document was laid
    /// out at 120 and then cropped to 80 by `Console.print`, so whole runs
    /// vanished and the surviving text read as if it were contiguous.
    ///
    /// Captured from rich-cli 1.8.1 driven by rich 15.0.0:
    /// `COLUMNS=80 rich -j long.json -w 40`.
    #[test]
    fn a_nested_document_is_cropped_rather_than_wrapped() {
        let payload = r#"{"k": "the quick brown fox jumps over the lazy dog and keeps running for a very long time indeed"}"#;
        let console = Console::builder().width(40).color_system(None).build();
        let json = Json::new(payload).expect("valid json").no_wrap(true);
        assert_eq!(
            console.render_to_string(&json),
            "{\n  \"k\": \"the quick brown fox jumps over t\n}"
        );

        // …and the wrap is still the default, because a bare
        // `Console.print(JSON(...))` loses the flag in `Text.join`.
        assert_eq!(
            render_plain(payload, 40),
            "{\n  \"k\": \"the quick brown fox jumps over \n\
             the lazy dog and keeps running for a \n\
             very long time indeed\"\n}"
        );
    }

    /// A crop must not cut a double-width character in half: `set_cell_size`
    /// drops the straddling character and the line comes out one cell short,
    /// never one cell over.
    #[test]
    fn cropping_never_splits_a_wide_character() {
        let console = Console::builder().width(12).color_system(None).build();
        let json = Json::new("{\"k\": \"\u{1f306}\u{1f306}\u{1f306}\"}")
            .expect("valid json")
            .no_wrap(true);
        for line in console.render_to_string(&json).lines() {
            assert!(
                crate::cells::cell_len(line) <= 12,
                "line {line:?} overflows the crop"
            );
        }
    }

    /// serde_json's default float parser takes a fast path that can land 1 ULP
    /// from the value in the file, so the rendered number parsed back to a
    /// *different* double. The `float_roundtrip` feature makes parsing exact.
    #[test]
    fn floats_round_trip_exactly() {
        for literal in [
            "-938371.9565467801",
            "0.1",
            "1.7976931348623157e308",
            "5e-324",
            "3.141592653589793",
        ] {
            let out = render_plain(&format!("{{\"v\": {literal}}}"), 120);
            let rendered: String = out
                .split(':')
                .nth(1)
                .expect("a value after the key")
                .trim()
                .trim_end_matches(['}', ' ', '\n'])
                .to_string();
            let want: f64 = literal.parse().expect("literal parses");
            let got: f64 = rendered
                .parse()
                .unwrap_or_else(|_| panic!("rendered {rendered:?}"));
            assert_eq!(
                got.to_bits(),
                want.to_bits(),
                "{literal} rendered as {rendered} — a different double"
            );
        }
    }

    /// serde_json stops at 128 levels, so a 200-deep document — which CPython
    /// parses without complaint — was reported as invalid JSON and the CLI
    /// exited 1 on a file upstream renders.
    #[test]
    fn deep_nesting_is_not_rejected() {
        for depth in [128, 129, 200, 1000] {
            let payload = format!("{}1{}", "[".repeat(depth), "]".repeat(depth));
            let json = Json::new(&payload)
                .unwrap_or_else(|error| panic!("depth {depth} rejected: {error}"));
            let console = Console::builder()
                .width(4 * depth + 8)
                .color_system(None)
                .build();
            let out = console.render_to_string(&json);
            assert_eq!(
                out.matches('[').count(),
                depth,
                "depth {depth} did not render every level"
            );
        }
    }

    /// Parsing and dropping must cost heap, not stack: a recursive parser (or
    /// the compiler's recursive drop glue) turns a deep document into a stack
    /// overflow, which kills the process without even an error message.
    #[test]
    fn very_deep_nesting_does_not_overflow_the_stack() {
        let payload = format!("{}1{}", "[".repeat(MAX_DEPTH), "]".repeat(MAX_DEPTH));
        let json = Json::new(&payload).expect("deep document parses");
        drop(json);
        // Drop glue on a tree far past the parse limit stays iterative too.
        let mut node = Node::Number("1".to_string());
        for _ in 0..100_000 {
            node = Node::Array(vec![node]);
        }
        drop(node);
    }

    /// Past `MAX_DEPTH` the document is refused with CPython's
    /// `RecursionError` message instead of being rendered: its indentation
    /// alone is quadratic in the depth, and a 100 000-deep array exhausted
    /// memory and was killed.
    #[test]
    fn nesting_past_the_limit_is_an_error() {
        for (open, close, what) in [("[", "]", "array"), ("{\"a\":", "}", "object")] {
            for depth in [MAX_DEPTH + 1, 100_000] {
                let payload = format!("{}1{}", open.repeat(depth), close.repeat(depth));
                let error = Json::new(&payload).err().expect("too deep");
                assert!(
                    error.to_string().contains(&format!(
                        "maximum recursion depth exceeded while decoding a JSON {what}"
                    )),
                    "{error}"
                );
            }
        }
        // An empty container counts as a level, as a call in CPython's scanner.
        let payload = format!("{}[]{}", "[".repeat(MAX_DEPTH), "]".repeat(MAX_DEPTH));
        assert!(Json::new(&payload).is_err());
        let payload = format!(
            "{}[]{}",
            "[".repeat(MAX_DEPTH - 1),
            "]".repeat(MAX_DEPTH - 1)
        );
        assert!(Json::new(&payload).is_ok());
    }

    /// Python's json accepts and emits `NaN` / `Infinity` / `-Infinity`
    /// (`allow_nan=True` is the default), and rich's JSONHighlighter has no
    /// rule that matches them, so upstream prints them *unstyled*. serde_json
    /// rejected the whole document.
    ///
    /// Captured from rich 15.0.0:
    /// `Console(width=40, force_terminal=True).print(JSON(...))`.
    #[test]
    fn non_finite_numbers_render_unstyled() {
        assert_eq!(
            render(r#"{"a": NaN, "b": Infinity, "c": -Infinity, "d": 1.5}"#),
            "\x1b[1m{\x1b[0m\n  \x1b[1;34m\"a\"\x1b[0m: NaN,\n  \x1b[1;34m\"b\"\x1b[0m: \
             Infinity,\n  \x1b[1;34m\"c\"\x1b[0m: -Infinity,\n  \x1b[1;34m\"d\"\x1b[0m: \
             \x1b[1;36m1.5\x1b[0m\n\x1b[1m}\x1b[0m"
        );
        // Python spells them with those exact capitalisations and nothing else.
        for rejected in [r#"{"a": nan}"#, r#"{"a": inf}"#, r#"{"a": -inf}"#] {
            assert!(Json::new(rejected).is_err(), "{rejected} should not parse");
        }
    }

    /// The hand-written reader must accept and reject exactly what serde_json
    /// does for bounded numbers — Python also accepts overflowing exponents.
    #[test]
    fn acceptance_matches_serde_json() {
        let samples = [
            "{}",
            "[]",
            "  {\t\"a\" :\n1 }  ",
            r#"{"a": 1, "a": 2}"#,
            r#"{"a": [1, {"b": null}], "c": "x"}"#,
            "0",
            "-0",
            "0.0",
            "1e10",
            "1E+10",
            "1e-7",
            "12345678901234567890",
            "-12345678901234567890123456789012345",
            "01",
            "1.",
            ".1",
            "+1",
            "1e",
            "-",
            "--1",
            "-i",
            "-Inf",
            "Infinit",
            "NAN",
            "1 2",
            "",
            "   ",
            "{",
            "[",
            "]",
            "}",
            "[,]",
            "[1,]",
            r#"{"a": 1,}"#,
            r#"{a: 1}"#,
            r#"{'a': 1}"#,
            r#"{"a" 1}"#,
            "[1 2]",
            "truex",
            "tru",
            "nul",
            r#""unterminated"#,
            r#""\q""#,
            r#""é""#,
            r#""😀""#,
            "\"raw\nnewline\"",
            r#""café ❤""#,
            "\"\u{e9}\\\"",
            "\u{feff}{}",
            "[[[[1]]]]",
        ];
        for sample in samples {
            let ours = Json::new(sample).is_ok();
            let theirs = serde_json::from_str::<serde_json::Value>(sample).is_ok();
            assert_eq!(ours, theirs, "disagreed about {sample:?}");
        }
    }

    #[test]
    fn python_numbers_preserve_large_integers_and_overflow() {
        for number in [
            "1234567890123456789012345678901234567890",
            "-1234567890123456789012345678901234567890",
        ] {
            assert_eq!(render_plain(number, 100), number);
        }
        assert_eq!(render_plain("-0", 100), "0");
        assert_eq!(render_plain("1e400", 100), "Infinity");
        assert_eq!(render_plain("-1e999", 100), "-Infinity");
        for invalid in ["01", "-01", "1.e2", "1e+", "1e400x", "--1", "1+2", ".1"] {
            assert!(Json::new(invalid).is_err(), "accepted {invalid}");
        }
    }

    /// And the tree it builds must be the tree serde_json would have built.
    /// `serde_json::to_string_pretty` happens to use the very layout upstream's
    /// `json.dumps(indent=2)` does, so it doubles as a reference dump: key
    /// order, repeated-key collapsing, string escaping and number formatting
    /// all have to agree.
    #[test]
    fn the_parsed_tree_matches_serde_json() {
        let samples = [
            r#"{"name": "Alice", "age": 30, "admin": true, "tags": ["a", "b"], "meta": null}"#,
            r#"{"a": 1, "b": 2, "a": 3}"#,
            r#"{"a": {"b": {"c": [1, [], {}, [[2]]]}}}"#,
            r#"{"k": "A\t\"x\"A\\\/é"}"#,
            r#"[0, 12345678901234567890, -7]"#,
            r#"{"café": "❤", "": ""}"#,
            "[]",
            "{}",
            "\"top level\"",
            "1234",
        ];
        for sample in samples {
            let reference: serde_json::Value =
                serde_json::from_str(sample).expect("sample is valid JSON");
            assert_eq!(
                render_plain(sample, 10_000),
                serde_json::to_string_pretty(&reference).expect("value re-serialises"),
                "diverged on {sample}"
            );
        }
    }

    /// Floats are written with Python's `float.__repr__`, as `json.dumps` does,
    /// not serde_json's spelling (`1e-7`, `10000000000.0`).
    #[test]
    fn floats_render_as_python_repr() {
        let out = render_plain(
            "[-0.5, 1e10, 1E+10, 1e-7, 1.7976931348623157e308, 1e16]",
            10_000,
        );
        let values: Vec<&str> = out
            .lines()
            .filter_map(|line| line.trim().strip_suffix(',').or(Some(line.trim())))
            .filter(|value| !matches!(*value, "[" | "]"))
            .collect();
        assert_eq!(
            values,
            [
                "-0.5",
                "10000000000.0",
                "10000000000.0",
                "1e-07",
                "1.7976931348623157e+308",
                "1e+16"
            ]
        );
    }

    /// A repeated key collapses to one entry — first position, last value —
    /// which is what both `dict` and serde_json's `preserve_order` produce.
    #[test]
    fn a_repeated_key_keeps_its_position_and_last_value() {
        assert_eq!(
            render_plain(r#"{"a": 1, "b": 2, "a": 3}"#, 40),
            "{\n  \"a\": 3,\n  \"b\": 2\n}"
        );
    }

    /// Escapes are decoded and re-encoded, because upstream re-serialises the
    /// parsed data with `json.dumps`.
    #[test]
    fn escapes_are_re_encoded_like_dumps() {
        assert_eq!(
            render_plain(r#"{"k": "A\t\"x\""}"#, 60),
            "{\n  \"k\": \"A\\t\\\"x\\\"\"\n}"
        );
    }
}