rustledger_parser/cst/

format.rs

//! Opinionated CST-backed formatter (phase 4.1 of #1262).
//!
//! [`format_source`] is a pure function `&str → String`: it
//! reparses the input into a CST and emits text in one canonical
//! form per AST shape. Two semantically-equivalent inputs produce
//! byte-identical output; idempotence (`f(f(x)) == f(x)`) follows
//! trivially.
//!
//! Replaces the pre-#1262 source-level formatter that took
//! `(source, ParseResult, FormatConfig)` and re-emitted via the
//! AST-driven `rustledger_core::format` path. Typed-directive
//! synthesis (`rustledger_core::format::format_directives`) still
//! lives in `rustledger-core` for callers that build a directive
//! from scratch (e.g., `rledger add`, importer extract, FFI
//! `format.entry`) — that's a different shape of input and is
//! out of scope here.
//!
//! # Typed-directive emit: known coupling
//!
//! The typed-directive path is a two-pass shim: callers run
//! `core::format::format_directives` to get bean-format-style text,
//! then run that text back through [`format_source`] for the
//! canonical pass. This keeps the FINAL byte sequence single-
//! sourced (always emitted by this module), but it means
//! `core::format` is permanently load-bearing as a parser-clean
//! intermediate and every canonical-form rule needs the legacy
//! emitter to produce SOMETHING the new parser accepts.
//!
//! Call sites (`rustledger-ffi-wasi::router::canonical_format_directives`,
//! `rustledger::cmd::add_cmd::canonical_format_directive`,
//! `rustledger::cmd::extract_cmd`) all guard the round-trip with
//! an explicit `parse(&raw)` step that bails on parse errors, so a
//! divergence between the two emitters surfaces as a hard error
//! instead of silently dropping content.
//!
//! The eventual fix is a typed-directive emit path on this module
//! (`format_directive(&Directive) -> String`) that bypasses the
//! source-string round-trip. Tracked in a follow-up issue.
//!
//! # Canonical form (locked in the PR-decision comment on #1262)
//!
//! - Indent inside a directive body: 2 spaces. Tabs converted.
//! - Blank lines between directives: preserved from the source
//!   (#1325). Grouped directives (consecutive `open`s, a `price`
//!   feed) stay grouped; the formatter does not insert or collapse
//!   blank lines, matching Python `bean-format`.
//! - Blank lines inside a directive: 0.
//! - Number lexical form: thousands separators dropped; user
//!   decimal-place count preserved.
//! - Comment content: verbatim.
//! - Comment positions: normalized to the attachment slot
//!   (header-trailing / inter-directive / body-internal /
//!   posting-trailing).
//! - Cost spec spacing: `{cost CCY}` (no inner padding).
//! - Tag/link order on a transaction header: source order, after
//!   the strings.
//! - Trailing newline at EOF: always exactly one.
//! - Line endings: LF; CRLF inputs normalized.
//! - Leading BOM: dropped.
//!
//! No `FormatConfig` parameter. One canonical form, no knobs.

use crate::cst::ast::{self, AstNode, AstToken, MetaEntry, SourceFile};

/// Pre-computed alignment data for a whole source file.
///
/// Bean-format-style two-axis alignment. The **number field** is a
/// fixed-width slot starting at column `number_col` and `number_width`
/// chars wide, into which each posting's number / arithmetic
/// expression is right-justified. Shorter numbers are left-padded
/// with spaces, so the currency column (right after the field) is
/// uniform across the whole file even when individual numbers have
/// different widths or signs.
///
/// - `number_col`   = INDENT + max(account width with optional `flag `) + 2
/// - `number_width` = max rendered width of any posting's number /
///   arithmetic expression (sign included)
///
/// `PostingAlignment` is `Copy` and `Default` (the all-zero state);
/// the default is the alignment used for files that contain no
/// postings (no transactions, or transactions with no AMOUNT).
/// Marked `#[non_exhaustive]` so that a future column-derivation
/// rule can add fields without breaking downstream consumers.
///
/// **Name choice.** The type is qualified by its semantic purpose
/// (posting layout column widths) so the public path
/// `rustledger_parser::format::PostingAlignment` doesn't compete
/// with future generic "alignment" types (text justification,
/// memory layout, etc.).
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
#[non_exhaustive]
pub struct PostingAlignment {
    /// 0-indexed column at which the right-justified number field
    /// starts.
    pub number_col: usize,
    /// Width of the number field; shorter numbers are left-padded
    /// with spaces so the currency column stays uniform.
    pub number_width: usize,
}

/// How numerals are grouped, resolved per currency.
///
/// `Copy` on purpose: it is threaded as a sibling of [`PostingAlignment`]
/// through the width pre-pass and the emitters, which must agree about a
/// numeral's width or the currency column drifts — the #1290 failure. It is
/// NOT a field of `PostingAlignment`: that type is `pub`, `Eq` and cached on
/// `ParseResult`, so it cannot carry a lifetime. The invariant that pairs a
/// cached alignment with the style it was measured under is stated on
/// `format_node_with_style` (crate-internal).
#[derive(Debug, Clone, Copy, Default)]
pub struct GroupingStyle<'a> {
    ctx: Option<&'a rustledger_core::DisplayContext>,
}

impl<'a> GroupingStyle<'a> {
    /// Grouping as declared by a ledger's display context.
    ///
    /// Returns the no-grouping style when the context groups nothing, so the
    /// overwhelmingly common case costs no per-numeral lookups.
    #[must_use]
    pub fn from_context(ctx: &'a rustledger_core::DisplayContext) -> Self {
        Self {
            ctx: ctx.renders_any_commas().then_some(ctx),
        }
    }

    /// Whether this style groups anything at all.
    ///
    /// Lets a caller keep the fast precomputed-alignment path when nothing is
    /// declared, which is the overwhelming majority of ledgers.
    #[must_use]
    pub const fn groups_anything(self) -> bool {
        self.ctx.is_some()
    }

    /// Whether a numeral in `currency` is grouped. `None` for numerals with no
    /// currency in scope (metadata values, `custom` values), which take the
    /// ledger-wide default.
    #[must_use]
    pub fn groups(self, currency: Option<&str>) -> bool {
        match (self.ctx, currency) {
            (None, _) => false,
            (Some(c), Some(cur)) => c.render_commas_for(cur),
            (Some(c), None) => c.render_commas(),
        }
    }
}

/// Two-space indent for directive bodies (postings, metadata).
const INDENT: &str = "  ";

/// Format a Beancount source file in opinionated canonical form.
///
/// Reparses internally — callers that already have a CST in hand
/// and want to avoid the double-parse can use [`format_node`].
///
/// Returns canonical text; output always ends with exactly one
/// trailing newline (even for an empty file, where the output is
/// just `"\n"`).
///
/// **Line-ending normalization runs BEFORE parsing.** The lexer
/// does not treat bare `\r` as a line terminator, so a classic-
/// Mac-authored `directive\r…\rdirective\r` would otherwise parse
/// as a single broken directive and the rest of the user's ledger
/// would be silently dropped. We normalize `\r\n` and bare `\r`
/// to `\n` first, then parse — matching the canonical-form
/// promise that line endings are LF-only on output.
#[must_use]
pub fn format_source(source: &str) -> String {
    format_source_grouped(source, GroupingStyle::default())
}

/// [`format_source`], with the thousands-separator rule supplied by the caller.
///
/// The style comes from the LEDGER, not from the tool: `option
/// "render_commas"` and per-commodity `render_commas:`, resolved by whoever
/// holds the options (the CLI after a load, the LSP from its ledger state, the
/// FFI session from its own). That is what keeps this a canonical form rather
/// than a knob — the output is still a function of the input, the declaration
/// simply travels with it.
///
/// `format_source` passes [`GroupingStyle::default`], so every existing caller
/// and every ledger that has not opted in is byte-for-byte unaffected.
#[must_use]
pub fn format_source_grouped(source: &str, style: GroupingStyle<'_>) -> String {
    let (stripped, _had_bom) = crate::bom::strip_leading(source);
    let normalized = crlf_to_lf_outside_strings(stripped);
    let parsed = SourceFile::parse(&normalized);
    format_node_grouped(parsed.syntax(), style)
}

/// Like [`format_source`] but reuses the caller's
/// [`crate::ParseResult`] instead of re-parsing `source`.
///
/// Skips both expensive pre-passes the bare `format_source` runs
/// every call: the lex+parse from `SourceFile::parse(&normalized)`,
/// and the `O(N_postings)` `compute_alignment` walk. Both pieces
/// are already on `parse_result` (in `syntax_root` and
/// `alignment` respectively, populated by `parse_via_cst`). For
/// any consumer that already holds a `ParseResult` — the LSP
/// `format_document` handler, the FFI `format.source` endpoint,
/// the WASM `ParsedLedger::format` bridge — this entry skips two
/// redundant traversals of the file.
///
/// **Output equivalence with `format_source`.** Pinned by
/// `parse_result_alignment_cache::format_source_with_parsed_matches_format_source_under_fallback`
/// (the fallback exercises broken sources) and
/// `cst::format::tests::format_source_with_parsed_matches_format_source`
/// (the cache path exercises clean sources) across LF / CRLF /
/// BOM / parse-error / mixed-line-ending fixtures. The cache-
/// path equivalence holds because the formatter rebuilds output
/// from each directive's typed values rather than echoing
/// trivia, so the CRLF-vs-LF difference in the underlying CST
/// trivia never reaches the output. The fallback path is
/// byte-trivially equivalent (it IS `format_source`).
///
/// **CRLF re-injection is still the caller's responsibility.**
/// Same as `format_source`: this function always returns LF;
/// LSP consumers that need to preserve CRLF for Windows-
/// authored files call [`lf_to_crlf_outside_strings`] on the
/// returned text.
///
/// **Parse-error fallback.** When `parse_result.errors` is
/// non-empty, this function delegates to `format_source(source)`
/// — losing the cache benefit but preserving byte-identity for
/// inputs whose CST diverges from what `format_source`'s
/// pre-parse normalization would produce. Concretely: bare-`\r`
/// (classic Mac) line terminators are normalized to LF by
/// `format_source` before parsing, but `parse_via_cst` does NOT
/// normalize them — so the cached CST treats them as broken
/// content and `parse_result.errors` is non-empty. The fallback
/// path keeps the byte-identity claim total instead of
/// "holds-only-when-clean".
///
/// **Stale `parse_result` is the caller's responsibility.** The
/// producer-side cache invariant (see
/// [`crate::ParseResult::alignment`] rustdoc) says
/// `parse_result` must come from a fresh `parse(source)` with
/// the same `source`. A `debug_assert_eq!` compares the CST's
/// text length against `source.len() - bom_offset` to catch the
/// most common mismatched-pair class (different documents have
/// different lengths) in debug builds; release builds skip the
/// check. Identical-length mismatches still pass silently —
/// the rustdoc-level contract remains the source of truth.
///
/// Formats under the DEFAULT grouping style, always: it reuses
/// `ParseResult::alignment`, which is measured ungrouped, and the two
/// must agree. A caller that needs a ledger's separators wants
/// [`format_node_grouped`], which measures its own alignment.
///
/// # Panics
///
/// Panics if `parse_result.syntax_root` is not a `SOURCE_FILE`
/// (always true for results produced by [`crate::parse`]).
///
/// In debug builds, panics on a `(parse_result, source)`
/// length-mismatch via `debug_assert_eq!`. Release builds
/// silently emit possibly-wrong output (pairing `source` with the
/// `parse_result` it came from is the caller's responsibility).
#[must_use]
pub fn format_source_with_parsed(parse_result: &crate::ParseResult, source: &str) -> String {
    // Parse-error fallback. See the function rustdoc for the
    // rationale: `parse_via_cst` does not run the same input
    // normalization `format_source` does (no CRLF/bare-CR
    // normalize), so for sources containing bare-`\r` line
    // terminators the cached CST is wrong-shaped and the cache
    // path would diverge from `format_source`. Delegating
    // preserves byte-identity unconditionally.
    if !parse_result.errors.is_empty() {
        return format_source(source);
    }
    let node = parse_result.syntax_node();
    // Defensive length check (debug-only). Catches the most
    // common form of `(parse_result, source)` mismatched pair —
    // different documents with different lengths. The CST's
    // text range is BOM-stripped, so we add back the BOM bytes
    // if the parser saw one.
    //
    // Computed outside the `debug_assert_eq!` to avoid clippy's
    // `debug_assert_with_mut_call` (`syntax_node()` does an Arc
    // bump, which clippy treats as state mutation in a debug
    // context).
    let cst_len =
        usize::from(node.text_range().len()) + if parse_result.has_leading_bom { 3 } else { 0 };
    debug_assert_eq!(
        cst_len,
        source.len(),
        "format_source_with_parsed called with a `source` whose length doesn't \
         match the CST stored in `parse_result`. The two arguments came from \
         different documents — the cache path will emit text for the wrong \
         buffer.",
    );
    format_node_with_alignment(&node, parse_result.alignment())
}

/// Like [`format_source`], but returns the parse errors instead
/// of silently formatting around them.
///
/// `format_source` is intentionally infallible — the canonical
/// formatter must still emit *something* for a file the parser
/// could only recover from. Tooling that wants to refuse to
/// rewrite a file with parse errors (the `rledger format` CLI,
/// the LSP `format` handler) previously had to call `parse`
/// out-of-band, inspect `errors`, then call `format_source` on
/// the SAME input — a contract two functions cooperated on
/// implicitly, and the kind of pairing a future caller could
/// easily forget. This helper makes the contract explicit.
///
/// Returns `Ok(formatted)` if and only if `parse(source).errors`
/// would be empty. Otherwise returns the parse errors verbatim,
/// in the same order the parser emitted them.
///
/// # Errors
///
/// Returns `Err(Vec<ParseError>)` containing every parse error
/// the underlying [`parse`](crate::parse) call would surface for
/// `source`. The caller decides whether to abort, render the
/// errors, or fall back to a non-canonical pass.
pub fn try_format_source(source: &str) -> Result<String, Vec<crate::ParseError>> {
    try_format_source_grouped(source, GroupingStyle::default())
}

/// [`try_format_source`], with the grouping style supplied by the caller.
///
/// # Errors
///
/// As [`try_format_source`].
pub fn try_format_source_grouped(
    source: &str,
    style: GroupingStyle<'_>,
) -> Result<String, Vec<crate::ParseError>> {
    let result = crate::parse(source);
    if !result.errors.is_empty() {
        return Err(result.errors);
    }
    // Reuse the parse we already produced for the error gate rather than
    // letting `format_source_grouped` re-parse. The alignment cached on
    // `ParseResult` is computed WITHOUT grouping, so it can only be reused for
    // the default style — see `format_node_with_style`'s invariant.
    if style.groups_anything() {
        return Ok(format_node_grouped(&result.syntax_node(), style));
    }
    Ok(format_source_with_parsed(&result, source))
}

/// Convert every `\n` line terminator OUTSIDE string literals back
/// to `\r\n`, leaving `\n` characters inside strings (and inside
/// comments… see below) untouched.
///
/// The canonical form emitted by [`format_source`] is LF-only.
/// Editors that round-trip Windows-authored files want to see CRLF
/// echoed back on every line. This helper bridges the two by
/// walking the canonical output with the shared `SourceState`
/// state machine. The walker respects:
///
/// - String literals: bytes pass through verbatim. The user's
///   original line endings inside a multi-line narration / note /
///   document string are preserved.
/// - Line comments (`;`, `%`, `#!`, `#+`): the comment's
///   terminating newline IS a real structural line terminator, so
///   it gets converted to CRLF; bytes inside the comment region
///   (which can include arbitrary characters, notably stray `"`)
///   pass through without flipping the in-string state. `#!` and
///   `#+` open a comment at any column — the lexer's
///   `SHEBANG` / `EMACS_DIRECTIVE` regexes carry no line-start
///   anchor, and the state machine matches that classification.
///
/// The helper lives in this module rather than the LSP crate
/// because its correctness depends on the lexer's `STRING` and
/// comment rules. Keep it co-located with the formatter so a
/// lexer change forces a co-evaluation here.
#[must_use]
pub fn lf_to_crlf_outside_strings(s: &str) -> String {
    let mut out = String::with_capacity(s.len() + s.matches('\n').count());
    // BOM is data, not classification input. We re-prepend it
    // verbatim and let the body start fresh in Code state. The
    // sibling crlf_to_lf_outside_strings does the same so the two
    // walkers handle a leading-BOM file identically.
    let (body, bom) = match s.strip_prefix('\u{FEFF}') {
        Some(rest) => (rest, "\u{FEFF}"),
        None => (s, ""),
    };
    out.push_str(bom);
    let mut chars = body.chars().peekable();
    let mut state = SourceState::Code;
    let mut prev_was_backslash = false;
    while let Some(ch) = chars.next() {
        let peek = chars.peek().copied();
        match state {
            SourceState::InString => out.push(ch),
            SourceState::InComment | SourceState::Code => {
                if ch == '\n' {
                    out.push_str("\r\n");
                } else {
                    out.push(ch);
                }
            }
        }
        state = advance_source_state(ch, peek, state, &mut prev_was_backslash);
    }
    out
}

/// Render typed Beancount `Directive`s in the canonical form
/// emitted by [`format_source`].
///
/// Two-pass pipeline:
///
/// 1. Synthesize a source string via the typed-directive emitter
///    in `rustledger_core::format::format_directives`. That
///    emitter is `Directive → text`; its output is bean-format-
///    style, parser-clean, and used here purely as an
///    intermediate.
/// 2. Re-parse the synthesized text. If the legacy emitter
///    produced something the new parser cannot fully accept,
///    return [`CanonicalizeError::ReparseFailed`] rather than
///    silently emitting the recoverable subset — that silent-loss
///    failure mode is what the older `crates/rustledger/tests/
///    format_compat.rs` (deleted in phase 4.1, distinct from the
///    phase 4.2 file-pair suite at `crates/rustledger-parser/
///    tests/format_compat/`) used to guard against. The new file-
///    pair suite exercises `format_source`, not this two-pass
///    shim; a future change to `canonicalize_directives`'s error
///    semantics needs its own dedicated regression test.
/// 3. Run the re-parsed text through [`format_source`] for the
///    canonical pass.
///
/// Single source of truth for the synthesize → canonicalize
/// shim. Every consumer that builds a typed `Directive` in memory
/// and wants canonical text — `rledger add`, `rledger extract`,
/// the FFI `format.entry` / `format.entries` endpoints — should
/// call this function instead of reinventing the pipeline.
pub fn canonicalize_directives<'a, I>(
    directives: I,
    config: &rustledger_core::format::FormatConfig,
) -> Result<String, CanonicalizeError>
where
    I: IntoIterator<Item = &'a rustledger_core::Directive>,
    I::IntoIter: ExactSizeIterator,
{
    // Take the count off the ExactSizeIterator without
    // collecting — the legacy emitter only walks the iterator
    // once, so we don't need to materialize a Vec just to know
    // how many directives the caller passed.
    let iter = directives.into_iter();
    let input_count = iter.len();
    let raw = rustledger_core::format::format_directives(iter, config);
    let parse_result = crate::parse(&raw);
    if !parse_result.errors.is_empty() {
        return Err(CanonicalizeError::ReparseFailed {
            errors: parse_result
                .errors
                .iter()
                .map(ToString::to_string)
                .collect(),
        });
    }
    // Count check covers the only Directive variants we have
    // today (12, all of which surface on parse_result.directives).
    // If a future `rustledger_core::Directive` variant is added
    // that the parser routes to a different `ParseResult`
    // collection (e.g., a typed Pushtag whose legacy text the
    // parser puts on a `pragmas` field), this check needs to
    // include that field too — otherwise a perfectly healthy
    // round-trip would always report DirectiveCountMismatch. The
    // compile-time `_directive_variant_fixture_coverage` match
    // pins the variant set we're committed to here; any new
    // variant breaks that match and surfaces this same
    // maintenance need.
    let reparsed_count = parse_result.directives.len();
    if reparsed_count != input_count {
        return Err(CanonicalizeError::DirectiveCountMismatch {
            input: input_count,
            reparsed: reparsed_count,
        });
    }
    // The second pass re-canonicalizes LAYOUT over the core emitter's text.
    // It must carry the same grouping rule, or it would strip the separators
    // the first pass just wrote — which is exactly why this shim used to
    // preserve precision but not grouping.
    let style = config
        .number_display
        .as_ref()
        .map_or_else(GroupingStyle::default, GroupingStyle::from_context);
    Ok(format_source_grouped(&raw, style))
}

/// Error returned by [`canonicalize_directives`].
///
/// Marked `#[non_exhaustive]` so that adding a future variant
/// (e.g. a `CanonicalizationTimeout` for an async path, or a new
/// guard for a future canonical-form rule) does not become a
/// SemVer-breaking change. Consumers must use a `_ => …` arm.
#[derive(Debug, Clone)]
#[non_exhaustive]
pub enum CanonicalizeError {
    /// The synthesized intermediate failed to re-parse cleanly.
    /// Carries the rendered error messages so callers can surface
    /// a diagnostic; the source text itself is not retained
    /// because it's an internal intermediate the caller has no
    /// control over.
    ReparseFailed {
        /// One rendered message per parse error from the
        /// intermediate text. Capped at the parser's own error
        /// limit so this field is bounded.
        errors: Vec<String>,
    },
    /// The synthesized intermediate parsed cleanly but produced a
    /// different directive count than the input. This indicates
    /// the legacy emitter and the new parser disagree on what
    /// constitutes a directive — typically a future
    /// `rustledger_core::Directive` variant whose legacy text the
    /// CST parser silently swallows as comments / error-recovery
    /// trivia. Without this guard, the call would round-trip to
    /// truncated text with no error returned.
    DirectiveCountMismatch {
        /// Number of directives the caller passed in.
        input: usize,
        /// Number of directives the parser recovered from the
        /// synthesized text.
        reparsed: usize,
    },
}

impl std::fmt::Display for CanonicalizeError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Self::ReparseFailed { errors } => {
                let preview: Vec<&str> = errors.iter().take(3).map(String::as_str).collect();
                write!(
                    f,
                    "canonical formatter failed to re-parse the synthesized \
                     directive text ({} error(s)): {}",
                    errors.len(),
                    preview.join("; ")
                )
            }
            Self::DirectiveCountMismatch { input, reparsed } => write!(
                f,
                "the canonical formatter could not emit {input} directive(s) \
                 without loss ({reparsed} survived the round-trip). This is \
                 an rledger bug; please report it with the input directives.",
            ),
        }
    }
}

impl std::error::Error for CanonicalizeError {}

/// Replace CRLF and bare-CR line terminators with LF, but ONLY
/// outside string literals.
///
/// String literals (`"…"`) can contain raw `\r` and `\n` per the
/// lexer's `STRING` rule; folding CR inside a string would mutate
/// the user's data. Uses the shared `SourceState` state machine
/// to track string / comment boundaries.
///
/// Cheap fast path: if the input contains no `\r`, returns the
/// source slice borrowed (no allocation). Used by
/// [`format_source`] before parsing so the lexer never has to see
/// legacy line endings. Exposed publicly under [`crlf_to_lf_outside_strings`]
/// for tooling (CLI `--diff`, format-equivalence checks) that
/// needs the same string-aware normalization.
pub fn crlf_to_lf_outside_strings(src: &str) -> std::borrow::Cow<'_, str> {
    if !src.contains('\r') {
        return std::borrow::Cow::Borrowed(src);
    }
    // Re-prepend the BOM verbatim and let the body start fresh in
    // Code state. The state machine no longer needs line-start
    // tracking — the lexer's `SHEBANG` / `EMACS_DIRECTIVE` regexes
    // have no line-start anchor, so `#!`/`#+` open a comment at
    // any column, and the state machine mirrors that.
    let (body, bom) = match src.strip_prefix('\u{FEFF}') {
        Some(rest) => (rest, "\u{FEFF}"),
        None => (src, ""),
    };
    let mut out = String::with_capacity(src.len());
    out.push_str(bom);
    let mut chars = body.chars().peekable();
    let mut state = SourceState::Code;
    let mut prev_was_backslash = false;
    while let Some(ch) = chars.next() {
        let peek = chars.peek().copied();
        match state {
            SourceState::InString => out.push(ch),
            _ => {
                if ch == '\r' {
                    out.push('\n');
                    if peek == Some('\n') {
                        chars.next();
                    }
                } else {
                    out.push(ch);
                }
            }
        }
        state = advance_source_state(ch, peek, state, &mut prev_was_backslash);
    }
    std::borrow::Cow::Owned(out)
}

/// `true` iff `src` contains at least one `\r` byte OUTSIDE a
/// string literal — i.e. the byte sequence the canonical
/// formatter would fold to `\n` via
/// [`crlf_to_lf_outside_strings`].
///
/// This is the explicit predicate companion to the Cow return of
/// [`crlf_to_lf_outside_strings`]. Tooling that only needs to
/// know whether the fold would change bytes (the CLI `--diff`
/// "CR-bearing line endings folded" cause line, the LSP
/// did-the-formatter-touch-this guard) should call this instead
/// of matching on `Cow::Owned`, which conflates allocation with
/// semantic change. A future optimization that pre-allocated the
/// Cow even on a no-op fold would silently invert that
/// match-on-Cow guard; this predicate keeps the question
/// answered by the bytes, not by allocation behavior.
#[must_use]
pub fn cr_outside_strings_present(src: &str) -> bool {
    if !src.contains('\r') {
        return false;
    }
    let body = src.strip_prefix('\u{FEFF}').unwrap_or(src);
    let mut chars = body.chars().peekable();
    let mut state = SourceState::Code;
    let mut prev_was_backslash = false;
    while let Some(ch) = chars.next() {
        let peek = chars.peek().copied();
        if matches!(state, SourceState::Code | SourceState::InComment) && ch == '\r' {
            return true;
        }
        state = advance_source_state(ch, peek, state, &mut prev_was_backslash);
    }
    false
}

/// Per-character walker state for line-ending normalization passes
/// that must respect string-literal and comment boundaries.
///
/// Used by both line-ending helpers: a flat `is_in_string` boolean
/// is not enough because a quote character inside a `;`/`%` /
/// `#!` / `#+` comment is data, not a string delimiter.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum SourceState {
    /// In normal code. `"` opens a string; `;` / `%` / `#!` /
    /// `#+` opens a comment; everything else is just bytes.
    Code,
    /// Inside `"…"`. Bytes pass through; an unescaped `"` exits.
    InString,
    /// Inside `;…\n`, `%…\n`, `#!…\n`, or `#+…\n`. Bytes pass
    /// through until LF/CR.
    InComment,
}

/// One-step state transition shared by both line-ending helpers.
///
/// Returns the state AFTER consuming `ch`. The string-escape
/// bookkeeping (`prev_was_backslash`) updates in place. Comment
/// opener detection covers all four line-comment lexemes: `;` and
/// `%` open a comment unconditionally; `#!` and `#+` open one at
/// any column — the lexer's `#![^\n\r]*` / `#\+[^\n\r]*` regexes
/// have NO line-start anchor, so a mid-line `#!` or `#+` is still
/// a `SHEBANG` / `EMACS_DIRECTIVE` token. A `#` followed by
/// anything else is a `TAG` / `HASH` token, not a comment.
const fn advance_source_state(
    ch: char,
    peek: Option<char>,
    state: SourceState,
    prev_was_backslash: &mut bool,
) -> SourceState {
    match state {
        SourceState::InString => {
            let is_close = ch == '"' && !*prev_was_backslash;
            *prev_was_backslash = ch == '\\' && !*prev_was_backslash;
            if is_close {
                SourceState::Code
            } else {
                SourceState::InString
            }
        }
        SourceState::InComment => {
            if matches!(ch, '\n' | '\r') {
                SourceState::Code
            } else {
                SourceState::InComment
            }
        }
        SourceState::Code => {
            let is_hash_line_comment = ch == '#' && matches!(peek, Some('!' | '+'));
            if ch == '"' {
                *prev_was_backslash = false;
                SourceState::InString
            } else if matches!(ch, ';' | '%') || is_hash_line_comment {
                SourceState::InComment
            } else {
                SourceState::Code
            }
        }
    }
}

/// Format a `SOURCE_FILE` syntax node in opinionated canonical form.
///
/// The bare-node entry for callers that already parsed the CST
/// (typically LSP formatting providers). Output rules are the
/// same as [`format_source`].
///
/// Internally runs [`compute_alignment`] on `node` to derive the
/// file-wide column targets. Hot paths that hold a precomputed
/// `PostingAlignment` (e.g., via [`crate::ParseResult::alignment`]) should
/// call [`format_node_with_alignment`] instead to skip the
/// per-call walk. Equivalence pinned by
/// `format_node_equals_format_node_with_alignment` in this file's
/// tests.
#[must_use]
pub fn format_node(node: &crate::SyntaxNode) -> String {
    // The SOURCE_FILE precondition is asserted by `format_node_grouped`.
    format_node_grouped(node, GroupingStyle::default())
}

/// [`format_node`], with the thousands-separator rule supplied by the caller.
#[must_use]
pub fn format_node_grouped(node: &crate::SyntaxNode, style: GroupingStyle<'_>) -> String {
    // Precondition as in `format_node`.
    #[allow(clippy::expect_used)]
    let source_file =
        SourceFile::cast(node.clone()).expect("format_node_grouped called on non-SOURCE_FILE node");
    let alignment = compute_alignment(&source_file, style);
    format_node_with_style(node, alignment, style)
}

/// Like [`format_node`] but skips the per-call
/// [`compute_alignment`] walk by accepting a precomputed
/// `PostingAlignment`.
///
/// The cache pattern: parse → take `ParseResult::alignment` (the
/// pre-computed file-wide alignment, populated by `parse_via_cst`)
/// → call this function. Subsequent formatting calls on the same
/// `ParseResult` pay only the per-call emit cost, not the
/// `O(N_postings)` pre-pass.
///
/// `alignment` MUST match what `compute_alignment(&SourceFile::cast(node).unwrap())` would
/// return for the given `node` — passing a mismatched alignment
/// is allowed but produces output with non-canonical column
/// widths. Use `PostingAlignment::default()` for files known to have no
/// postings (no transactions, or transactions with no AMOUNT).
///
/// # Panics
///
/// Panics if `node`'s kind is not `SOURCE_FILE`.
#[must_use]
pub fn format_node_with_alignment(node: &crate::SyntaxNode, alignment: PostingAlignment) -> String {
    format_node_with_style(node, alignment, GroupingStyle::default())
}

/// [`format_node_with_alignment`] with an explicit grouping style.
///
/// Deliberately NOT public. It is the one shape that can be called wrongly —
/// see the invariant below — and every public entry point either measures the
/// alignment itself (`format_node_grouped`) or is fixed to the default style,
/// which is what `ParseResult::alignment` is measured under. Keeping this
/// private makes the mismatch unstatable from outside the crate rather than
/// merely documented.
///
/// INVARIANT: `alignment` must have been produced by [`compute_alignment`] with
/// this same `style`. Grouping changes a numeral's width, so an alignment
/// measured under a different style puts the currency column in the wrong place
/// — the #1290 non-convergence, reintroduced. The `_grouped` entry points
/// derive both from one value and cannot mismatch; a caller reusing a cached
/// `ParseResult::alignment` (computed at parse time, before any options are
/// known) must pass `GroupingStyle::default()`.
#[must_use]
fn format_node_with_style(
    node: &crate::SyntaxNode,
    alignment: PostingAlignment,
    group: GroupingStyle<'_>,
) -> String {
    // Precondition check (debug-only). The bare `format_node`
    // delegate already validated the kind via the
    // `SourceFile::cast` it performs for `compute_alignment`, so
    // for the most common call path (bare → with_alignment) the
    // debug_assert is a redundant no-op in release. External
    // direct callers of this entry point (FFI, future LSP
    // handlers calling `format_node_with_alignment` with a
    // `parse_result.alignment()` cache) get the panic in debug
    // builds; in release, a wrong-kind `node` produces empty or
    // malformed output rather than panicking — acceptable for
    // a precondition that's guaranteed by the call's typed
    // contract.
    debug_assert_eq!(
        node.kind(),
        crate::SyntaxKind::SOURCE_FILE,
        "format_node_with_alignment called on non-SOURCE_FILE node (got {:?})",
        node.kind(),
    );
    let mut out = String::new();
    // Walk every direct child in source order so file-level comments
    // (file-leading per phase-2.0 trivia attachment, plus file-
    // trailing) interleave correctly with directives. Inter-directive
    // and same-line trailing comments live INSIDE the next/owning
    // directive and surface from `emit_directive`'s leading-trivia
    // pass.
    //
    // Blank-line policy at the top level: PRESERVE the author's blank
    // lines between directives rather than normalizing to exactly one.
    // Between two directives, emit as many blank lines as the source
    // had — including zero, so deliberately grouped runs (consecutive
    // `open`s, a dense `price` feed) stay grouped instead of being
    // double-spaced (#1325). This matches Python `bean-format` and the
    // rest of the beancount formatter lineage (fava,
    // beancount-language-server, beancount-mode), all of which leave
    // blank-line structure untouched and only realign amounts.
    //
    // Adjacent file-level comments still stay tight as a group (so a
    // `; ====\n; HEADER\n; ====` section header keeps its visual
    // grouping), and a comment group sitting against a directive on
    // either side stays flush.
    let mut prev_was_directive = false;
    for el in node.children_with_tokens() {
        match el {
            rowan::NodeOrToken::Node(n) => {
                if let Some(directive) = ast::Directive::cast(n.clone()) {
                    if prev_was_directive {
                        for _ in 0..leading_blank_lines(directive.syntax()) {
                            out.push('\n');
                        }
                    }
                    emit_directive(&directive, alignment, group, &mut out);
                    prev_was_directive = true;
                } else if n.kind() == crate::SyntaxKind::ERROR_NODE {
                    // Preserve unparsable content verbatim (#1335): `format`
                    // must never delete the author's text. Org-mode `*`
                    // section headers (and any comments grouped with them)
                    // parse into ERROR_NODEs; emit them as-is rather than
                    // dropping them. Treated like a directive for spacing — an
                    // ERROR_NODE is a top-level content block, so the author's
                    // blank lines around it (before it, and before the next
                    // directive) are preserved, not flushed.
                    if prev_was_directive {
                        for _ in 0..leading_blank_lines(&n) {
                            out.push('\n');
                        }
                    }
                    emit_error_node(&n, &mut out);
                    prev_was_directive = true;
                }
                // Any other non-directive node: nothing to emit.
            }
            rowan::NodeOrToken::Token(t) => {
                if matches!(
                    t.kind(),
                    crate::SyntaxKind::COMMENT
                        | crate::SyntaxKind::PERCENT_COMMENT
                        | crate::SyntaxKind::SHEBANG
                        | crate::SyntaxKind::EMACS_DIRECTIVE
                ) {
                    out.push_str(t.text().trim_end_matches(['\n', '\r']));
                    out.push('\n');
                    prev_was_directive = false;
                }
            }
        }
    }
    if !out.ends_with('\n') {
        out.push('\n');
    }
    out
}

/// Format the subset of `node`'s top-level children that intersect
/// `range`, returning the snapped byte range and the canonical-form
/// replacement text.
///
/// This is the building block for the LSP `textDocument/rangeFormatting`
/// provider: the client sends a `Range`, the server snaps it up to
/// the smallest set of top-level structural nodes (directives or
/// standalone comments) that intersect the selection, formats those
/// nodes the same way [`format_node`] formats the whole file, and
/// returns a single `TextEdit` replacing the snapped range. The
/// alternative — formatting a substring of the source — would have
/// to either invent a partial canonical form (creating a second
/// truth alongside the whole-file canonical form, the failure mode
/// that bit #1252) or refuse to format anything that crosses a
/// structural boundary. Snapping up to top-level boundaries is the
/// only choice that lets the same canonical-form rules apply.
///
/// **Frame.** `range` is in the *CST* byte frame — the same frame
/// the syntax node's `TextRange`s use. The LSP handler is
/// responsible for shifting `bom_offset` at the input/output
/// boundary (mirrors the [`super::super::SyntaxNode`] /
/// `selection_range` handler convention; see
/// `ParseResult::syntax_root` rustdoc for the rationale).
///
/// **Behavior.**
///
/// - If `range` intersects no top-level Directive or standalone
///   COMMENT/SHEBANG/EMACS token, returns `None`. The LSP handler
///   surfaces `None` directly (serialized as `null` per LSP, not
///   as `[]`); the client treats it as "nothing to format".
/// - If the computed snap range would cover any top-level
///   `ERROR_NODE` byte, returns `None`. **Range formatting refuses
///   to delete user content the parser couldn't classify.** This
///   diverges from [`format_node`], which silently drops
///   `ERROR_NODE` children on the whole-file path; the rationale
///   is the per-handler asymmetry the LSP exposes — the user
///   pressing "Format Selection" expects either a clean
///   reformat or a no-op, never a silent partial delete of an
///   in-progress directive. Tooling that genuinely wants to drop
///   broken regions can still call [`format_node`] on the same
///   node.
/// - Otherwise returns `Some((snap, text))` where `snap` is the
///   union of the included children's text ranges (so it begins at
///   the first included child's start and ends at the last
///   included child's end, including each child's leading-trivia
///   prefix per the phase-2.0 Directive-Terminator Rule) and
///   `text` is the canonical-form replacement.
/// - Cursor-only selection (`range.is_empty()`): the child at the
///   cursor is included if the cursor is strictly inside it OR is
///   exactly at the child's start. Boundary at the child's end
///   belongs to the next child, not the previous one — matches
///   the standard "end-of-line cursor is start-of-next-line"
///   convention.
///
/// **Posting alignment.** The pre-pass uses the FULL `SourceFile`, not
/// the selected subset. A selection that formats one transaction
/// in a file with many other transactions inherits the file's
/// alignment columns, so the formatted output stays visually
/// aligned with un-formatted postings elsewhere. The opposite
/// policy (per-selection alignment) would create a jarring
/// visual jump every time the user re-formats a sub-range.
///
/// **Round-trip invariant.** For any `range` that contains every
/// top-level child, the returned text equals the result of
/// [`format_node`] on the same node. Pinned by
/// `format_node_range_full_range_matches_format_node` in this
/// file's test module.
///
/// Returns `None` if `node`'s kind is not `SOURCE_FILE` (the precondition is
/// still that callers pass the parse root) or if `range` intersects no top-level
/// child.
#[must_use]
pub fn format_node_range(
    node: &crate::SyntaxNode,
    range: rowan::TextRange,
) -> Option<(rowan::TextRange, String)> {
    // This returns `Option`, so a non-SOURCE_FILE node is a clean `None` rather
    // than a panic (the precondition is still that callers pass the parse root).
    let source_file = SourceFile::cast(node.clone())?;
    // File-wide alignment pre-pass: see rustdoc above for the
    // rationale. The selected subset always uses the full file's
    // alignment columns. Hot paths with a precomputed `PostingAlignment`
    // should call `format_node_range_with_alignment` instead.
    let alignment = compute_alignment(&source_file, GroupingStyle::default());
    format_node_range_with_alignment(node, range, alignment)
}

/// Like [`format_node_range`] but skips the per-call
/// [`compute_alignment`] walk by accepting a precomputed
/// `PostingAlignment`.
///
/// The cache pattern is identical to
/// [`format_node_with_alignment`]: parse → take
/// `ParseResult::alignment` → call this function. The hot path the
/// cache addresses is the LSP `textDocument/rangeFormatting`
/// fallback (CST-snap path that fires on parse-error files), which
/// can be invoked per-keystroke through format-on-type clients.
/// Without the cache the per-call cost is
/// `O(N_postings_in_file)`; with the cache it's
/// `O(N_cst_nodes covered by range)`.
///
/// `alignment` MUST match what `compute_alignment(&SourceFile::cast(node).unwrap())` would
/// return for the given `node`; pinned by
/// `format_node_range_matches_format_node_range_with_alignment`. Same
/// `range` semantics, `ERROR_NODE` policy, snap rules, and
/// `# Panics` precondition as [`format_node_range`].
#[must_use]
pub fn format_node_range_with_alignment(
    node: &crate::SyntaxNode,
    range: rowan::TextRange,
    alignment: PostingAlignment,
) -> Option<(rowan::TextRange, String)> {
    format_node_range_with_style(node, range, alignment, GroupingStyle::default())
}

/// [`format_node_range`] under an explicit grouping style, measuring the
/// alignment itself.
///
/// Prefer this over pairing a caller-supplied alignment with a
/// alignment: the alignment MUST have been measured under the same style, and
/// doing both here makes that unfalsifiable. Costs one `compute_alignment`
/// walk, which is why the ungrouped hot path still uses the cached alignment.
pub fn format_node_range_grouped(
    node: &crate::SyntaxNode,
    range: rowan::TextRange,
    style: GroupingStyle<'_>,
) -> Option<(rowan::TextRange, String)> {
    let source_file = SourceFile::cast(node.clone())?;
    let alignment = compute_alignment(&source_file, style);
    format_node_range_with_style(node, range, alignment, style)
}

/// [`format_node_range_with_alignment`] with an explicit grouping style.
///
/// Same `alignment`/`style` agreement invariant as
/// `format_node_with_style`: `alignment` must have been measured by
/// `compute_alignment` under this same `style`, or the currency column will be
/// off by the separators' width. The cached `ParseResult::alignment` is
/// measured UNGROUPED, so it may only be paired with the default style.
fn format_node_range_with_style(
    node: &crate::SyntaxNode,
    range: rowan::TextRange,
    alignment: PostingAlignment,
    style: GroupingStyle<'_>,
) -> Option<(rowan::TextRange, String)> {
    // Precondition check (debug-only). Same rationale as
    // `format_node_with_alignment`: the bare delegate already
    // validated the kind, so the most common call path (bare →
    // with_alignment) gets no release-build cost from this
    // assert. External direct callers — the LSP range_formatting
    // fallback, FFI, future format-on-type — get a debug-build
    // panic; release-build wrong-kind input produces no output
    // (rather than panicking).
    debug_assert_eq!(
        node.kind(),
        crate::SyntaxKind::SOURCE_FILE,
        "format_node_range_with_alignment called on non-SOURCE_FILE node (got {:?})",
        node.kind(),
    );

    // First pass: identify the included children and the snap range.
    // We pick:
    //   - Directive nodes whose `text_range` intersects `range`
    //   - top-level COMMENT/PERCENT_COMMENT/SHEBANG/EMACS_DIRECTIVE
    //     tokens whose range intersects `range`
    // ERROR_NODE and other non-Directive nodes are skipped (matches
    // `format_node`); a selection that lands only on them returns
    // None below.
    let mut snap_start: Option<rowan::TextSize> = None;
    let mut snap_end: Option<rowan::TextSize> = None;
    let mut any_included = false;
    for el in node.children_with_tokens() {
        let (kind, child_range) = (el.kind(), el.text_range());
        let is_formattable = match &el {
            rowan::NodeOrToken::Node(n) => ast::Directive::cast(n.clone()).is_some(),
            rowan::NodeOrToken::Token(_) => matches!(
                kind,
                crate::SyntaxKind::COMMENT
                    | crate::SyntaxKind::PERCENT_COMMENT
                    | crate::SyntaxKind::SHEBANG
                    | crate::SyntaxKind::EMACS_DIRECTIVE
            ),
        };
        if !is_formattable {
            continue;
        }
        if !range_intersects(child_range, range) {
            continue;
        }
        any_included = true;
        snap_start = Some(snap_start.map_or(child_range.start(), |s| s.min(child_range.start())));
        snap_end = Some(snap_end.map_or(child_range.end(), |e| e.max(child_range.end())));
    }
    if !any_included {
        return None;
    }
    // `any_included` guarantees both bounds were set in the loop above; bail
    // (return `None`) rather than `unwrap` if somehow not.
    let (Some(snap_start), Some(snap_end)) = (snap_start, snap_end) else {
        return None;
    };
    let snap = rowan::TextRange::new(snap_start, snap_end);

    // ERROR_NODE intersection bail: if the snap range covers any
    // top-level ERROR_NODE byte, refuse to format and return None.
    // Range formatting must not silently delete content the parser
    // could not classify — without this guard, a selection
    // spanning two valid directives with an ERROR_NODE between
    // them would emit a TextEdit that replaces all three with
    // just the two formatted directives, deleting the user's
    // in-progress source bytes.
    //
    // This is the deliberate divergence from `format_node`'s
    // whole-file policy: the whole-file path runs on the
    // assumption that the caller (CLI / FFI / `try_format_source`)
    // has already decided to accept content loss; the per-handler
    // LSP path has no such opt-in. The cost is occasional
    // "format-selection did nothing" UX while a parse error sits
    // inside the snap; the benefit is no data loss.
    for el in node.children_with_tokens() {
        if !matches!(el.kind(), crate::SyntaxKind::ERROR_NODE) {
            continue;
        }
        let er = el.text_range();
        // Strict-overlap check: an ERROR_NODE whose end touches
        // snap.start (or start touches snap.end) is adjacent, not
        // overlapping — those are safe to emit alongside.
        if er.end() > snap.start() && er.start() < snap.end() {
            return None;
        }
    }

    // Second pass: emit only the children whose range falls
    // inside `snap`. We re-walk rather than caching the first
    // pass because the second pass needs to maintain the
    // `prev_was_directive` blank-line state in source order, and
    // the child set is small enough that the second walk is
    // cheap. (Re-walking also keeps the data-flow obvious: snap
    // computation and emission are two distinct concerns.)
    let mut out = String::new();
    let mut prev_was_directive = false;
    for el in node.children_with_tokens() {
        let child_range = el.text_range();
        // Use the snap range (not the input `range`) so we emit
        // every child WITHIN the snap, even those that the
        // original selection didn't directly intersect but that
        // sit between two intersecting children. Without this,
        // ERROR_NODE-free trivia between two selected directives
        // would be re-formatted into our output (the comment
        // pass picks them up), which matches `format_node`.
        if child_range.end() <= snap.start() || child_range.start() >= snap.end() {
            continue;
        }
        match el {
            rowan::NodeOrToken::Node(n) => {
                // ERROR_NODEs never reach here: the range path bails out
                // above (returns None) when the snap covers one, so it
                // refuses to format rather than risk touching unparsable
                // content. Only the whole-file path preserves them verbatim.
                let Some(directive) = ast::Directive::cast(n) else {
                    continue;
                };
                // Preserve the author's inter-directive blank lines
                // (#1325), identically to `format_node_with_alignment`,
                // so range formatting and whole-file formatting agree.
                //
                // The FIRST directive emitted from the snap needs care:
                // its predecessor may sit OUTSIDE the selection, but the
                // blank lines between them are this directive's leading
                // trivia (the Directive-Terminator Rule), so they fall
                // INSIDE the snapped range. Dropping them would delete
                // the blank line above the selection. Emit them whenever
                // a directive precedes this one in the file — the same
                // condition the whole-file path expresses as
                // `prev_was_directive`. For the file's first directive
                // (no predecessor) there is nothing to preserve.
                let preceded_by_directive = prev_was_directive
                    || directive
                        .syntax()
                        .prev_sibling()
                        .and_then(ast::Directive::cast)
                        .is_some();
                if preceded_by_directive {
                    for _ in 0..leading_blank_lines(directive.syntax()) {
                        out.push('\n');
                    }
                }
                emit_directive(&directive, alignment, style, &mut out);
                prev_was_directive = true;
            }
            rowan::NodeOrToken::Token(t) => {
                if matches!(
                    t.kind(),
                    crate::SyntaxKind::COMMENT
                        | crate::SyntaxKind::PERCENT_COMMENT
                        | crate::SyntaxKind::SHEBANG
                        | crate::SyntaxKind::EMACS_DIRECTIVE
                ) {
                    out.push_str(t.text().trim_end_matches(['\n', '\r']));
                    out.push('\n');
                    prev_was_directive = false;
                }
            }
        }
    }
    if !out.ends_with('\n') {
        out.push('\n');
    }
    Some((snap, out))
}

/// Whether `child` (a CST node's text range) intersects the
/// caller's selection. Zero-width selections (a cursor with no
/// extent) are handled specially: the cursor counts as "inside"
/// a child if the cursor is strictly inside the child's range or
/// is exactly at the child's start. Boundary at the child's end
/// is NOT a match — it belongs to the next child, matching
/// editors' "end-of-line cursor = start of next line" convention.
fn range_intersects(child: rowan::TextRange, sel: rowan::TextRange) -> bool {
    if sel.is_empty() {
        child.contains(sel.start()) || sel.start() == child.start()
    } else {
        child.start() < sel.end() && sel.start() < child.end()
    }
}

/// Compute the file-wide alignment columns for a parsed `SourceFile`.
///
/// Walks every Transaction's postings once, takes the max LHS
/// width (account + optional `flag `) and max number-text width,
/// and derives the column targets from them.
///
/// **`O(N_postings)`.** Public so consumers can pre-compute the
/// alignment once (typically at parse time) and pass the cached
/// `PostingAlignment` into [`format_node_with_alignment`] or
/// [`format_node_range_with_alignment`] — eliminates the per-call
/// walk in hot formatting paths (LSP format-on-type through a
/// parse error, repeat-format scripts, etc.).
///
/// **Tree-shape precondition.** `sf` must be a `SourceFile` whose
/// CST was produced by `parse_structured` (directly or transitively
/// via `parse_via_cst` / `parse`). Hand-built partial trees (e.g.,
/// a `GreenNodeBuilder` invocation for snippet formatting) silently
/// return `PostingAlignment::default()` because their wrapping
/// nodes fail the `ast::Directive::Transaction::cast` check.
/// Likewise, transactions wrapped in `ERROR_NODE` by mid-edit
/// error recovery are excluded — see
/// `parse_result_alignment_cache::mid_transaction_error_node` for
/// the pinned behavior. The function never panics on a partial
/// tree; it just returns the all-zero alignment for the no-postings
/// case.
///
/// **Pinning the contract.** `ParseResult::alignment` is populated
/// by calling this function during `parse_via_cst`; the equivalence
/// between the cached value and a fresh call is guaranteed by the
/// `parse_result_alignment_cache::*` regression tests (7 fixtures) in
/// this module.
#[must_use]
pub fn compute_alignment(sf: &SourceFile, style: GroupingStyle<'_>) -> PostingAlignment {
    let mut max_lhs: usize = 0;
    let mut max_num: usize = 0;
    // Tracks postings that actually render a number — the only ones that
    // participate in alignment. A file whose postings render no numbers
    // gets `PostingAlignment::default()`, matching the type docs.
    let mut any_aligned_posting = false;
    for directive in sf.directives() {
        let ast::Directive::Transaction(t) = directive else {
            continue;
        };
        for child in t.syntax().children() {
            let Some(p) = ast::Posting::cast(child) else {
                continue;
            };
            let mut lhs = 0usize;
            if let Some(flag) = p.flag() {
                lhs += flag.text().chars().count() + 1; // `! ` etc.
            }
            if let Some(account) = p.account() {
                lhs += account.text().chars().count();
            }

            // Only postings that render a number drive the alignment
            // column. `bean-format` computes the number column from the
            // prefixes of number-bearing lines only, so two kinds of
            // posting must NOT push the column right:
            //   - amount-less postings (the elided balancing leg, or a
            //     long account with no amount), and
            //   - currency-only amounts (`Assets:Cash USD`), which
            //     `emit_posting` prints with no number at all.
            // Counting either is why `rledger format` and `bean-format`
            // disagreed and round-tripping never converged (issue #1290).
            // `amount_number_text` is the shared predicate that keeps
            // this pre-pass in lockstep with `emit_posting`.
            if let Some(amt) = p.amount()
                && let Some(text) = amount_number_text(&amt, style)
            {
                any_aligned_posting = true;
                max_lhs = max_lhs.max(lhs);
                max_num = max_num.max(text.chars().count());
            }
        }
    }
    if !any_aligned_posting {
        return PostingAlignment::default();
    }
    // 2 spaces between the longest account end and the number field,
    // matching the conventional Beancount layout.
    PostingAlignment {
        number_col: INDENT.len() + max_lhs + 2,
        number_width: max_num,
    }
}

/// The rendered number / arithmetic-expression text of an amount *if it
/// renders a number*, or `None` when it renders nothing (a currency-only
/// amount like `USD`, whose value text is empty). EXCLUDES the trailing
/// currency; sign (if any) is included.
///
/// This is the single source of truth for "does this posting line have a
/// number?". Both the file-wide alignment pre-pass ([`compute_alignment`])
/// and the emitter ([`emit_posting`]) consult it, so they can never
/// disagree about which postings participate in alignment — the bug
/// class behind #1290 (amount-less postings) and its currency-only
/// sibling.
fn amount_number_text(amt: &ast::Amount, group: GroupingStyle<'_>) -> Option<String> {
    let text = amount_value_text(amt, group);
    (!text.is_empty()).then_some(text)
}

/// Render an amount's value portion (number or arithmetic
/// expression) as a string, EXCLUDING the trailing currency.
/// Mirrors the value half of [`format_amount`].
fn amount_value_text(amt: &ast::Amount, group: GroupingStyle<'_>) -> String {
    let mut buf = String::new();
    if amt.is_arithmetic() {
        emit_amount_subnode_expression(amt.syntax(), group, &mut buf);
        return buf;
    }
    if let Some(sign) = amt.sign()
        && sign.is_minus()
    {
        buf.push('-');
    }
    if let Some(n) = amt.number() {
        // `amt.currency()` is a CST child lookup, and `groups()` ignores the
        // currency entirely under the default style — so resolving it first
        // costs a tree walk to answer a question already settled. That is per
        // posting, per PARSE, because `compute_alignment` runs on every parse
        // (see `convert.rs`), long before anything asks to be formatted. It
        // measured 1.35% of the load pipeline's instructions.
        //
        // Same short-circuit `emit_amount_expression` already applies to
        // `run_currency`; these two sites were missed when that was added.
        let grouped = group.groups_anything()
            && group.groups(amt.currency().as_ref().map(ast::CurrencyName::text));
        buf.push_str(&canonical_number(n.text(), grouped));
    }
    buf
}

fn emit_directive(
    d: &ast::Directive,
    align: PostingAlignment,
    group: GroupingStyle<'_>,
    out: &mut String,
) {
    // Leading inter-directive trivia: COMMENT tokens that sit
    // BEFORE the directive's first content token. Per phase-2.0
    // trivia attachment, these live inside the directive's syntax
    // node — emit them as their own lines BEFORE the canonical
    // content.
    emit_leading_comments(d.syntax(), out);

    // Capture an optional same-line trailing comment so we can
    // splice it back in immediately before the directive's
    // terminating NEWLINE — see the comment-aware emit loop at
    // the bottom of this function.
    let trailing = collect_trailing_comment(d.syntax());

    let len_before = out.len();
    match d {
        ast::Directive::Open(d) => emit_open(d, group, out),
        ast::Directive::Close(d) => emit_close(d, group, out),
        ast::Directive::Commodity(d) => emit_commodity(d, group, out),
        ast::Directive::Note(d) => emit_note(d, group, out),
        ast::Directive::Event(d) => emit_event(d, group, out),
        ast::Directive::Query(d) => emit_query(d, group, out),
        ast::Directive::Pad(d) => emit_pad(d, group, out),
        ast::Directive::Document(d) => emit_document(d, group, out),
        ast::Directive::Price(d) => emit_price(d, group, out),
        ast::Directive::Balance(d) => emit_balance(d, group, out),
        ast::Directive::Custom(d) => emit_custom(d, group, out),
        ast::Directive::Option(d) => emit_option(d, out),
        ast::Directive::Include(d) => emit_include(d, out),
        ast::Directive::Plugin(d) => emit_plugin(d, out),
        ast::Directive::Pushtag(d) => emit_pushtag(d, out),
        ast::Directive::Poptag(d) => emit_poptag(d, out),
        ast::Directive::Pushmeta(d) => emit_pushmeta(d, group, out),
        ast::Directive::Popmeta(d) => emit_popmeta(d, out),
        ast::Directive::Transaction(d) => emit_transaction(d, align, group, out),
    }
    // Splice the same-line trailing comment in: find the FIRST '\n'
    // after `len_before` (= end of the directive's header line in
    // the emitted bytes) and insert `" ; comment"` before it. For
    // single-line directives the first '\n' is also the only one
    // and this lands the comment on the directive line. For multi-
    // line transactions it lands the comment on the header line
    // (where the source had it), not after the body.
    if let Some(c) = trailing
        && let Some(newline_rel) = out[len_before..].find('\n')
    {
        let insert_at = len_before + newline_rel;
        let mut splice = String::with_capacity(c.len() + 1);
        splice.push(' ');
        splice.push_str(&c);
        out.insert_str(insert_at, &splice);
    }
}

/// Emit an `ERROR_NODE`'s text verbatim, so `format` never deletes content it
/// could not parse (#1335) — chiefly org-mode `*` section headers and the
/// comments grouped with them. Only trailing whitespace per line is stripped
/// (the formatter's no-trailing-space policy) and the node's trailing newlines
/// are collapsed to one; everything else — including blank lines, comments and
/// the unparsable lines themselves — is preserved exactly as written.
fn emit_error_node(node: &crate::SyntaxNode, out: &mut String) {
    let text = node.text().to_string();
    // Trim leading AND trailing blank lines: the caller emits the leading
    // blank lines (via `leading_blank_lines`) so emitting them here too would
    // double-count them and break idempotence. Internal blank lines and the
    // content (org headers, grouped comments) are preserved.
    for line in text.trim_matches(['\n', '\r']).split('\n') {
        out.push_str(line.trim_end());
        out.push('\n');
    }
}

/// Number of blank lines the author left immediately before this
/// directive's first visible line (its leading comment, if any, else
/// its content). Each NEWLINE in the leading trivia that precedes the
/// first comment / content token is exactly one blank line: the
/// previous directive owns its own terminator NEWLINE (the Directive-
/// Terminator Rule), so this node's leading NEWLINEs are purely the
/// blank gap, with no off-by-one. WHITESPACE-only "blank" lines count
/// too (the NEWLINE that ends them is included). Scanning stops at the
/// first comment or content token, so a blank line sitting *between* a
/// leading comment and the directive's content is not counted here
/// (that gap is collapsed by `emit_leading_comments`, as before).
fn leading_blank_lines(node: &crate::SyntaxNode) -> usize {
    let mut blanks = 0;
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            break;
        };
        match t.kind() {
            crate::SyntaxKind::NEWLINE => blanks += 1,
            crate::SyntaxKind::WHITESPACE => {}
            // First comment or content token — past the leading gap.
            _ => break,
        }
    }
    blanks
}

/// Walk the directive's direct-child tokens until the first
/// non-trivia token, emitting each `COMMENT` (and `PERCENT_COMMENT`)
/// on its own line. Whitespace and newlines in the leading region
/// are ignored — the canonical form controls inter-directive
/// blank-line spacing separately.
fn emit_leading_comments(node: &crate::SyntaxNode, out: &mut String) {
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            break;
        };
        match t.kind() {
            crate::SyntaxKind::COMMENT | crate::SyntaxKind::PERCENT_COMMENT => {
                out.push_str(t.text().trim_end_matches(['\n', '\r']));
                out.push('\n');
            }
            crate::SyntaxKind::WHITESPACE | crate::SyntaxKind::NEWLINE => {}
            _ => break,
        }
    }
}

/// Return the directive's same-line trailing comment (if any) —
/// the COMMENT token that appears between the LAST non-trivia
/// content token and the directive-terminating NEWLINE on the
/// header line. Returns the verbatim comment text (no trailing
/// newline).
fn collect_trailing_comment(node: &crate::SyntaxNode) -> Option<String> {
    // Find the directive-header terminating NEWLINE: the FIRST
    // direct-child NEWLINE that follows at least one non-trivia
    // content token. (For single-line directives there's only one
    // NEWLINE; for transactions the header line is the first
    // NEWLINE, after which postings/metadata follow.)
    let mut header_nl_idx: Option<usize> = None;
    let mut saw_content = false;
    let tokens: Vec<crate::SyntaxToken> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .collect();
    for (i, t) in tokens.iter().enumerate() {
        let k = t.kind();
        if k == crate::SyntaxKind::NEWLINE && saw_content {
            header_nl_idx = Some(i);
            break;
        }
        if !matches!(
            k,
            crate::SyntaxKind::WHITESPACE
                | crate::SyntaxKind::NEWLINE
                | crate::SyntaxKind::COMMENT
                | crate::SyntaxKind::PERCENT_COMMENT
        ) {
            saw_content = true;
        }
    }
    // EOF-without-newline fallback: if there is no header-
    // terminating NEWLINE, the directive runs to the end of the
    // file. Scan from the LAST token instead. A `?` early-return
    // here previously dropped same-line trailing comments at the
    // final line of a file that lacked a trailing newline, e.g.
    // `2024-01-15 open Assets:A ; trailing` (no `\n`). The
    // canonical formatter restores the trailing newline, but the
    // comment was already gone.
    let nl_idx = header_nl_idx.unwrap_or(tokens.len());
    // Scan backwards from the header NEWLINE (or EOF): the
    // trailing comment is the last COMMENT before the NEWLINE
    // separated only by WHITESPACE.
    for i in (0..nl_idx).rev() {
        let k = tokens[i].kind();
        if matches!(
            k,
            crate::SyntaxKind::COMMENT | crate::SyntaxKind::PERCENT_COMMENT
        ) {
            return Some(tokens[i].text().to_string());
        }
        if k != crate::SyntaxKind::WHITESPACE {
            return None;
        }
    }
    None
}

// ---- Single-line directives ------------------------------------

fn emit_open(d: &ast::OpenDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let account = d
        .account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" open ");
    out.push_str(&account);
    // The currency constraint list is comma-separated (`USD,EUR`), not
    // space-separated — emitting spaces produces invalid beancount (#1405).
    for (i, currency) in d.currencies().enumerate() {
        out.push_str(if i == 0 { " " } else { "," });
        out.push_str(currency.text());
    }
    if let Some(booking) = d.booking_method() {
        // `booking.text()` includes the surrounding quotes.
        out.push(' ');
        out.push_str(booking.text());
    }
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_close(d: &ast::CloseDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let account = d
        .account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" close ");
    out.push_str(&account);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_commodity(d: &ast::CommodityDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let currency = d
        .currency()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" commodity ");
    out.push_str(&currency);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_note(d: &ast::NoteDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let account = d
        .account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    let text = d.text().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str(&date);
    out.push_str(" note ");
    out.push_str(&account);
    out.push(' ');
    out.push_str(&text);
    // beancount v3 accepts tags and links on a note header, and dropping them
    // here silently deleted user data on every `rledger format` run (#2184).
    emit_header_tags_and_links(d.syntax(), out);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_event(d: &ast::EventDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let event_type = d
        .event_type()
        .map(|s| s.text().to_string())
        .unwrap_or_default();
    let value = d.value().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str(&date);
    out.push_str(" event ");
    out.push_str(&event_type);
    out.push(' ');
    out.push_str(&value);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_query(d: &ast::QueryDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let name = d.name().map(|s| s.text().to_string()).unwrap_or_default();
    let query = d.query().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str(&date);
    out.push_str(" query ");
    out.push_str(&name);
    out.push(' ');
    out.push_str(&query);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_pad(d: &ast::PadDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let target = d
        .target_account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    let source = d
        .source_account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" pad ");
    out.push_str(&target);
    out.push(' ');
    out.push_str(&source);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

/// Re-emit the trailing `#tag` / `^link` tokens of a directive header.
///
/// The typed AST has no accessor for them, so this walks direct-child tokens.
/// Two things make that harder than it sounds, and both are bugs that have
/// been fixed here before:
///
/// - Skip LEADING trivia. A blank line before a non-first directive attaches
///   its NEWLINE inside the node, so a walk that stopped at "the first
///   NEWLINE" would stop before the header and drop every tag (#1321 in the
///   transaction path, and #2189 in the converter, which had the same shape).
/// - Skip leading COMMENT lines too, not just whitespace. A comment before a
///   non-first directive attaches inside this node (Directive-Terminator
///   Rule); treating it as content would flip `seen_content`, break at the
///   comment's own NEWLINE, and drop the real header tags.
///
/// `document` carried this walk alone while `note` had none, so formatting a
/// note deleted its tags and links (#2184). Sharing it is what keeps the two
/// from drifting apart again.
fn emit_header_tags_and_links(node: &crate::SyntaxNode, out: &mut String) {
    let mut seen_content = false;
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            // A child NODE is header content, not a reason to stop. For today's
            // two callers it can only be a META_ENTRY, which sits past the
            // header newline the walk has already broken on -- but stopping
            // here would mean a header that ever gains a child node silently
            // loses the tags after it, and nothing would say so.
            seen_content = true;
            continue;
        };
        match t.kind() {
            crate::SyntaxKind::TAG | crate::SyntaxKind::LINK => {
                out.push(' ');
                out.push_str(t.text());
                seen_content = true;
            }
            crate::SyntaxKind::NEWLINE if seen_content => break,
            k if k.is_trivia() => {}
            _ => seen_content = true,
        }
    }
}

fn emit_document(d: &ast::DocumentDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let account = d
        .account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    let path = d.path().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str(&date);
    out.push_str(" document ");
    out.push_str(&account);
    out.push(' ');
    out.push_str(&path);
    emit_header_tags_and_links(d.syntax(), out);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_price(d: &ast::PriceDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let base = d
        .base_currency()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    let quote = d
        .quote_currency()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" price ");
    out.push_str(&base);
    out.push(' ');
    emit_amount_expression(d.syntax(), group, out);
    out.push(' ');
    out.push_str(&quote);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_balance(d: &ast::BalanceDirective, group: GroupingStyle<'_>, out: &mut String) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let account = d
        .account()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    let currency = d
        .currency()
        .map(|t| t.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" balance ");
    out.push_str(&account);
    out.push(' ');
    emit_amount_expression(d.syntax(), group, out);
    // `balance ACCOUNT AMOUNT [~ TOLERANCE] CURRENCY` — ONE currency, trailing,
    // covering both numbers. The tolerance's own `CURRENCY` token (if the source
    // repeated it) is deliberately dropped: emitting it as well produced
    // `0.00 USD ~ 1234.5 USD`, which is not the beancount form.
    if let Some((tolerance, _tol_currency)) = balance_tolerance(d.syntax(), group) {
        out.push_str(" ~ ");
        out.push_str(&tolerance);
    }
    out.push(' ');
    out.push_str(&currency);
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

fn emit_custom(d: &ast::CustomDirective, group: GroupingStyle<'_>, out: &mut String) {
    // `custom` / `pushmeta` values are not denominated in anything, so
    // they take the ledger-wide default.
    let run_group = group.groups(None);
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    let custom_type = d
        .custom_type()
        .map(|s| s.text().to_string())
        .unwrap_or_default();
    out.push_str(&date);
    out.push_str(" custom ");
    out.push_str(&custom_type);
    // Walk raw tokens after the type STRING and emit each value
    // with single-space separation. NUMBER + CURRENCY adjacent
    // counts as an Amount; emitted together with one space.
    let tokens: Vec<crate::SyntaxToken> = d
        .syntax()
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        .collect();
    // `seen_type` skips the leading DATE + CUSTOM_KW + type-STRING
    // tokens (already emitted above as the directive header); once
    // it flips true, every subsequent non-trivia token is a value
    // argument and gets emitted with single-space separation. An
    // adjacent NUMBER + CURRENCY pair is glued with a single space
    // (canonical Amount shape); the CURRENCY is NOT eaten as a
    // standalone arg next iteration.
    //
    // Beancount custom directives accept any mix of value kinds
    // including DATE — a `custom "type" 2024-06-15 100.00 USD`
    // shape has a DATE in value position. The previous version
    // skipped every DATE after seen_type, silently dropping such
    // user-provided date arguments.
    let mut seen_type = false;
    let mut i = 0;
    while i < tokens.len() {
        let t = &tokens[i];
        if !seen_type {
            if t.kind() == crate::SyntaxKind::STRING {
                seen_type = true;
            }
            i += 1;
            continue;
        }
        out.push(' ');
        if t.kind() == crate::SyntaxKind::NUMBER {
            out.push_str(&canonical_number(t.text(), run_group));
            if matches!(
                tokens.get(i + 1).map(rowan::SyntaxToken::kind),
                Some(crate::SyntaxKind::CURRENCY)
            ) {
                out.push(' ');
                out.push_str(tokens[i + 1].text());
                i += 2;
                continue;
            }
        } else {
            out.push_str(t.text());
        }
        i += 1;
    }
    out.push('\n');
    emit_meta_entries_of(d.syntax(), group, out);
}

// ---- Top-level non-dated directives -----------------------------

fn emit_option(d: &ast::OptionDirective, out: &mut String) {
    let key = d.key().map(|s| s.text().to_string()).unwrap_or_default();
    let value = d.value().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str("option ");
    out.push_str(&key);
    out.push(' ');
    out.push_str(&value);
    out.push('\n');
}

fn emit_include(d: &ast::IncludeDirective, out: &mut String) {
    let path = d.path().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str("include ");
    out.push_str(&path);
    out.push('\n');
}

fn emit_plugin(d: &ast::PluginDirective, out: &mut String) {
    let module = d.module().map(|s| s.text().to_string()).unwrap_or_default();
    out.push_str("plugin ");
    out.push_str(&module);
    if let Some(config) = d.config() {
        out.push(' ');
        out.push_str(config.text());
    }
    out.push('\n');
}

// ---- State directives (no metadata) -----------------------------

fn emit_pushtag(d: &ast::PushtagDirective, out: &mut String) {
    let tag = d.tag().map(|t| t.text().to_string()).unwrap_or_default();
    out.push_str("pushtag ");
    out.push_str(&tag);
    out.push('\n');
}

fn emit_poptag(d: &ast::PoptagDirective, out: &mut String) {
    let tag = d.tag().map(|t| t.text().to_string()).unwrap_or_default();
    out.push_str("poptag ");
    out.push_str(&tag);
    out.push('\n');
}

fn emit_pushmeta(d: &ast::PushmetaDirective, group: GroupingStyle<'_>, out: &mut String) {
    // `custom` / `pushmeta` values are not denominated in anything, so
    // they take the ledger-wide default.
    let run_group = group.groups(None);
    let key = d.key().map(|t| t.text().to_string()).unwrap_or_default();
    out.push_str("pushmeta ");
    out.push_str(&key);
    // Walk the value tokens after META_KEY, single-space separated.
    let mut past_key = false;
    for el in d.syntax().children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        if !past_key {
            if t.kind() == crate::SyntaxKind::META_KEY {
                past_key = true;
            }
            continue;
        }
        if is_trivia_kind(t.kind()) {
            continue;
        }
        out.push(' ');
        if t.kind() == crate::SyntaxKind::NUMBER {
            out.push_str(&canonical_number(t.text(), run_group));
        } else {
            out.push_str(t.text());
        }
    }
    out.push('\n');
}

fn emit_popmeta(d: &ast::PopmetaDirective, out: &mut String) {
    let key = d.key().map(|t| t.text().to_string()).unwrap_or_default();
    out.push_str("popmeta ");
    out.push_str(&key);
    out.push('\n');
}

// ---- Transaction + Posting --------------------------------------

fn emit_transaction(
    d: &ast::Transaction,
    align: PostingAlignment,
    group: GroupingStyle<'_>,
    out: &mut String,
) {
    let date = d.date().map(|t| t.text().to_string()).unwrap_or_default();
    out.push_str(&date);
    out.push(' ');
    out.push_str(&transaction_flag_string(d));
    if let Some(payee) = d.payee() {
        out.push(' ');
        out.push_str(payee.text());
    }
    if let Some(narration) = d.narration() {
        out.push(' ');
        out.push_str(narration.text());
    }
    // Header-region tags/links — emitted in source order
    // (typed `.tags()` / `.links()` accessors return each kind
    // grouped, which loses interleaving like `#a ^l #b`). Walk
    // direct-child tokens, stopping at the header-terminating
    // NEWLINE.
    //
    // `seen_content` guards against LEADING trivia: for any directive
    // after the first, the preceding blank line's NEWLINE attaches
    // inside this node before the date (the Directive-Terminator Rule).
    // The header terminator is the first NEWLINE *after* the date, not
    // a leading one — otherwise this loop would break immediately and
    // emit no header tags (#1321).
    let mut seen_content = false;
    for el in d.syntax().children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            break;
        };
        match t.kind() {
            crate::SyntaxKind::TAG | crate::SyntaxKind::LINK => {
                out.push(' ');
                out.push_str(t.text());
                seen_content = true;
            }
            crate::SyntaxKind::NEWLINE if seen_content => break,
            // Leading trivia before the date: whitespace, blank-line
            // NEWLINEs, AND comment lines (a comment before a non-first
            // directive attaches inside this node per the Directive-
            // Terminator Rule). Skipping only WHITESPACE/NEWLINE would
            // let a leading comment flip `seen_content`, break at the
            // comment's NEWLINE, and drop the real header tags/links.
            k if k.is_trivia() => {}
            // DATE / flag / STRING etc. — header content has begun.
            _ => seen_content = true,
        }
    }
    out.push('\n');
    // Body: a single source-order walk over the transaction's children,
    // emitting — in the order they appear — POSTING / META_ENTRY nodes, any
    // body-internal COMMENT lines (#1332: the formatter must not delete the
    // author's comments), and trailing body-line TAG / LINK continuation
    // tokens (valid Beancount per the body-line exemption).
    //
    // `seen_content` / `past_header` skip the header region exactly as the
    // header loop above does, so the header-trailing comment (spliced onto
    // the header line by `emit_directive`) and the header tags/links (already
    // emitted inline above) are not duplicated here. A leading blank-line
    // NEWLINE for any directive past the first is trivia and must not flip
    // `past_header` early (#1321).
    let mut past_header = false;
    let mut seen_content = false;
    for el in d.syntax().children_with_tokens() {
        match el {
            rowan::NodeOrToken::Node(n) => {
                // A POSTING / META_ENTRY node is definitively past the header.
                past_header = true;
                if let Some(p) = ast::Posting::cast(n.clone()) {
                    emit_posting(&p, align, group, out);
                } else if let Some(m) = ast::MetaEntry::cast(n) {
                    emit_meta_entry(&m, INDENT, group, out);
                }
            }
            rowan::NodeOrToken::Token(t) => {
                if !past_header {
                    match t.kind() {
                        crate::SyntaxKind::NEWLINE if seen_content => past_header = true,
                        k if k.is_trivia() => {}
                        // DATE / flag / STRING / header TAG / LINK: still header.
                        _ => seen_content = true,
                    }
                    continue;
                }
                // Body tokens: preserve comment-only lines and emit
                // continuation tags/links, each on its own indented line.
                match t.kind() {
                    crate::SyntaxKind::COMMENT | crate::SyntaxKind::PERCENT_COMMENT => {
                        out.push_str(INDENT);
                        out.push_str(t.text().trim_end_matches(['\n', '\r']));
                        out.push('\n');
                    }
                    crate::SyntaxKind::TAG | crate::SyntaxKind::LINK => {
                        out.push_str(INDENT);
                        out.push_str(t.text());
                        out.push('\n');
                    }
                    _ => {}
                }
            }
        }
    }
}

fn transaction_flag_string(d: &ast::Transaction) -> String {
    use crate::cst::ast::TransactionFlagKind;
    match d.flag() {
        None => "*".to_string(),
        Some(f) => match f.classify() {
            TransactionFlagKind::Star | TransactionFlagKind::Txn => "*".to_string(),
            TransactionFlagKind::Pending => "!".to_string(),
            TransactionFlagKind::Hash => "#".to_string(),
            TransactionFlagKind::Letter | TransactionFlagKind::CurrencyLetter => {
                f.text().to_string()
            }
        },
    }
}

fn emit_posting(
    p: &ast::Posting,
    align: PostingAlignment,
    group: GroupingStyle<'_>,
    out: &mut String,
) {
    // Posting-trailing comment (same-line, before the posting-line
    // NEWLINE) — capture upfront so we can splice it back in just
    // before that NEWLINE, preserving the user's attachment intent.
    let trailing = collect_trailing_comment(p.syntax());
    let posting_start = out.len();

    out.push_str(INDENT);
    let mut col = INDENT.len();
    if let Some(flag) = p.flag() {
        out.push_str(flag.text());
        out.push(' ');
        col += flag.text().chars().count() + 1;
    }
    let account_text = p
        .account()
        .map(|a| a.text().to_string())
        .unwrap_or_default();
    out.push_str(&account_text);
    col += account_text.chars().count();

    if let Some(amt) = p.amount() {
        // `amount_number_text` is the shared "does this render a number?"
        // predicate (see `compute_alignment`); a currency-only amount
        // returns `None` and prints no number.
        if let Some(value) = amount_number_text(&amt, group) {
            // Two stages of padding:
            //   1) Account end → start of number field (`number_col`).
            //      Fall back to 2 spaces when the LHS already exceeds
            //      the file-wide max (over-long account name).
            //   2) Inside the number field, left-pad to right-justify
            //      to `number_width`. Effect: the currency column
            //      lands at a single uniform position file-wide even
            //      when numbers have different widths or signs.
            let field_pad = align.number_col.saturating_sub(col).max(2);
            let justify_pad = align.number_width.saturating_sub(value.chars().count());
            for _ in 0..(field_pad + justify_pad) {
                out.push(' ');
            }
            out.push_str(&value);
            if let Some(c) = amt.currency() {
                out.push(' ');
                out.push_str(c.text());
            }
            if let Some(cs) = p.cost_spec() {
                out.push(' ');
                out.push_str(&format_cost_spec(&cs, group));
            }
            if let Some(pa) = p.price_annotation() {
                out.push(' ');
                out.push_str(&format_price_annotation(&pa, group));
            }
        } else {
            // No NUMBER, but the posting may still carry a currency, a cost
            // spec or a price. `Assets:Other USD` is valid beancount that
            // constrains interpolation to USD, and a units-less posting with
            // a cost or price is malformed but is still what the author
            // wrote. Emitting nothing here deleted all of it (#2142).
            //
            // Two spaces rather than the aligned number column: there is no
            // number to align, and `compute_alignment` deliberately excludes
            // these postings from the column width so a long currency-only
            // account cannot widen the file.
            for part in [
                amt.currency().map(|c| c.text().to_string()),
                p.cost_spec().map(|cs| format_cost_spec(&cs, group)),
                p.price_annotation()
                    .map(|pa| format_price_annotation(&pa, group)),
            ]
            .into_iter()
            .flatten()
            {
                out.push_str("  ");
                out.push_str(&part);
            }
        }
    } else {
        // No amount node at all, yet a cost spec or price annotation parsed:
        // the CST attaches them after the account whether or not an AMOUNT
        // was recognized. Same rule as above, keep whatever is there.
        //
        // An earlier version of this fix handled only the cost spec, so
        // `Assets:MSFT @@ 2000.00 USD` still lost its price. Both are listed
        // here so neither can be forgotten again.
        for part in [
            p.cost_spec().map(|cs| format_cost_spec(&cs, group)),
            p.price_annotation()
                .map(|pa| format_price_annotation(&pa, group)),
        ]
        .into_iter()
        .flatten()
        {
            out.push_str("  ");
            out.push_str(&part);
        }
    }
    out.push('\n');
    // Splice the trailing comment in BEFORE the posting-line
    // NEWLINE (the first '\n' in the emitted posting region).
    if let Some(c) = trailing
        && let Some(rel) = out[posting_start..].find('\n')
    {
        let mut splice = String::with_capacity(c.len() + 1);
        splice.push(' ');
        splice.push_str(&c);
        out.insert_str(posting_start + rel, &splice);
    }
    // Posting body: emit attached metadata AND posting-internal comment
    // lines in source order, indented 4 (deeper than the posting's 2).
    // Comment-only lines inside a posting attach as COMMENT tokens of the
    // POSTING node; walking children-with-tokens preserves them (#1337)
    // instead of dropping them. The posting's own header line is skipped via
    // the seen_content/past_header guard, so the same-line trailing comment
    // (spliced above) is not duplicated here.
    let mut past_header = false;
    let mut seen_content = false;
    for el in p.syntax().children_with_tokens() {
        match el {
            rowan::NodeOrToken::Node(n) => {
                // Header child nodes (AMOUNT / COST_SPEC / PRICE_ANNOTATION)
                // are emitted inline above and must NOT flip `past_header` —
                // only the posting-line NEWLINE does. Otherwise the same-line
                // trailing comment, which follows the AMOUNT node, would be
                // re-emitted here as a body comment. META_ENTRY nodes only
                // appear in the body, after `past_header` is already set.
                if let Some(m) = ast::MetaEntry::cast(n) {
                    emit_meta_entry(&m, "    ", group, out);
                }
            }
            rowan::NodeOrToken::Token(t) => {
                if !past_header {
                    match t.kind() {
                        crate::SyntaxKind::NEWLINE if seen_content => past_header = true,
                        k if k.is_trivia() => {}
                        _ => seen_content = true,
                    }
                    continue;
                }
                if matches!(
                    t.kind(),
                    crate::SyntaxKind::COMMENT | crate::SyntaxKind::PERCENT_COMMENT
                ) {
                    out.push_str("    ");
                    out.push_str(t.text().trim_end_matches(['\n', '\r']));
                    out.push('\n');
                }
            }
        }
    }
}

/// Format an `AMOUNT` (units + currency) in canonical form. For
/// arithmetic shapes, emits the expression with single-space
/// separators (parens tight); for plain shapes, emits
/// `NUMBER CURRENCY` with thousands separators stripped.
fn format_amount(amt: &ast::Amount, group: GroupingStyle<'_>) -> String {
    let mut out = String::new();
    if amt.is_arithmetic() {
        emit_amount_subnode_expression(amt.syntax(), group, &mut out);
        if let Some(c) = amt.currency() {
            if !out.is_empty() {
                out.push(' ');
            }
            out.push_str(c.text());
        }
        return out;
    }
    if let Some(sign) = amt.sign()
        && sign.is_minus()
    {
        out.push('-');
    }
    if let Some(n) = amt.number() {
        // `amt.currency()` is a CST child lookup, and `groups()` ignores the
        // currency entirely under the default style — so resolving it first
        // costs a tree walk to answer a question already settled. That is per
        // posting, per PARSE, because `compute_alignment` runs on every parse
        // (see `convert.rs`), long before anything asks to be formatted. It
        // measured 1.35% of the load pipeline's instructions.
        //
        // Same short-circuit `emit_amount_expression` already applies to
        // `run_currency`; these two sites were missed when that was added.
        let grouped = group.groups_anything()
            && group.groups(amt.currency().as_ref().map(ast::CurrencyName::text));
        out.push_str(&canonical_number(n.text(), grouped));
    }
    if let Some(c) = amt.currency() {
        if !out.is_empty() && !out.ends_with('-') {
            out.push(' ');
        }
        out.push_str(c.text());
    }
    out
}

/// Canonical form for cost specs: `{cost CCY}` (single-brace
/// per-unit), `{{cost CCY}}` (double-brace total), `{# cost CCY}`
/// (per-unit + total via opener), or the in-brace `{N # T CCY}`
/// shape preserved as-is with single-space normalization.
///
/// Commas separating cost components (`{N CCY, DATE, "label"}`)
/// stay tight against the preceding token; every other adjacent
/// token pair is joined with a single space.
fn format_cost_spec(cs: &ast::CostSpec, group: GroupingStyle<'_>) -> String {
    let (open, close) = if cs.is_total() {
        ("{{", "}}")
    } else if cs.is_per_unit_plus_total() {
        ("{#", "}")
    } else {
        ("{", "}")
    };
    // Collect inner content tokens (skip opener/closer/whitespace),
    // then route through write_canonical_token_sequence so the spacing rule
    // is identical to balance/price/AMOUNT-subnode arithmetic — most
    // importantly, unary `+`/`-` stays tight (`{-500 USD}`, not
    // `{- 500 USD}`) and COMMA stays tight.
    let inner_tokens: Vec<crate::SyntaxToken> = cs
        .syntax()
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::L_BRACE
                    | crate::SyntaxKind::R_BRACE
                    | crate::SyntaxKind::L_DOUBLE_BRACE
                    | crate::SyntaxKind::R_DOUBLE_BRACE
                    | crate::SyntaxKind::L_BRACE_HASH
                    | crate::SyntaxKind::WHITESPACE
                    | crate::SyntaxKind::NEWLINE
            )
        })
        .collect();
    let mut inner = String::new();
    write_canonical_token_sequence(&inner_tokens, group, &mut inner);
    // The `{#` opener is a two-character marker; canonical form
    // separates it from the first inner token with a single space
    // (matching the rendering in this function's rustdoc). `{` and
    // `{{` don't get inner padding per the canonical-form spec.
    if cs.is_per_unit_plus_total() && !inner.is_empty() {
        format!("{open} {inner}{close}")
    } else {
        format!("{open}{inner}{close}")
    }
}

/// Canonical price annotation: `@ amount` (per-unit) or
/// `@@ amount` (total).
fn format_price_annotation(pa: &ast::PriceAnnotation, group: GroupingStyle<'_>) -> String {
    let op = if pa.is_total() { "@@" } else { "@" };
    match pa.amount() {
        Some(a) => format!("{op} {}", format_amount(&a, group)),
        None => op.to_string(),
    }
}

// ---- Helpers ---------------------------------------------------

/// True for tokens that don't contribute content to the canonical
/// form: whitespace, newlines, every comment kind, and the
/// leading-file `BOM` token.
const fn is_trivia_kind(kind: crate::SyntaxKind) -> bool {
    matches!(
        kind,
        crate::SyntaxKind::WHITESPACE
            | crate::SyntaxKind::NEWLINE
            | crate::SyntaxKind::COMMENT
            | crate::SyntaxKind::PERCENT_COMMENT
            | crate::SyntaxKind::SHEBANG
            | crate::SyntaxKind::EMACS_DIRECTIVE
            | crate::SyntaxKind::BOM
    )
}

/// Render a `NUMBER` token in canonical form: the user's decimal-place count
/// is preserved, and digit grouping is imposed by `group` rather than by what
/// the source happened to contain.
///
/// `group == false` → `1,000.00` becomes `1000.00`. `group == true` → the
/// reverse: `1000.00` becomes `1,000.00`. Either way this is a TOTAL rewrite of
/// the grouping, so the formatter still yields one form per value — the rule
/// changes, not the guarantee.
///
/// Groups are always three digits, because that is the only shape the lexer
/// accepts (`(\d{1,3}(,\d{3})*|\d+)(\.\d*)?`). Anything else — Indian lakh
/// grouping, say — would emit text this parser then REJECTS, so widening this
/// needs a lexer change first, not just a formatter one.
fn canonical_number(text: &str, group: bool) -> std::borrow::Cow<'_, str> {
    // The overwhelmingly common numeral is already canonical: no separators to
    // strip and no grouping requested. Borrow it rather than allocating a copy
    // per numeral on the default formatter path.
    if !group && !text.contains(',') {
        return std::borrow::Cow::Borrowed(text);
    }
    let bare = text.replace(',', "");
    if !group {
        return std::borrow::Cow::Owned(bare);
    }
    let (int_part, frac) = match bare.split_once('.') {
        Some((i, f)) => (i, Some(f)),
        None => (bare.as_str(), None),
    };
    // Defensive: a non-digit integer part is not ours to regroup. Unreachable
    // for a lexed NUMBER, whose sign is a separate token.
    if int_part.is_empty() || !int_part.bytes().all(|b| b.is_ascii_digit()) {
        return std::borrow::Cow::Owned(bare);
    }
    let n = int_part.len();
    let mut out = String::with_capacity(bare.len() + n / 3);
    for (i, c) in int_part.chars().enumerate() {
        if i > 0 && (n - i) % 3 == 0 {
            out.push(',');
        }
        out.push(c);
    }
    if let Some(f) = frac {
        out.push('.');
        out.push_str(f);
    }
    std::borrow::Cow::Owned(out)
}

/// Emit the arithmetic expression of a `PRICE` / `BALANCE`
/// directive: tokens from the first expression-starting token
/// (`NUMBER`, unary `+`/`-`, or `(`) up to (but not including) the
/// first `CURRENCY` at paren-depth 0. Spacing rules per
/// [`write_canonical_token_sequence`].
///
/// **Why the predicate must allow `PLUS` / `MINUS` / `L_PAREN`,
/// not just `NUMBER`.** A previous version skipped tokens until
/// it hit a `NUMBER`, which silently dropped leading unary signs
/// and opening parens — flipping the sign on inputs like
/// `2024-01-15 price USD -1.00 EUR` (formatted to `1.00 EUR`) and
/// corrupting parenthesized expressions like
/// `2024-01-15 balance Assets:A (1 + 2) USD` (formatted to
/// `1 + 2) USD USD`). Sign drift in BALANCE / PRICE is silent data
/// corruption — a balance assertion that previously asserted a
/// debit would assert a credit after a round-trip.
fn emit_amount_expression(node: &crate::SyntaxNode, group: GroupingStyle<'_>, out: &mut String) {
    let raw: Vec<crate::SyntaxToken> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        .skip_while(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::NUMBER
                    | crate::SyntaxKind::PLUS
                    | crate::SyntaxKind::MINUS
                    | crate::SyntaxKind::L_PAREN
            )
        })
        .collect();
    let mut depth: i32 = 0;
    let mut first_currency_idx: Option<usize> = None;
    for (i, t) in raw.iter().enumerate() {
        match t.kind() {
            crate::SyntaxKind::L_PAREN => depth += 1,
            crate::SyntaxKind::R_PAREN => depth -= 1,
            // A `~ tolerance` clause ENDS the asserted amount. `emit_balance`
            // emits it separately via `balance_tolerance`, so running past the
            // tilde here emitted it twice: `0.00 ~ 1234.5 USD` came out as
            // `0.00 ~ 1234.5 USD ~ 1234.5 USD`. Stable but wrong, and very
            // likely not valid beancount — its balance grammar takes at most
            // one tolerance.
            crate::SyntaxKind::TILDE if depth == 0 && first_currency_idx.is_none() => {
                first_currency_idx = Some(i);
            }
            crate::SyntaxKind::CURRENCY if depth == 0 && first_currency_idx.is_none() => {
                first_currency_idx = Some(i);
            }
            _ => {}
        }
    }
    let end = first_currency_idx.unwrap_or(raw.len());
    write_canonical_token_sequence(&raw[..end], group, out);
}

/// Emit an `AMOUNT` subnode's expression region: every non-trivia
/// token minus the trailing `CURRENCY` (caller re-emits the
/// currency itself). Used by [`format_amount`] for arithmetic
/// posting amounts like `-(1.00 + 2.00) USD`.
fn emit_amount_subnode_expression(
    node: &crate::SyntaxNode,
    group: GroupingStyle<'_>,
    out: &mut String,
) {
    let mut tokens: Vec<crate::SyntaxToken> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        .collect();
    if let Some(last) = tokens.last()
        && last.kind() == crate::SyntaxKind::CURRENCY
    {
        tokens.pop();
    }
    write_canonical_token_sequence(&tokens, group, out);
}

/// Single dispatcher for the canonical spacing rules used by EVERY
/// token-sequence emit path: balance / price arithmetic, AMOUNT
/// subnodes, cost-spec interiors, and metadata values. There is no
/// separate path; each call site collects the relevant non-trivia
/// tokens and routes them through here so the rules cannot drift
/// between contexts.
///
/// Rules:
///
/// - single space between adjacent operands / binary operators
/// - no space after `(` or before `)` (parens stay tight)
/// - no space after a unary `+` / `-` (one that opens the run
///   or follows `(` or another operator)
/// - no space before `,` (commas in cost-spec component lists
///   stay tight against the preceding token)
///
/// **Adding a new `SyntaxKind` to the formatter implies thinking
/// about its effect on every call site of this function.** A new
/// operator-like kind added to `is_op` will silently change cost-
/// spec and metadata spacing too; a new bracket-like kind needs
/// its own rule. The corpus-level idempotence test
/// (`idempotence_corpus_sweep`) is the safety net that catches
/// drifts.
/// The currency a token run is denominated in: the LAST `CURRENCY` token in it.
///
/// A cost spec (`{1234.56 USD}`) and a balance tolerance carry their own
/// currency, and it is the one whose declaration governs their numerals. Taking
/// the ledger default instead is how `{1,234,567.89 USD}` came out grouped
/// while a plain `1234567.89 USD` posting in the same file stayed bare — USD
/// had declared `render_commas: FALSE` and only one of the two honored it.
fn run_currency(tokens: &[crate::SyntaxToken]) -> Option<&str> {
    tokens
        .iter()
        .rev()
        .find(|t| t.kind() == crate::SyntaxKind::CURRENCY)
        .map(rowan::SyntaxToken::text)
}

fn write_canonical_token_sequence(
    tokens: &[crate::SyntaxToken],
    group: GroupingStyle<'_>,
    out: &mut String,
) {
    // `&&` short-circuits, so a ledger that declares no grouping never pays
    // for the currency scan. `format` walks whole ledgers; see the
    // `profile_format` example.
    let run_group = group.groups_anything() && group.groups(run_currency(tokens));
    let is_op = |k: crate::SyntaxKind| {
        matches!(
            k,
            crate::SyntaxKind::PLUS
                | crate::SyntaxKind::MINUS
                | crate::SyntaxKind::STAR
                | crate::SyntaxKind::SLASH
        )
    };
    let mut prev_kind: Option<crate::SyntaxKind> = None;
    let mut prev_was_unary = false;
    for t in tokens {
        let kind = t.kind();
        let is_unary = is_op(kind)
            && match prev_kind {
                None => true,
                Some(p) => p == crate::SyntaxKind::L_PAREN || is_op(p),
            };
        let need_space = match prev_kind {
            None => false,
            Some(prev) => {
                prev != crate::SyntaxKind::L_PAREN
                    && kind != crate::SyntaxKind::R_PAREN
                    && kind != crate::SyntaxKind::COMMA
                    && !prev_was_unary
            }
        };
        if need_space {
            out.push(' ');
        }
        if kind == crate::SyntaxKind::NUMBER {
            out.push_str(&canonical_number(t.text(), run_group));
        } else {
            out.push_str(t.text());
        }
        prev_kind = Some(kind);
        prev_was_unary = is_unary;
    }
}

/// Extract a balance directive's optional tolerance — the
/// `NUMBER` after the first `TILDE`, plus an optional trailing
/// `CURRENCY` at paren-depth 0.
fn balance_tolerance(
    node: &crate::SyntaxNode,
    group: GroupingStyle<'_>,
) -> Option<(String, Option<String>)> {
    // `balance Assets:A 100.00 ~ 0.05 USD` — one currency covers the asserted
    // amount and the tolerance, and it trails both, so resolve it up front
    // rather than mid-walk.
    let run_group = group.groups_anything() && {
        // Only materialized when something groups — see above.
        let toks: Vec<crate::SyntaxToken> = node
            .children_with_tokens()
            .filter_map(rowan::NodeOrToken::into_token)
            .collect();
        group.groups(run_currency(&toks))
    };
    // The tolerance is an EXPRESSION, not a number: `~ 0.005 + 0.005 USD` and
    // `~ 0.005 * 2 USD` are both legal and both mean 0.010. Keeping only the
    // first NUMBER -- which this did -- rewrote them as `~ 0.005 USD`, halving
    // the tolerance. That reparsed cleanly and asserted something else, so
    // `rledger format` could turn a passing ledger into a failing one with no
    // diagnostic on either side. Same bug as #1944 in the parser's
    // `extract_balance_tolerance`, which was fixed there and left here.
    let mut past_tilde = false;
    let mut expr: Vec<String> = Vec::new();
    let mut currency: Option<String> = None;
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        if !past_tilde {
            if t.kind() == crate::SyntaxKind::TILDE {
                past_tilde = true;
            }
            continue;
        }
        match t.kind() {
            crate::SyntaxKind::NUMBER => {
                expr.push(canonical_number(t.text(), run_group).into_owned());
            }
            crate::SyntaxKind::PLUS
            | crate::SyntaxKind::MINUS
            | crate::SyntaxKind::STAR
            | crate::SyntaxKind::SLASH => expr.push(t.text().to_string()),
            // Parens bind to their neighbors rather than taking spaces, so
            // they are joined below rather than pushed as operands.
            crate::SyntaxKind::L_PAREN => expr.push("(".to_string()),
            crate::SyntaxKind::R_PAREN => expr.push(")".to_string()),
            crate::SyntaxKind::CURRENCY if !expr.is_empty() && currency.is_none() => {
                currency = Some(t.text().to_string());
            }
            _ => {}
        }
    }
    if expr.is_empty() {
        return None;
    }
    // A `+`/`-` is UNARY when nothing that can end an operand precedes it, and
    // a unary sign binds to its number: `~ -0.01`, not `~ - 0.01`. The amount
    // on the same line already renders `-1.00` tight, so spacing it here made
    // one directive disagree with itself (`-1.00 ~ - 0.01 USD`).
    let ends_operand = |t: &str| t == ")" || t.starts_with(|c: char| c.is_ascii_digit());
    let mut rendered = String::new();
    let mut tight_next = false;
    for (i, tok) in expr.iter().enumerate() {
        let prev = i.checked_sub(1).and_then(|j| expr.get(j));
        let unary = matches!(tok.as_str(), "+" | "-") && prev.is_none_or(|p| !ends_operand(p));
        let needs_space = i > 0 && tok != ")" && !tight_next && prev.is_none_or(|p| p != "(");
        if needs_space {
            rendered.push(' ');
        }
        rendered.push_str(tok);
        tight_next = unary;
    }
    Some((rendered, currency))
}

// ---- Metadata --------------------------------------------------

/// Walk a directive's direct-child `META_ENTRY` nodes and emit
/// each on its own indented line in canonical form (`indent + KEY:
/// value\n`). Most directive types don't have a `.meta_entries()`
/// accessor on their typed wrapper; we walk the syntax node
/// directly to stay uniform.
fn emit_meta_entries_of(node: &crate::SyntaxNode, group: GroupingStyle<'_>, out: &mut String) {
    // Source-order walk so body-internal COMMENT lines are preserved
    // alongside the metadata entries (#1332). The header region (up to and
    // including the header-terminating NEWLINE) is skipped so the
    // header-trailing comment — spliced onto the header line by
    // `emit_directive` — is not duplicated here.
    let mut past_header = false;
    let mut seen_content = false;
    for el in node.children_with_tokens() {
        match el {
            rowan::NodeOrToken::Node(n) => {
                past_header = true;
                if let Some(entry) = MetaEntry::cast(n) {
                    emit_meta_entry(&entry, INDENT, group, out);
                }
            }
            rowan::NodeOrToken::Token(t) => {
                if !past_header {
                    match t.kind() {
                        crate::SyntaxKind::NEWLINE if seen_content => past_header = true,
                        k if k.is_trivia() => {}
                        _ => seen_content = true,
                    }
                    continue;
                }
                if matches!(
                    t.kind(),
                    crate::SyntaxKind::COMMENT | crate::SyntaxKind::PERCENT_COMMENT
                ) {
                    out.push_str(INDENT);
                    out.push_str(t.text().trim_end_matches(['\n', '\r']));
                    out.push('\n');
                }
            }
        }
    }
}

/// Canonical emit for a single `META_ENTRY`. Walks non-trivia
/// tokens, prints them with single-space separation, and
/// normalizes numbers via [`canonical_number`]. The `META_KEY`
/// token already includes the trailing colon (e.g. `note:`); the
/// value side gets the same NUMBER + CURRENCY gluing rule the
/// rest of the formatter uses elsewhere.
///
/// Two semantically-equivalent inputs (e.g. `foo: "bar"` and
/// `foo:    "bar"`) produce byte-identical output — the
/// gofmt-style invariant the file rustdoc promises.
fn emit_meta_entry(m: &MetaEntry, indent: &str, group: GroupingStyle<'_>, out: &mut String) {
    out.push_str(indent);
    // Split the META_ENTRY's non-trivia tokens into [META_KEY,
    // value*]. The META_KEY token already includes the trailing
    // colon (e.g. `note:`); the value tokens go through
    // write_canonical_token_sequence so the spacing rules — unary +/-
    // tight, COMMA tight, paren-tight, NUMBER canonicalized — are
    // shared with the balance/price/cost-spec/posting-amount paths.
    let content: Vec<crate::SyntaxToken> = m
        .syntax()
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::WHITESPACE | crate::SyntaxKind::NEWLINE
            )
        })
        .collect();
    let mut iter = content.iter();
    if let Some(key) = iter.next() {
        out.push_str(key.text());
    }
    let value_tokens: Vec<crate::SyntaxToken> = iter.cloned().collect();
    if !value_tokens.is_empty() {
        out.push(' ');
        write_canonical_token_sequence(&value_tokens, group, out);
    }
    out.push('\n');
}

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

    #[test]
    fn empty_input_yields_single_newline() {
        assert_eq!(format_source(""), "\n");
    }

    #[test]
    fn open_directive_canonical() {
        let src = "2024-01-15   open    Assets:Cash\n";
        assert_eq!(format_source(src), "2024-01-15 open Assets:Cash\n");
    }

    #[test]
    fn open_with_currencies_and_booking_canonical() {
        // The currency constraint list is comma-separated; emitting spaces
        // produced invalid beancount (#1405).
        let src = "2024-01-15 open Assets:Brokerage USD,EUR \"STRICT\"\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 open Assets:Brokerage USD,EUR \"STRICT\"\n"
        );
    }

    /// Regression for #1405: `format` must keep the open currency list
    /// comma-separated, not rewrite it space-separated (invalid syntax), and
    /// the result must be idempotent.
    #[test]
    fn open_currency_list_stays_comma_separated() {
        let src = "2026-01-01 open Assets:Wallet USD,EUR\n";
        let once = format_source(src);
        assert_eq!(once, "2026-01-01 open Assets:Wallet USD,EUR\n");
        assert_eq!(format_source(&once), once, "format must be idempotent");
    }

    #[test]
    fn close_directive_canonical() {
        let src = "2024-12-31 close Assets:Cash\n";
        assert_eq!(format_source(src), "2024-12-31 close Assets:Cash\n");
    }

    #[test]
    fn commodity_directive_canonical() {
        let src = "2024-01-01 commodity HOOL\n";
        assert_eq!(format_source(src), "2024-01-01 commodity HOOL\n");
    }

    #[test]
    fn blank_lines_between_directives_preserved() {
        // #1325: the formatter preserves the author's inter-directive
        // blank lines rather than normalizing to exactly one (matching
        // Python bean-format and the rest of the beancount lineage).

        // Grouped (no blank in source) stays grouped — not double-spaced.
        let grouped = "2024-01-01 open Assets:A\n2024-01-02 open Assets:B\n";
        assert_eq!(format_source(grouped), grouped);

        // One blank is preserved as one.
        let one = "2024-01-01 open Assets:A\n\n2024-01-02 open Assets:B\n";
        assert_eq!(format_source(one), one);

        // Two blanks are preserved as two (not collapsed).
        let two = "2024-01-01 open Assets:A\n\n\n2024-01-02 open Assets:B\n";
        assert_eq!(format_source(two), two);

        // A whitespace-only "blank" line still counts as one blank line
        // (its trailing whitespace is stripped, leaving an empty line).
        let ws_blank = "2024-01-01 open Assets:A\n   \n2024-01-02 open Assets:B\n";
        assert_eq!(
            format_source(ws_blank),
            "2024-01-01 open Assets:A\n\n2024-01-02 open Assets:B\n"
        );
    }

    #[test]
    fn trailing_newline_always_present() {
        let src = "2024-01-01 open Assets:A";
        let formatted = format_source(src);
        assert!(formatted.ends_with('\n'));
        assert!(!formatted.ends_with("\n\n"));
    }

    #[test]
    fn idempotent_on_canonical_input() {
        let src = "2024-01-01 open Assets:A\n\n2024-01-02 close Assets:A\n";
        let once = format_source(src);
        let twice = format_source(&once);
        assert_eq!(once, twice);
    }

    #[test]
    fn note_canonical() {
        let src = "2024-01-15   note   Assets:Cash   \"a note\"\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 note Assets:Cash \"a note\"\n"
        );
    }

    #[test]
    fn event_canonical() {
        let src = "2024-01-15  event  \"location\"   \"NYC\"\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 event \"location\" \"NYC\"\n"
        );
    }

    #[test]
    fn query_canonical() {
        let src = "2024-01-15 query \"q1\" \"SELECT account\"\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 query \"q1\" \"SELECT account\"\n"
        );
    }

    #[test]
    fn pad_canonical() {
        let src = "2024-01-15  pad   Assets:A   Equity:Opening\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 pad Assets:A Equity:Opening\n"
        );
    }

    #[test]
    fn document_with_tags_and_links_canonical() {
        let src = "2024-06-01 document Assets:Bank \"stmt.pdf\" #q1 ^scan42 #urgent\n";
        assert_eq!(
            format_source(src),
            "2024-06-01 document Assets:Bank \"stmt.pdf\" #q1 ^scan42 #urgent\n"
        );
    }

    #[test]
    fn issue_1321_document_tags_links_idempotent_across_directives() {
        // Same class as the transaction case, in `document` directives:
        // the 2nd+ document's trailing tags/links were dropped on a
        // reformat (found by the #1323 corpus idempotence check). Assert
        // the fixed-point property: re-formatting must not change (and
        // must not drop the tags/links of the second document).
        let src = "\
2013-05-18 document Assets:Bank \"/a.pdf\" #tag1 ^link1
2013-05-19 document Assets:Bank \"/b.pdf\" #tag2 ^link2
";
        let once = format_source(src);
        assert_eq!(format_source(&once), once, "format must be idempotent");
        assert!(
            once.contains("#tag2") && once.contains("^link2"),
            "the second document's tags/links must survive formatting; got:\n{once}"
        );
    }

    #[test]
    fn issue_1321_header_tags_links_idempotent_across_transactions() {
        // Header tags/links must stay on the header line for EVERY
        // transaction, not just the first. Regression for #1321 where
        // the 2nd+ transaction's header tags/links got migrated to
        // continuation lines.
        let src = "\
2024-01-15 * \"x\" #tag1 ^link1 #tag2 ^link2
  Assets:Cash    -1.00 USD
  Expenses:Misc   1.00 USD

2024-01-16 * \"x\" #tag1 ^link1 #tag2 ^link2
  Assets:Cash    -1.00 USD
  Expenses:Misc   1.00 USD
";
        assert_eq!(
            format_source(src),
            src,
            "format must be a no-op (idempotent)"
        );
    }

    #[test]
    fn issue_1321_comment_before_transaction_keeps_header_tags() {
        // A comment line before a transaction is leading trivia attached
        // inside the transaction node (Directive-Terminator Rule), exactly
        // like a blank line. Skipping only WHITESPACE/NEWLINE let the
        // comment flip `seen_content`, break at the comment's NEWLINE, and
        // migrate the real header tags/links to continuation lines. The
        // header tags/links must stay on the header line. (Found by the
        // Copilot review of the #1321 fix.)
        let src = "\
2024-01-15 * \"first\" #h1 ^l1
  Assets:Cash    -1.00 USD
  Expenses:Misc   1.00 USD

; a comment before the second transaction
2024-01-16 * \"second\" #tag1 ^link1
  Assets:Cash    -2.00 USD
  Expenses:Misc   2.00 USD
";
        assert_eq!(
            format_source(src),
            src,
            "a leading comment must not migrate header tags/links to continuation lines"
        );
    }

    #[test]
    fn issue_1321_comment_before_document_keeps_tags() {
        // Document-directive variant of the comment-trivia case above.
        let src = "\
2013-05-18 document Assets:Bank \"/a.pdf\" #tag1 ^link1
; a comment before the second document
2013-05-19 document Assets:Bank \"/b.pdf\" #tag2 ^link2
";
        let once = format_source(src);
        assert_eq!(format_source(&once), once, "format must be idempotent");
        assert!(
            once.contains("\"/b.pdf\" #tag2 ^link2"),
            "the second document's tags/links must stay on its header line; got:\n{once}"
        );
    }

    #[test]
    fn issue_1332_body_comments_in_metadata_preserved() {
        // The formatter must NOT delete comment-only lines inside a
        // directive body (#1332). Here two commented-out `; price:` lines
        // sit between metadata entries in a `commodity` body; they must
        // survive, interleaved in source order, and the result is idempotent.
        let src = "\
2023-06-04 commodity EAM-VEUR ; cSpell: word VEUR
  name: \"Vanguard FTSE Developed Europe UCITS ETF EUR Dist\"
  ; price: \"EUR:alphavantage/price:VEUR.AS:EUR\"
  ; price: \"EUR:yahoo/VEUR.AS\"
  price: \"EUR:pricehist.beanprice.yahoo/VEUR.AS\"
";
        assert_eq!(
            format_source(src),
            src,
            "body comments must be preserved verbatim"
        );
        assert_eq!(format_source(&format_source(src)), format_source(src));
    }

    #[test]
    fn issue_1332_body_comments_between_postings_preserved() {
        // Same class, inside a transaction body: a comment-only line between
        // postings must survive (in source order, 2-space indent). Asserted
        // via preservation + idempotence rather than an exact match, since
        // amount alignment is also canonicalized.
        let src = "\
2024-01-15 * \"Cafe\" \"Latte\"
  Expenses:Coffee   4.50 USD
  ; was 5.00 before the discount
  Assets:Checking
";
        let out = format_source(src);
        assert!(
            out.contains("\n  ; was 5.00 before the discount\n"),
            "the body comment must be preserved on its own indented line; got:\n{out}"
        );
        // Order: the comment stays between the two postings.
        let coffee = out.find("Expenses:Coffee").unwrap();
        let comment = out.find("; was 5.00").unwrap();
        let checking = out.find("Assets:Checking").unwrap();
        assert!(
            coffee < comment && comment < checking,
            "comment must stay between postings:\n{out}"
        );
        assert_eq!(format_source(&out), out, "format must be idempotent");
    }

    #[test]
    fn issue_1335_org_headers_and_grouped_comments_preserved() {
        // The formatter must not delete unparsable content (#1335).
        // Org-mode `*` section headers parse into ERROR_NODEs, and comments
        // grouped with them get swallowed into the same node — previously all
        // dropped. They must survive, and the result must be idempotent.
        let src = "\
* Section A
;; comment between headers
;; second line
* Section B
2013-01-01 open Assets:X
";
        let out = format_source(src);
        // Use the exact `;;` needles: a single-`;` substring would still match
        // `;; ...` even if one `;` were dropped, weakening the regression.
        for needle in [
            "* Section A",
            ";; comment between headers",
            ";; second line",
            "* Section B",
            "2013-01-01 open Assets:X",
        ] {
            assert!(
                out.contains(needle),
                "lost {needle:?} on format; got:\n{out}"
            );
        }
        assert_eq!(format_source(&out), out, "format must be idempotent");
    }

    #[test]
    fn issue_1335_org_header_then_directive_keeps_header() {
        // A lone org header before a directive: the header is an ERROR_NODE
        // and must be kept (the comment here attaches to the directive and
        // was already preserved).
        let src = "* Accounts\n2013-01-01 open Assets:X\n";
        let out = format_source(src);
        assert!(
            out.contains("* Accounts"),
            "org header dropped; got:\n{out}"
        );
        assert_eq!(format_source(&out), out);
    }

    #[test]
    fn issue_1335_blank_lines_around_org_header_preserved() {
        // An ERROR_NODE is a top-level content block: the author's blank line
        // between an org header and the following directive is preserved (it
        // is not flushed), and the result is idempotent.
        let src = "* Accounts\n\n2013-01-01 open Assets:X\n";
        assert_eq!(
            format_source(src),
            src,
            "blank around org header must be kept"
        );
        assert_eq!(format_source(&format_source(src)), format_source(src));
    }

    #[test]
    fn issue_1337_posting_internal_comments_preserved() {
        // A comment on its own line inside a posting attaches as a COMMENT
        // token of the POSTING node; it must be preserved (#1337), not
        // dropped, and stay between its posting and the next.
        let src = "\
2024-01-15 * \"x\"
  Assets:A   1.00 USD
    ; posting-internal note
  Assets:B
";
        let out = format_source(src);
        assert!(
            out.contains("; posting-internal note"),
            "posting-internal comment dropped; got:\n{out}"
        );
        let a = out.find("Assets:A").unwrap();
        let c = out.find("; posting-internal note").unwrap();
        let b = out.find("Assets:B").unwrap();
        assert!(a < c && c < b, "comment must stay between postings:\n{out}");
        assert_eq!(format_source(&out), out, "format must be idempotent");
    }

    #[test]
    fn price_canonical_strips_thousands_separators() {
        let src = "2024-01-15 price USD  1,234.56 EUR\n";
        assert_eq!(format_source(src), "2024-01-15 price USD 1234.56 EUR\n");
    }

    #[test]
    fn price_arithmetic_canonicalizes_spacing() {
        let src = "2024-01-15 price USD 1/2 EUR\n";
        assert_eq!(format_source(src), "2024-01-15 price USD 1 / 2 EUR\n");
    }

    #[test]
    fn balance_canonical() {
        let src = "2024-01-15  balance  Assets:Cash   100.00  USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:Cash 100.00 USD\n"
        );
    }

    #[test]
    fn balance_with_tolerance_canonical() {
        // Beancount's form is `AMOUNT ~ TOLERANCE CURRENCY` — ONE trailing
        // currency covering both numbers, per its Precision & Tolerances docs
        // (`319.020 ~ 0.002 RGAGX`). This test previously asserted
        // `100.00 USD ~ 0.01 USD`, repeating the currency; that is not the
        // beancount form, and the emitter produced it by running the amount
        // expression past the tilde and then emitting the tolerance again.
        // Input that repeats the currency now normalizes to the canonical form.
        let src = "2024-01-15 balance Assets:Cash 100.00 USD ~ 0.01 USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:Cash 100.00 ~ 0.01 USD\n"
        );
    }

    #[test]
    fn balance_arithmetic_canonical() {
        let src = "2024-01-15 balance Assets:Cash  0.25 + 0.75  USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:Cash 0.25 + 0.75 USD\n"
        );
    }

    #[test]
    fn custom_canonical() {
        let src = "2024-01-01 custom \"budget\" Expenses:Food 500.00 USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-01 custom \"budget\" Expenses:Food 500.00 USD\n"
        );
    }

    #[test]
    fn option_canonical() {
        let src = "option   \"title\"   \"My Ledger\"\n";
        assert_eq!(format_source(src), "option \"title\" \"My Ledger\"\n");
    }

    #[test]
    fn include_canonical() {
        let src = "include  \"other.beancount\"\n";
        assert_eq!(format_source(src), "include \"other.beancount\"\n");
    }

    #[test]
    fn plugin_canonical_with_config() {
        let src = "plugin  \"beancount.plugins.unrealized\"  \"Unrealized\"\n";
        assert_eq!(
            format_source(src),
            "plugin \"beancount.plugins.unrealized\" \"Unrealized\"\n"
        );
    }

    #[test]
    fn plugin_canonical_without_config() {
        let src = "plugin   \"my.plugin\"\n";
        assert_eq!(format_source(src), "plugin \"my.plugin\"\n");
    }

    #[test]
    fn pushtag_poptag_canonical() {
        // No blank line in the source — preserved as grouped (#1325).
        let src = "pushtag  #active\npoptag  #active\n";
        assert_eq!(format_source(src), "pushtag #active\npoptag #active\n");
    }

    #[test]
    fn pushmeta_popmeta_canonical() {
        // No blank line in the source — preserved as grouped (#1325).
        let src = "pushmeta location: \"NYC\"\npopmeta location:\n";
        assert_eq!(
            format_source(src),
            "pushmeta location: \"NYC\"\npopmeta location:\n"
        );
    }

    // ---- Transaction tests ------------------------------------

    #[test]
    fn transaction_minimal_two_postings_aligns_amounts() {
        let src = "\
2024-01-15 * \"Coffee\"
  Assets:Cash       -5.00 USD
  Expenses:Coffee    5.00 USD
";
        // max LHS = 15 (Expenses:Coffee); number_col = 17.
        // max number width = 6 (`-5.00`); number_width = 6.
        // Posting 1: account end at col 13, pad 4 → `-5.00` (width 6,
        //   no left-pad) → currency at col 24.
        // Posting 2: account end at col 17, pad 2 → ` 5.00` (width
        //   5 left-padded by 1) → currency at col 24.
        let expected = "\
2024-01-15 * \"Coffee\"
  Assets:Cash      -5.00 USD
  Expenses:Coffee   5.00 USD
";
        assert_eq!(format_source(src), expected);
    }

    /// Regression for #1290: an amount-less posting (the common elided
    /// balancing leg) must NOT widen the number column, even when its
    /// account is longer than every amount-bearing account. `bean-format`
    /// computes the column only from number-bearing lines, so counting
    /// `Expenses:Food` here would make `rledger format` and `bean-format`
    /// disagree and never converge on round-trip.
    #[test]
    fn transaction_elided_posting_does_not_widen_amount_column() {
        let src = "\
2024-01-15 * \"Coffee\"
  Assets:Cash  -5.00 USD
  Expenses:Food
";
        // Only Assets:Cash (11) bears an amount; Expenses:Food (13) is
        // elided and is ignored for alignment. number_col = 2+11+2 = 15.
        let expected = "\
2024-01-15 * \"Coffee\"
  Assets:Cash  -5.00 USD
  Expenses:Food
";
        assert_eq!(format_source(src), expected);
        // Idempotent: re-formatting the output is a no-op.
        assert_eq!(format_source(expected), expected);
    }

    /// Regression for #1290 using the reporter's exact fixture: a long
    /// elided account (`Expenses:Thingamabobs`) alongside a short
    /// amount-bearing one (`Assets:Money`). Pre-fix the number was
    /// pushed right to clear the long account; `bean-format` keeps it
    /// two spaces after `Assets:Money`. Also confirms the thousands
    /// separator is stripped.
    #[test]
    fn transaction_long_elided_account_matches_bean_format() {
        let src = "\
2024-07-20 * \"Commas should stay\"
  Assets:Money  -1,024 USD
  Expenses:Thingamabobs
";
        let expected = "\
2024-07-20 * \"Commas should stay\"
  Assets:Money  -1024 USD
  Expenses:Thingamabobs
";
        assert_eq!(format_source(src), expected);
        assert_eq!(format_source(expected), expected);
    }

    /// Regression for the currency-only gap (#1307, found in review): a
    /// currency-only posting (`... USD`, no number) renders no number,
    /// so — like an elided posting — it must not widen the alignment
    /// column even when its account is the longest. Only `Assets:Bank`
    /// bears a number here, so the number stays two spaces after it. The
    /// assertion checks the numbered line directly, independent of how
    /// the currency-only line itself renders.
    #[test]
    fn transaction_currency_only_posting_does_not_widen_amount_column() {
        let out = format_source(
            "2024-01-15 * \"x\"\n  Assets:Bank  -5.00 USD\n  Assets:LongCashReserve USD\n",
        );
        assert!(
            out.contains("  Assets:Bank  -5.00 USD"),
            "number column must align to the numbered posting, not the longer \
             currency-only one; got:\n{out}"
        );
    }

    /// A posting with no NUMBER keeps its currency, cost and price.
    ///
    /// `emit_posting` renders the amount only when `amount_number_text`
    /// yields a number, and everything else was inside that branch, so a
    /// posting without one lost its currency, its cost spec and its price
    /// annotation. The tokens were not reformatted, they were deleted, and
    /// what the author wrote could not be recovered from the output (#2142).
    ///
    /// `Assets:Other USD` is VALID beancount that constrains interpolation to
    /// USD, so this was data loss on correct input, not only on malformed
    /// input. The malformed shapes matter too: `format` runs on editor save,
    /// so mistyping a units number and saving deleted the rest of the line.
    #[test]
    fn issue_2142_numberless_posting_keeps_currency_cost_and_price() {
        // Each expectation names the WHOLE posting line, not a fragment.
        // An earlier version asserted only `"USD"` for the currency-only
        // case, which the OTHER posting (`-5.00 USD`) satisfies, so the test
        // would have passed while the currency-only posting still lost its
        // currency. A test written to prove content survives must not be
        // satisfiable by unrelated content.
        for (name, src, must_contain) in [
            (
                "currency only",
                "2024-01-15 * \"x\"\n  Assets:Bank  -5.00 USD\n  Assets:Other USD\n",
                "Assets:Other  USD",
            ),
            (
                "price with no amount node at all",
                "2013-05-18 * \"x\"\n  Assets:MSFT @@ 2000.00 USD\n  Assets:Cash  -2000.00 USD\n",
                "Assets:MSFT  @@ 2000.00 USD",
            ),
            (
                "cost with no amount node at all",
                "2013-05-18 * \"x\"\n  Assets:MSFT {12.00 USD}\n  Assets:Cash  -120.00 USD\n",
                "Assets:MSFT  {12.00 USD}",
            ),
            (
                "total price without units",
                "2013-05-18 * \"x\"\n  Assets:MSFT  MSFT @@ 2000.00 USD\n  Assets:Cash  -2000.00 USD\n",
                "Assets:MSFT  MSFT  @@ 2000.00 USD",
            ),
            (
                "cost without units",
                "2013-05-18 * \"x\"\n  Assets:MSFT  MSFT {12.00 USD}\n  Assets:Cash  -120.00 USD\n",
                "Assets:MSFT  MSFT  {12.00 USD}",
            ),
            (
                "per-unit price without units",
                "2013-05-18 * \"x\"\n  Assets:MSFT  MSFT @ 12.00 USD\n  Assets:Cash  -120.00 USD\n",
                "Assets:MSFT  MSFT  @ 12.00 USD",
            ),
        ] {
            let out = format_source(src);
            assert!(
                out.contains(must_contain),
                "{name}: formatting deleted {must_contain:?}; got:\n{out}"
            );
            // And formatting the result again must not change it further: a
            // fix that re-emitted the tokens in a shape it could not reparse
            // would show up here rather than in the assertion above.
            assert_eq!(
                format_source(&out),
                out,
                "{name}: not idempotent; got:\n{out}"
            );
        }
    }

    #[test]
    fn transaction_payee_and_narration() {
        let src =
            "2024-01-15 * \"Starbucks\" \"Coffee\"\n  Assets:Cash -5.00 USD\n  Expenses:Coffee\n";
        let out = format_source(src);
        assert!(
            out.contains("2024-01-15 * \"Starbucks\" \"Coffee\"\n"),
            "got: {out}"
        );
    }

    #[test]
    fn transaction_pending_flag() {
        let src = "2024-01-15 ! \"Pending\"\n  Assets:Cash -5.00 USD\n  Expenses:Misc\n";
        let out = format_source(src);
        assert!(out.starts_with("2024-01-15 ! \"Pending\"\n"), "got: {out}");
    }

    #[test]
    fn transaction_txn_keyword_normalized_to_star() {
        // The `txn` keyword form is canonical-form equivalent to `*`.
        let src = "2024-01-15 txn \"x\"\n  Assets:Cash -1.00 USD\n  Expenses:Misc\n";
        let out = format_source(src);
        assert!(out.starts_with("2024-01-15 * \"x\"\n"), "got: {out}");
    }

    #[test]
    fn transaction_header_tags_and_links() {
        let src =
            "2024-01-15 * \"x\" #tag1 ^link1 #tag2\n  Assets:Cash -1.00 USD\n  Expenses:Misc\n";
        let out = format_source(src);
        assert!(
            out.starts_with("2024-01-15 * \"x\" #tag1 ^link1 #tag2\n"),
            "got: {out}"
        );
    }

    #[test]
    fn transaction_auto_balance_posting_no_amount() {
        let src = "2024-01-15 * \"x\"\n  Assets:Cash  -5.00 USD\n  Expenses:Misc\n";
        let out = format_source(src);
        // The auto-balance posting has no amount; should just be
        // the indented account name.
        assert!(out.contains("\n  Expenses:Misc\n"), "got: {out}");
    }

    #[test]
    fn transaction_posting_with_cost_spec() {
        let src = "2024-01-15 * \"buy\"\n  Assets:Brokerage  10 HOOL {500.00 USD}\n  Assets:Cash  -5000.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("10 HOOL {500.00 USD}"), "got: {out}");
    }

    #[test]
    fn transaction_posting_with_total_cost_spec() {
        let src = "2024-01-15 * \"buy\"\n  Assets:Brokerage  10 HOOL {{5000.00 USD}}\n  Assets:Cash  -5000.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("10 HOOL {{5000.00 USD}}"), "got: {out}");
    }

    #[test]
    fn transaction_posting_with_per_unit_price() {
        let src = "2024-01-15 * \"buy\"\n  Assets:Brokerage  10 HOOL @ 500.00 USD\n  Assets:Cash  -5000.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("10 HOOL @ 500.00 USD"), "got: {out}");
    }

    #[test]
    fn transaction_posting_with_total_price() {
        let src = "2024-01-15 * \"buy\"\n  Assets:Brokerage  10 HOOL @@ 5000.00 USD\n  Assets:Cash  -5000.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("10 HOOL @@ 5000.00 USD"), "got: {out}");
    }

    #[test]
    fn transaction_posting_with_flag() {
        let src = "2024-01-15 * \"x\"\n  ! Assets:Cash  -5.00 USD\n  Expenses:Misc  5.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("\n  ! Assets:Cash"), "got: {out}");
    }

    #[test]
    fn transaction_negative_amount() {
        let src = "2024-01-15 * \"x\"\n  Assets:Cash -5.00 USD\n  Expenses:Misc 5.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("-5.00 USD"), "got: {out}");
        assert!(out.contains(" 5.00 USD"), "got: {out}");
    }

    #[test]
    fn transaction_strips_thousands_separators_in_postings() {
        let src = "2024-01-15 * \"x\"\n  Assets:Cash -1,000.00 USD\n  Expenses:Misc 1,000.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("-1000.00 USD"), "got: {out}");
        assert!(!out.contains("1,000"), "got: {out}");
    }

    #[test]
    fn transaction_arithmetic_amount() {
        let src =
            "2024-01-15 * \"x\"\n  Assets:Cash  -(1.00 + 2.00) USD\n  Expenses:Misc 3.00 USD\n";
        let out = format_source(src);
        // The arithmetic expression should render with single
        // spaces around binary ops and tight parens.
        assert!(
            out.contains("(1.00 + 2.00) USD") || out.contains("-(1.00 + 2.00) USD"),
            "got: {out}"
        );
    }

    #[test]
    fn transaction_idempotent() {
        let src = "\
2024-01-15 * \"Coffee\"
  Assets:Cash       -5.00 USD
  Expenses:Coffee    5.00 USD
";
        let once = format_source(src);
        let twice = format_source(&once);
        assert_eq!(once, twice);
    }

    #[test]
    fn transaction_file_wide_alignment_across_transactions() {
        let src = "\
2024-01-15 * \"x\"
  Assets:Cash -5.00 USD
  Expenses:Misc 5.00 USD

2024-01-16 * \"y\"
  Liabilities:CreditCard:Visa  -100.00 USD
  Expenses:Big  100.00 USD
";
        let out = format_source(src);
        // Cross-posting invariant: the currency column (USD here)
        // lands at the same column on every posting line, even when
        // individual numbers differ in width or sign. The number
        // field is right-justified so the currency column is uniform.
        let usd_cols: Vec<usize> = out
            .lines()
            .filter(|l| l.starts_with("  ") && l.contains(" USD"))
            .filter_map(|l| l.find("USD"))
            .collect();
        assert!(
            usd_cols.len() >= 4,
            "expected ≥4 posting lines, got {usd_cols:?} in {out}"
        );
        let first = usd_cols[0];
        assert!(
            usd_cols.iter().all(|&c| c == first),
            "expected USD column uniform at {first}, got {usd_cols:?} in:\n{out}"
        );
    }

    #[test]
    fn transaction_posting_metadata_indented_four() {
        let src =
            "2024-01-15 * \"x\"\n  Assets:Cash -5.00 USD\n    foo: \"bar\"\n  Expenses:Misc\n";
        let out = format_source(src);
        assert!(out.contains("\n    foo: \"bar\"\n"), "got: {out}");
    }

    // ---- Code-review regression tests -----------------------------
    //
    // Each test pins a bug surfaced by the high-effort code review of
    // PR #1284 and verified at runtime against the unfixed formatter.

    #[test]
    fn cost_spec_per_unit_plus_total_opener_preserved() {
        // Bug: format_cost_spec only branched on is_total() and emitted
        // `{` for the `{#` opener too, dropping the `#` marker and
        // changing semantics from per-unit-plus-total to plain
        // per-unit cost.
        let src = "2024-01-01 * \"buy\"\n  Assets:Brokerage 10 HOOL {# 500.00 USD}\n  Assets:Cash -5000.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("{# 500.00 USD}"),
            "expected `{{#` opener preserved; got:\n{out}"
        );
        assert!(!out.contains("{500.00 USD}"), "got:\n{out}");
    }

    #[test]
    fn cost_spec_comma_stays_tight_to_prev_token() {
        // Bug: format_cost_spec's catch-all arm inserted a space
        // before every non-trivia token including COMMA, producing
        // `{500.00 USD , 2024-01-15}` instead of the canonical
        // `{500.00 USD, 2024-01-15}`.
        let src = "2024-01-01 * \"buy\"\n  Assets:Brokerage 10 HOOL {500.00 USD, 2024-01-15}\n  Assets:Cash -5000.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("{500.00 USD, 2024-01-15}"),
            "comma must stay tight to USD; got:\n{out}"
        );
        assert!(
            !out.contains("USD ,"),
            "no space allowed before comma; got:\n{out}"
        );
    }

    #[test]
    fn custom_directive_preserves_date_value_arguments() {
        // Bug: emit_custom's post-seen_type match skipped every DATE
        // token, silently dropping legitimate date-typed value
        // arguments. The leading directive date is already skipped
        // via the seen_type=false phase.
        let src = "2024-01-01 custom \"budget\" \"name\" 2024-06-15 100.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("2024-06-15"),
            "value-position DATE must survive; got: {out}"
        );
    }

    #[test]
    fn file_level_adjacent_comments_stay_tight() {
        // Bug: format_node's top-level walk inserted a blank `\n`
        // separator before every emitted item including comments,
        // breaking section-header blocks like `; ====\n; HEADER\n; ====`
        // by injecting blanks between every adjacent comment line.
        let src = "; ====\n; HEADER\n; ====\n2024-01-01 open Assets:A\n";
        let expected = "; ====\n; HEADER\n; ====\n2024-01-01 open Assets:A\n";
        assert_eq!(format_source(src), expected);
    }

    #[test]
    fn metadata_internal_whitespace_normalized() {
        // Bug: emit_meta_entries_of passed META_ENTRY source text
        // through verbatim, so `foo: "bar"` and `foo:    "bar"` —
        // identical typed ASTs — produced different formatter
        // output, violating the gofmt-style invariant the rustdoc
        // declares.
        let a = "2024-01-01 open Assets:Bank\n  starting: \"foo\"\n";
        let b = "2024-01-01 open Assets:Bank\n  starting:    \"foo\"\n";
        assert_eq!(format_source(a), format_source(b));
    }

    #[test]
    fn metadata_number_thousands_separator_stripped() {
        // Same invariant: numbers inside metadata values share the
        // canonical thousands-separator policy with posting numbers
        // (otherwise the same file would emit inconsistent numeric
        // forms in postings vs. metadata).
        let src = "2024-01-01 open Assets:Bank\n  starting_balance: 1,000.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("1000.00 USD"),
            "thousands-sep should strip in metadata too; got: {out}"
        );
        assert!(!out.contains("1,000"), "got: {out}");
    }

    #[test]
    fn bare_cr_line_endings_normalized_to_lf_before_parse() {
        // Bug: the lexer doesn't treat bare CR as a line terminator,
        // so a classic-Mac-authored `directive\r…\rdirective\r`
        // parsed as one broken directive and the rest were silently
        // dropped. format_source normalizes line endings BEFORE
        // parsing so bare CR (and CRLF) are treated as LF.
        let src = "2024-01-01 open Assets:A\r2024-01-02 open Assets:B\r";
        let out = format_source(src);
        assert!(
            out.contains("2024-01-01 open Assets:A"),
            "first directive lost: {out:?}"
        );
        assert!(
            out.contains("2024-01-02 open Assets:B"),
            "second directive lost on bare-CR input: {out:?}"
        );
    }

    #[test]
    fn crlf_input_canonicalizes_to_lf() {
        // CRLF and bare CR both fold to LF on the way through the
        // canonical pass (the canonical form is LF-only).
        let src = "2024-01-01 open Assets:A\r\n2024-01-02 open Assets:B\r\n";
        let out = format_source(src);
        assert!(
            !out.contains('\r'),
            "canonical output must be LF-only: {out:?}"
        );
        assert!(out.contains("2024-01-01 open Assets:A\n"), "got: {out:?}");
        assert!(out.contains("2024-01-02 open Assets:B\n"), "got: {out:?}");
    }

    #[test]
    fn metadata_value_with_unary_minus_stays_tight() {
        // Bug: emit_meta_entry's tokenized walk inserted a space
        // after a unary `+`/`-`, breaking `key: -5.00 USD` →
        // `key: - 5.00 USD`. Routed through write_canonical_token_sequence
        // so unary detection matches the balance/price/posting paths.
        let src = "2024-01-01 open Assets:Bank\n  threshold: -5.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("threshold: -5.00 USD"),
            "unary minus must stay tight in metadata; got: {out}"
        );
        assert!(
            !out.contains("- 5.00"),
            "no space after unary minus; got: {out}"
        );
    }

    #[test]
    fn metadata_value_with_unary_plus_stays_tight() {
        let src = "2024-01-01 open Assets:Bank\n  min: +1.00 USD\n";
        let out = format_source(src);
        assert!(out.contains("min: +1.00 USD"), "got: {out}");
        assert!(!out.contains("+ 1.00"), "got: {out}");
    }

    #[test]
    fn cost_spec_negative_cost_stays_tight() {
        // Bug: format_cost_spec catch-all had no unary-operator
        // handling. `{-500 USD}` formatted to `{- 500 USD}`. Now
        // routes through write_canonical_token_sequence.
        let src = "2024-01-01 * \"x\"\n  Assets:Brokerage 10 HOOL {-500 USD}\n  Assets:Cash -5000.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("{-500 USD}"),
            "negative cost spec must stay tight; got:\n{out}"
        );
        assert!(!out.contains("{- "), "got:\n{out}");
    }

    #[test]
    fn cost_spec_arithmetic_with_unary_stays_tight() {
        // `{500 * -2 USD}` formerly emitted `{500 * - 2 USD}` because
        // the cost-spec catch-all didn't understand unary +/-.
        let src = "2024-01-01 * \"x\"\n  Assets:Brokerage 10 HOOL {500 * -2 USD}\n  Assets:Cash -1000.00 USD\n";
        let out = format_source(src);
        assert!(
            out.contains("{500 * -2 USD}"),
            "cost-spec arithmetic unary must stay tight; got:\n{out}"
        );
    }

    // ---- Property tests -------------------------------------------
    //
    // Two invariants the rustdoc's gofmt-style promise depends on,
    // pinned over a hand-curated input matrix:
    //
    // - **Idempotence:** `format_source(format_source(x)) == format_source(x)`.
    // - **Round-trip stability for canonicalize_directives:** the
    //   synthesize-then-canonicalize shim produces text that, when
    //   parsed back, yields the same Directive count and zero parse
    //   errors.
    //
    // The matrix covers every directive kind plus the high-risk
    // edge cases the prior reviews surfaced (unary +/- in metadata,
    // cost-spec arithmetic, CRLF, bare CR, multi-line strings,
    // comments containing quotes, non-Latin accounts). When the
    // upstream compatibility corpus is fetched into
    // `tests/compatibility/files/` the per-file sweep at the bottom
    // also runs; otherwise the file-based test is skipped.

    const IDEMPOTENCE_MATRIX: &[(&str, &str)] = &[
        ("empty", ""),
        ("only_comment", "; header comment\n"),
        ("only_directive", "2024-01-01 open Assets:Cash\n"),
        (
            "two_open_directives",
            "2024-01-01 open Assets:A\n2024-01-02 open Assets:B\n",
        ),
        (
            "transaction_with_cost_and_price",
            "2024-01-15 * \"buy\"\n  Assets:Brokerage 10 HOOL {500.00 USD} @ 510.00 USD\n  Assets:Cash -5000.00 USD\n",
        ),
        (
            "transaction_with_per_unit_plus_total_cost",
            "2024-01-15 * \"x\"\n  Assets:Brokerage 10 HOOL {# 500.00 USD}\n  Assets:Cash -5000.00 USD\n",
        ),
        (
            "transaction_with_arithmetic_amount",
            "2024-01-15 * \"x\"\n  Assets:Cash  -(1.00 + 2.00) USD\n  Expenses:Misc 3.00 USD\n",
        ),
        (
            "balance_with_arithmetic_and_tolerance",
            "2024-01-15 balance Assets:Cash 0.25 + 0.75 USD ~ 0.01 USD\n",
        ),
        // Regression for Copilot #2: a previous emit_amount_expression
        // skipped tokens until the first NUMBER, which dropped a
        // leading unary `-` and silently flipped the sign — a
        // balance assertion that asserted a debit would assert a
        // credit after a round-trip. These fixtures pin the
        // sign / paren preservation explicitly.
        (
            "balance_leading_unary_minus",
            "2024-01-15 balance Assets:A -1.00 USD\n",
        ),
        (
            "balance_leading_parenthesized_expression",
            "2024-01-15 balance Assets:A (1 + 2) USD\n",
        ),
        (
            "price_leading_unary_minus",
            "2024-01-15 price USD -1.00 EUR\n",
        ),
        (
            "price_with_thousands_separator",
            "2024-01-15 price USD 1,234.56 EUR\n",
        ),
        (
            "metadata_unary_minus",
            "2024-01-01 open Assets:Bank\n  threshold: -5.00 USD\n",
        ),
        (
            "metadata_arithmetic",
            "2024-01-01 open Assets:Bank\n  total: 1000 + 500 USD\n",
        ),
        (
            "cost_spec_with_comma_and_date",
            "2024-01-15 * \"x\"\n  Assets:Brokerage 10 HOOL {500.00 USD, 2024-01-15}\n  Assets:Cash -5000.00 USD\n",
        ),
        (
            "cost_spec_with_negative",
            "2024-01-15 * \"x\"\n  Assets:Brokerage 10 HOOL {-500 USD}\n  Assets:Cash 5000.00 USD\n",
        ),
        (
            "transaction_with_tags_and_links",
            "2024-01-15 * \"x\" #tag1 ^link1 #tag2\n  Assets:Cash -1.00 USD\n  Expenses:Misc 1.00 USD\n",
        ),
        (
            "custom_with_date_value",
            "2024-01-01 custom \"budget\" \"name\" 2024-06-15 100.00 USD\n",
        ),
        (
            "non_latin_account_name",
            "2024-01-15 * \"x\"\n  Активы:Банк -5.00 USD\n  Expenses:Misc 5.00 USD\n",
        ),
        (
            "section_header_comments",
            "; ====\n; HEADER\n; ====\n2024-01-01 open Assets:A\n",
        ),
        (
            "multiline_note_string",
            "2024-01-15 note Assets:Bank \"line 1\nline 2\"\n",
        ),
        (
            "comment_containing_quote",
            "; comment with \"a quote\n2024-01-01 open Assets:A\n",
        ),
        (
            "crlf_input",
            "2024-01-01 open Assets:A\r\n2024-01-02 open Assets:B\r\n",
        ),
        (
            "bare_cr_input",
            "2024-01-01 open Assets:A\r2024-01-02 open Assets:B\r",
        ),
        (
            "file_with_trailing_newlines",
            "2024-01-01 open Assets:A\n\n\n",
        ),
        ("file_without_trailing_newline", "2024-01-01 open Assets:A"),
        // Regression for Copilot #1: collect_trailing_comment
        // previously returned None for a directive with no
        // header-terminating NEWLINE token, which silently dropped
        // a same-line trailing comment at EOF when the file lacked
        // a trailing newline. The canonical formatter restores the
        // trailing newline, but the dropped comment was already
        // gone.
        (
            "trailing_comment_no_final_newline",
            "2024-01-15 open Assets:A ; trailing",
        ),
        (
            "posting_with_trailing_comment",
            "2024-01-15 * \"x\"\n  Assets:Cash -5.00 USD ; pocket\n  Expenses:Misc 5.00 USD\n",
        ),
        (
            "balance_assertion_with_meta",
            "2024-01-15 balance Assets:Cash 100.00 USD\n  source: \"bank\"\n",
        ),
        (
            "options_and_includes",
            "option \"title\" \"My Ledger\"\ninclude \"sub.beancount\"\nplugin \"my.plugin\" \"cfg\"\n",
        ),
        // ---- per-variant coverage ---------------------------------
        ("close_directive", "2024-12-31 close Assets:Cash\n"),
        ("commodity_directive", "2024-01-01 commodity HOOL\n"),
        // Tagged and linked on purpose. Untagged, this fixture could not see
        // a note losing its tags -- which is exactly what happened: the
        // emitters dropped them and every harness in this file passed
        // (#2184).
        (
            "note_directive",
            "2024-01-15 note Assets:Cash \"a note\" #n1 ^l1\n",
        ),
        ("event_directive", "2024-01-15 event \"location\" \"NYC\"\n"),
        (
            "query_directive",
            "2024-01-15 query \"q1\" \"SELECT account\"\n",
        ),
        ("pad_directive", "2024-01-15 pad Assets:A Equity:Opening\n"),
        (
            "document_directive",
            "2024-06-01 document Assets:Bank \"stmt.pdf\" #q1 ^d1\n",
        ),
        // Note: `#!` and `#+` anywhere on a line, not just at
        // line start, open the lexer's SHEBANG / EMACS_DIRECTIVE
        // tokens. The fixture places `#+` mid-line and tails it
        // with an unbalanced `"`: an incorrect state machine that
        // gated the opener on `at_line_start` would stay in Code
        // when it hit the `#+`, then flip to InString on the next
        // `"` and trap there for the remainder of the file. The
        // lexer-agreement property test catches that divergence,
        // and the round-trip body runs too because the parser
        // treats the mid-line EMACS_DIRECTIVE as same-line
        // trailing trivia under the directive-terminator rule.
        (
            "emacs_directive_mid_line_with_quote",
            "2024-01-15 open Assets:A #+stray \"q\n",
        ),
        ("pushtag_directive", "pushtag #active\n"),
        ("poptag_directive", "poptag #active\n"),
        ("pushmeta_directive", "pushmeta location: \"NYC\"\n"),
        ("popmeta_directive", "popmeta location:\n"),
    ];

    /// Number of fixtures in [`IDEMPOTENCE_MATRIX`] that legitimately
    /// produce zero typed directives — comment-only / empty /
    /// pragma-only inputs. The round-trip property test skips these
    /// (they have nothing to emit), but every OTHER fixture MUST
    /// exercise the body. Bumping this constant when adding such a
    /// fixture is the only manual maintenance the coverage floor
    /// needs; otherwise the floor (`IDEMPOTENCE_MATRIX.len() -
    /// ROUNDTRIP_KNOWN_ZERO_DIRECTIVE_FIXTURES`) tracks the matrix
    /// automatically.
    ///
    /// Today's zero-directive fixtures (skipped by the round-trip
    /// body), verified by an exhaustive probe against the live
    /// parser:
    ///
    /// - `empty`, `only_comment` — no directives at all.
    /// - `bare_cr_input` — the parser does not recognize bare CR
    ///   (without a following LF) as a directive terminator, so
    ///   the file's two would-be directives never surface as
    ///   structured tokens. The fixture's purpose is the
    ///   line-ending state-machine pass, not the round-trip body.
    /// - `pushtag_directive`, `poptag_directive`,
    ///   `pushmeta_directive`, `popmeta_directive` — pragma
    ///   directives don't surface as `Directive` variants on the
    ///   typed-AST side (the parser also rejects them today, so
    ///   they produce parse errors and the skip-on-errors guard
    ///   triggers).
    /// - `options_and_includes` — option / include / plugin lines
    ///   live on separate `ParseResult` collections, not on
    ///   `.directives`.
    ///
    /// Note: `comment_containing_quote` and
    /// `emacs_directive_mid_line_with_quote` BOTH exercise the
    /// body — each is paired with a parseable directive on the
    /// same line or an adjacent line, and the trivia token
    /// (comment / `EMACS_DIRECTIVE`) attaches as same-line or
    /// inter-directive trivia under the directive-terminator
    /// rule. Their purpose is the state-machine / lexer agreement
    /// property on a comment with an unbalanced `"`, not the
    /// zero-directive case.
    const ROUNDTRIP_KNOWN_ZERO_DIRECTIVE_FIXTURES: usize = 8;

    #[test]
    fn lf_to_crlf_outside_strings_preserves_string_interior() {
        // Bug: a flat in_string-only state machine would re-inject
        // CRLF inside multi-line strings, mutating the user's bytes.
        let s = "2024-01-15 note Assets:Bank \"line 1\nline 2\"\n";
        let out = lf_to_crlf_outside_strings(s);
        assert!(out.contains("line 1\nline 2"), "got: {out:?}");
        assert!(out.ends_with("\r\n"), "got: {out:?}");
    }

    #[test]
    fn lf_to_crlf_outside_strings_handles_comment_with_quote() {
        // Bug: an unbalanced `"` inside a `;` comment formerly flipped
        // in_string=true for the rest of the file, leaving every
        // subsequent newline as LF.
        let s = "; comment with \"a quote\n2024-01-01 open Assets:A\n";
        let out = lf_to_crlf_outside_strings(s);
        assert_eq!(
            out,
            "; comment with \"a quote\r\n2024-01-01 open Assets:A\r\n",
        );
    }

    #[test]
    fn lf_to_crlf_outside_strings_handles_percent_comment_with_quote() {
        let s = "% percent \"quote\n2024-01-01 open Assets:A\n";
        let out = lf_to_crlf_outside_strings(s);
        assert_eq!(out, "% percent \"quote\r\n2024-01-01 open Assets:A\r\n");
    }

    #[test]
    fn crlf_to_lf_preserves_crlf_inside_strings() {
        // Bug fix mirror: a Windows-authored multi-line string had
        // its CRLF folded to LF by the pre-parse normalizer too,
        // which silently mutated the user's bytes.
        let s = "2024-01-15 note Assets:Bank \"line1\r\nline2\"\r\n";
        let normalized = crlf_to_lf_outside_strings(s);
        // Outside the string, the trailing CRLF folds to LF; inside
        // the string, CRLF stays CRLF (user's bytes preserved).
        assert!(
            normalized.contains("\"line1\r\nline2\""),
            "got: {:?}",
            &*normalized
        );
        assert!(normalized.ends_with('\n') && !normalized.ends_with("\r\n"));
    }

    /// Formatting must not change what the file MEANS.
    ///
    /// The idempotence matrix next door checks that formatting twice equals
    /// formatting once. That property cannot see data loss: dropping a note's
    /// tags is perfectly idempotent, and it passed for as long as the bug
    /// lived (#2184). Stability is not fidelity.
    ///
    /// So: parse, format, parse again, and compare the directive models. Spans
    /// move and are not compared; everything the model holds must survive.
    ///
    /// Known blind spot, worth stating so a future reader does not over-trust
    /// this: a model comparison cannot see data the model never held. When the
    /// PARSER drops something, both sides are equally lossy and compare equal.
    /// `balance X 1.00 USD ~ 0.01 EUR` formats to `1.00 ~ 0.01 USD` -- the EUR
    /// is erased from the file -- and this test passes, because the parser
    /// discarded it before either comparison (#2193). That is why the
    /// note-tag test next door also asserts on output TEXT.
    ///
    /// This could not land with #2184: it failed on
    /// `balance_with_arithmetic_and_tolerance`, where the formatter emitted a
    /// balance that no longer parsed. That was a parser inconsistency (#2191),
    /// fixed in the commit this test arrives with -- the harness found it, and
    /// waited for it rather than hiding it behind an allow-list.
    #[test]
    fn formatting_preserves_the_parsed_directives() {
        let mut checked = 0;
        for (name, src) in IDEMPOTENCE_MATRIX {
            let before = crate::parse(src);
            if !before.errors.is_empty() {
                continue; // error fixtures pin diagnostics, not round-trips
            }
            let formatted = format_source(src);
            let after = crate::parse(&formatted);

            assert!(
                after.errors.is_empty(),
                "{name}: formatter produced output that no longer parses: {:?}\n{formatted}",
                after.errors,
            );
            let b: Vec<_> = before.directives.iter().map(|d| &d.value).collect();
            let a: Vec<_> = after.directives.iter().map(|d| &d.value).collect();
            assert_eq!(
                b, a,
                "{name}: formatting changed the directives it parsed from\n\
                 --- source ---\n{src}\n--- formatted ---\n{formatted}",
            );
            checked += 1;
        }
        // A matrix that stopped yielding clean fixtures would make every
        // assertion above vacuous.
        // 37 of 42 round-trip today; the rest are error fixtures. The floor
        // is just under that rather than a round number a third of the way
        // down: at `> 20` a change that made ten fixtures stop parsing would
        // slip through, and skipping is exactly how this test goes quiet.
        assert!(
            checked >= 35,
            "only {checked} fixtures round-tripped; this test is not covering \
             what it claims to"
        );
    }

    /// Every fixture must survive the TYPED emitter, not just `format_source`.
    ///
    /// `canonicalize_directives` is the path `rledger add`, `rledger extract`
    /// and the FFI `format.entry` endpoints take, and it renders through
    /// `rustledger_core::format::format_directives` -- a different emitter
    /// with its own set of fields it might forget. It forgot a note's and a
    /// document's tags and links, and the fidelity test next door could not
    /// see it because that one exercises `format_source`.
    ///
    /// Same shape as that test, aimed at the other emitter: parse, render,
    /// parse, compare models.
    #[test]
    fn canonicalizing_every_fixture_preserves_the_model() {
        let cfg = rustledger_core::format::FormatConfig::default();
        let mut checked = 0;
        for (name, src) in IDEMPOTENCE_MATRIX {
            let before = crate::parse(src);
            if !before.errors.is_empty() || before.directives.is_empty() {
                continue;
            }
            let ds: Vec<_> = before.directives.iter().map(|d| d.value.clone()).collect();
            let out = match canonicalize_directives(ds.iter(), &cfg) {
                Ok(out) => out,
                // The shim reports a re-parse failure rather than emitting a
                // recoverable subset; that is its contract, not a silent loss.
                Err(e) => panic!("{name}: canonicalize failed: {e:?}"),
            };
            let after = crate::parse(&out);
            assert!(
                after.errors.is_empty(),
                "{name}: typed emitter produced unparsable text: {:?}\n{out}",
                after.errors,
            );
            let a: Vec<_> = after.directives.iter().map(|d| &d.value).collect();
            let b: Vec<_> = ds.iter().collect();
            assert_eq!(
                b, a,
                "{name}: the typed emitter changed the directives\n\
                 --- source ---\n{src}\n--- rendered ---\n{out}",
            );
            checked += 1;
        }
        // Fixtures that parse to no directive at all (comments, options) are
        // skipped above; the floor keeps that from quietly becoming all of them.
        assert!(
            checked >= 30,
            "only {checked} fixtures reached the typed emitter"
        );
    }

    /// The typed-directive emitter must not delete tags either.
    ///
    /// There are TWO live formatters. `format_source` is what `rledger
    /// format` runs; `canonicalize_directives` is what `rledger add`,
    /// `rledger extract` and the FFI `format.entry` endpoints run, and it
    /// synthesizes its intermediate text through
    /// `rustledger_core::format::format_directives`.
    ///
    /// Fixing only the first left the second deleting the same data -- and
    /// deleting MORE of it, since the CST emitter at least kept a document's
    /// tags while the typed one dropped those too. A round-trip through this
    /// shim is what a caller building directives in memory actually gets.
    #[test]
    fn canonicalizing_typed_directives_keeps_tags_and_links() {
        let src = "2024-01-05 note Assets:A \"n\" #ntag ^nlink\n\
                   2024-01-06 document Assets:A \"/x.pdf\" #dtag ^dlink\n\
                   2024-01-07 * \"t\" #ttag ^tlink\n\
                  \x20 Assets:A  1 USD\n\
                  \x20 Equity:O\n";
        let parsed = crate::parse(src);
        assert!(parsed.errors.is_empty(), "{:?}", parsed.errors);
        let directives: Vec<_> = parsed.directives.iter().map(|d| d.value.clone()).collect();

        let out = canonicalize_directives(
            directives.iter(),
            &rustledger_core::format::FormatConfig::default(),
        )
        .expect("fixture must canonicalize");

        for marker in ["#ntag", "^nlink", "#dtag", "^dlink", "#ttag", "^tlink"] {
            assert!(
                out.contains(marker),
                "typed emitter dropped {marker}\n{out}"
            );
        }

        // And the text it produced still means the same thing.
        let after = crate::parse(&out);
        assert!(after.errors.is_empty(), "{:?}\n{out}", after.errors);
        let a: Vec<_> = after.directives.iter().map(|d| &d.value).collect();
        let b: Vec<_> = directives.iter().collect();
        assert_eq!(b, a, "canonicalizing changed the directives\n{out}");
    }

    /// Formatting a note must not change what it means.
    ///
    /// The idempotence matrix next door cannot catch this class: dropping a
    /// note's tags is perfectly idempotent, and it passed for as long as
    /// #2184 lived. Stability is not fidelity, so this parses, formats,
    /// parses again, and compares the models.
    ///
    /// Narrower than the matrix tests above on purpose: those compare models,
    /// and a model comparison is satisfied by two equally-empty sides. This
    /// one names the tags it expects in the output TEXT.
    #[test]
    fn formatting_preserves_a_notes_tags_and_links() {
        // A note past the first, after a blank line and after a comment:
        // both attach trivia inside the directive node, which is what broke
        // the equivalent walk in the converter (#2189).
        let src = "2024-01-01 open Assets:A USD\n\
                   \n\
                   2024-01-05 note Assets:A \"first\" #n1 ^l1\n\
                   \n\
                   2024-01-06 note Assets:A \"after blank\" #n2 ^l2\n\
                   ; a comment line\n\
                   2024-01-07 note Assets:A \"after comment\" #n3\n\
                   2024-01-08 note Assets:A \"with meta\" #n4\n\
                  \x20 key: \"v\"\n\
                   \n\
                   2024-01-09 document Assets:A \"/x.pdf\" #d1 ^l3\n";

        let before = crate::parse(src);
        assert!(before.errors.is_empty(), "{:?}", before.errors);
        let formatted = format_source(src);
        let after = crate::parse(&formatted);
        assert!(
            after.errors.is_empty(),
            "formatted output no longer parses: {:?}\n{formatted}",
            after.errors,
        );

        let b: Vec<_> = before.directives.iter().map(|d| &d.value).collect();
        let a: Vec<_> = after.directives.iter().map(|d| &d.value).collect();
        assert_eq!(
            b, a,
            "formatting changed the directives it parsed from\n\
             --- source ---\n{src}\n--- formatted ---\n{formatted}",
        );

        // The model comparison above is NOT sufficient on its own, and this
        // is not a hypothetical: run it against a tree whose CONVERTER also
        // drops these tags (#2189) and both sides come back tagless and
        // equal, while every tag in the file has been deleted. Assert on the
        // formatter's own output text, which is the thing under change here
        // and cannot be satisfied by a matching pair of empty models.
        for tag in ["#n1", "#n2", "#n3", "#n4", "#d1"] {
            assert!(
                formatted.contains(tag),
                "formatter dropped {tag}\n{formatted}"
            );
        }
        for link in ["^l1", "^l2", "^l3"] {
            assert!(
                formatted.contains(link),
                "formatter dropped {link}\n{formatted}"
            );
            // Every one of these is already canonical, so formatting must be a
            // no-op on the text too. Without this the value assertion above
            // would accept any spelling that happens to evaluate the same.
            assert_eq!(formatted, src, "already-canonical input was rewritten");
        }

        // A unary sign binds to its number; a binary one keeps its spaces.
        // Rendering `~ - 0.01` evaluates the same but makes one directive
        // disagree with itself, since the amount renders `-1.00` tight.
        assert_eq!(
            format_source("2024-01-15 balance Assets:C -1.00 ~ -0.01 USD\n"),
            "2024-01-15 balance Assets:C -1.00 ~ -0.01 USD\n",
        );
        assert_eq!(
            format_source("2024-01-15 balance Assets:C 1.00 ~ (-0.005 + 0.015) USD\n"),
            "2024-01-15 balance Assets:C 1.00 ~ (-0.005 + 0.015) USD\n",
        );
    }

    #[test]
    fn idempotence_matrix() {
        // The gofmt invariant in the file rustdoc: f(f(x)) == f(x)
        // on every accepted input. Each fixture below covers one
        // axis of the canonical-form spec; together they exercise
        // every directive kind and every spacing rule shared via
        // write_canonical_token_sequence.
        for (name, src) in IDEMPOTENCE_MATRIX {
            let once = format_source(src);
            let twice = format_source(&once);
            assert_eq!(
                once, twice,
                "idempotence broken on fixture `{name}`\n--- once ---\n{once}\n--- twice ---\n{twice}",
            );
        }
    }

    /// The number-display context (#1766) threads through the
    /// two-pass canonicalize shim: precision pads. Thousands
    /// separators are deliberately absent — canonical ledger text has
    /// none (this canonicalizer strips them by definition, and
    /// `render_number` agrees so direct emitters match the shim).
    #[test]
    fn canonicalize_directives_honors_number_display_context() {
        use rustledger_core::format::FormatConfig;

        let source = "2024-01-15 balance Assets:Bank 1234.5 USD\n";
        let parsed = crate::parse(source);
        assert!(parsed.errors.is_empty(), "{:?}", parsed.errors);

        let mut ctx = rustledger_core::DisplayContext::new();
        ctx.set_fixed_precision("USD", 2);
        ctx.set_render_commas(true);
        let config = FormatConfig {
            number_display: Some(ctx),
            ..FormatConfig::default()
        };
        let out = canonicalize_directives(parsed.directives.iter().map(|d| &d.value), &config)
            .expect("comma-grouped canonical text must survive the reparse");
        // REVERSED from "precision pads, no separators". The shim's second
        // pass now carries the grouping rule, so a context that asks for
        // separators gets them in ledger text — matching beancount, whose
        // `render_commas` is documented to affect its PRINT command. The
        // machine boundary is the parser, and the grammar admits grouped
        // numerals; csv/json (whose consumers have no grammar) are unaffected.
        assert!(
            out.contains("1,234.50 USD"),
            "precision AND grouping both flow through the shim: {out}"
        );

        // And the default config stays byte-faithful to the value's scale.
        let out = canonicalize_directives(
            parsed.directives.iter().map(|d| &d.value),
            &FormatConfig::default(),
        )
        .expect("canonicalizes");
        assert!(out.contains("1234.5 USD"), "own scale preserved: {out}");
    }

    /// Padded COST and PRICE-annotation numbers survive the shim's
    /// pass 2 (the CST re-canonicalization preserves trailing zeros),
    /// and the pass-1 emitter and the shim agree on every rendered
    /// number — the cross-surface drift guard the `render_number` doc
    /// claims (deep review of #1807).
    #[test]
    fn canonicalize_pads_costs_and_prices_and_agrees_with_pass_one() {
        use rustledger_core::format::FormatConfig;

        let source = "2024-01-10 * \"buy\"\n  Assets:Broker  2 AAPL {150 USD} @ 155.5 USD\n  Assets:Cash  -310.00 USD\n";
        let parsed = crate::parse(source);
        assert!(parsed.errors.is_empty(), "{:?}", parsed.errors);

        let mut ctx = rustledger_core::DisplayContext::new();
        ctx.set_fixed_precision("USD", 2);
        let config = FormatConfig {
            number_display: Some(ctx),
            ..FormatConfig::default()
        };

        let pass_one = rustledger_core::format::format_directives(
            parsed.directives.iter().map(|d| &d.value),
            &config,
        );
        let canonical =
            canonicalize_directives(parsed.directives.iter().map(|d| &d.value), &config)
                .expect("padded cost/price text must survive the reparse");

        for padded in ["{150.00 USD}", "@ 155.50 USD"] {
            assert!(
                pass_one.contains(padded),
                "pass-1 emitter pads {padded}: {pass_one}"
            );
            assert!(
                canonical.contains(padded),
                "the shim preserves the padded {padded}: {canonical}"
            );
        }
    }

    #[test]
    fn canonicalize_directives_roundtrips_every_synthesized_directive() {
        // For each canonical-form fixture: parse → take the typed
        // directives → run them through canonicalize_directives →
        // re-parse the canonical text → assert the parser reports
        // zero errors and the directive count is preserved.
        //
        // This is the proper end-to-end test of the two-pass shim
        // the FFI format.entry and rledger add/extract commands all
        // depend on. Without it, a future Directive variant added
        // to rustledger-core without matching coverage in
        // cst::format would silently round-trip to truncated text.
        //
        // Counter + assertion guards against silent-skip: if the
        // guard at the top of the loop ever filters too many
        // fixtures (e.g. a parser regression that drops directives
        // from previously-clean fixtures), the test fails instead
        // of silently passing with zero coverage.
        use rustledger_core::format::FormatConfig;
        let cfg = FormatConfig::default();
        let mut exercised = 0usize;
        for (name, src) in IDEMPOTENCE_MATRIX {
            let parsed = crate::parse(src);
            if parsed.errors.is_empty() && !parsed.directives.is_empty() {
                let dirs: Vec<&rustledger_core::Directive> =
                    parsed.directives.iter().map(|s| &s.value).collect();
                let formatted = super::canonicalize_directives(dirs.iter().copied(), &cfg)
                    .unwrap_or_else(|e| {
                        panic!("canonicalize_directives error on fixture `{name}`: {e}")
                    });
                let reparsed = crate::parse(&formatted);
                assert!(
                    reparsed.errors.is_empty(),
                    "round-trip parse errors on fixture `{name}`:\n--- formatted ---\n{formatted}\n--- errors ---\n{:?}",
                    reparsed.errors,
                );
                assert_eq!(
                    parsed.directives.len(),
                    reparsed.directives.len(),
                    "directive count drifted on fixture `{name}`\n--- formatted ---\n{formatted}",
                );
                exercised += 1;
            }
        }
        let expected = IDEMPOTENCE_MATRIX
            .len()
            .saturating_sub(ROUNDTRIP_KNOWN_ZERO_DIRECTIVE_FIXTURES);
        assert!(
            exercised >= expected,
            "only {exercised} fixtures exercised the round-trip body, \
             expected at least {expected} (= IDEMPOTENCE_MATRIX.len() - \
             {ROUNDTRIP_KNOWN_ZERO_DIRECTIVE_FIXTURES}). A parser \
             regression or a broken fixture is silently dropping coverage."
        );
    }

    /// `SHEBANG` / `EMACS_DIRECTIVE` lines (`#!…` / `#+…` at line
    /// start) also count as comments for the LSP-CRLF state
    /// machine. A stray quote inside such a line used to flip
    /// `in_string=true` for the rest of the file just like the
    /// `;` / `%` comment case the round-3 fix covered.
    #[test]
    fn lf_to_crlf_outside_strings_handles_emacs_directive_with_quote() {
        let s = "#+title: \"My Book\n2024-01-01 open Assets:A\n";
        let out = lf_to_crlf_outside_strings(s);
        assert_eq!(out, "#+title: \"My Book\r\n2024-01-01 open Assets:A\r\n");
    }

    #[test]
    fn lf_to_crlf_outside_strings_handles_shebang_with_quote() {
        let s = "#!shebang \"quote\n2024-01-01 open Assets:A\n";
        let out = lf_to_crlf_outside_strings(s);
        assert_eq!(out, "#!shebang \"quote\r\n2024-01-01 open Assets:A\r\n");
    }

    /// `#` NOT at line start is a TAG / HASH token; the state
    /// machine must NOT treat it as a comment opener.
    #[test]
    fn lf_to_crlf_outside_strings_hash_mid_line_is_not_comment() {
        let s = "2024-01-15 * \"x\" #tag1\n  Assets:A 1 USD\n";
        let out = lf_to_crlf_outside_strings(s);
        // Every LF outside strings becomes CRLF — including the
        // one ending the tag-bearing line.
        assert!(out.contains("#tag1\r\n"), "got: {out:?}");
        assert!(out.ends_with("\r\n"), "got: {out:?}");
    }

    /// Regression for Copilot #2 inline review on PR #1284: a
    /// previous `emit_amount_expression` dropped leading unary
    /// signs and parens, flipping the sign on
    /// `2024-01-15 balance Assets:A
    /// -1.00 USD` to `1.00 USD` — silent data corruption (a debit
    /// asserted as a credit). Byte-exact pins on every shape.
    #[test]
    fn balance_price_preserve_leading_unary_and_parens() {
        // Bare leading minus on balance.
        let src = "2024-01-15 balance Assets:A -1.00 USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:A -1.00 USD\n"
        );

        // Bare leading minus on price (sign flip would change
        // every quote on the user's commodity).
        let src = "2024-01-15 price USD -1.00 EUR\n";
        assert_eq!(format_source(src), "2024-01-15 price USD -1.00 EUR\n");

        // Leading parenthesized expression. The previous code
        // dropped the `(`, which made the trailing `)` unbalanced
        // AND made the first-CURRENCY scan find the wrong token.
        let src = "2024-01-15 balance Assets:A (1 + 2) USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:A (1 + 2) USD\n"
        );

        // Leading minus on a parenthesized arithmetic expression.
        let src = "2024-01-15 balance Assets:A -(1 + 2) USD\n";
        assert_eq!(
            format_source(src),
            "2024-01-15 balance Assets:A -(1 + 2) USD\n"
        );
    }

    /// Regression for Copilot #1 inline review on PR #1284:
    /// `collect_trailing_comment` used `?` on the header-terminating
    /// NEWLINE, silently dropping same-line trailing comments at
    /// EOF when the file had no final newline. The canonical
    /// formatter restores the trailing newline, but the dropped
    /// comment was already gone — a real-world case for editors
    /// that don't insert a trailing newline on save.
    #[test]
    fn trailing_comment_preserved_at_eof_without_newline() {
        let src = "2024-01-15 open Assets:A ; trailing";
        assert_eq!(format_source(src), "2024-01-15 open Assets:A ; trailing\n");
    }

    #[test]
    fn try_format_source_returns_ok_on_clean_input() {
        let src = "2024-01-15 open Assets:Cash\n";
        let out = super::try_format_source(src).expect("clean input should format");
        assert_eq!(out, super::format_source(src));
    }

    #[test]
    fn try_format_source_returns_err_on_parse_error() {
        // Bare `unparsable` text triggers parser errors. The
        // helper must surface them instead of silently emitting
        // canonical text around a broken file.
        let src = "this is not a directive at all\n";
        let err = super::try_format_source(src).expect_err("garbage should error");
        assert!(!err.is_empty(), "errors must not be empty");
    }

    #[test]
    fn cr_outside_strings_present_distinguishes_in_string_cr() {
        // CR inside a multi-line string literal must NOT count —
        // the formatter wouldn't fold it.
        let in_string_only = "2024-01-15 note Assets:Bank \"line1\r\nline2\"\n";
        assert!(!super::cr_outside_strings_present(in_string_only));

        // CR outside any string literal (CRLF line terminator)
        // counts — that's what crlf_to_lf_outside_strings would
        // fold.
        let crlf_terminator = "2024-01-01 open Assets:A\r\n";
        assert!(super::cr_outside_strings_present(crlf_terminator));

        // No `\r` at all — fast path.
        let lf_only = "2024-01-01 open Assets:A\n";
        assert!(!super::cr_outside_strings_present(lf_only));

        // CR inside a `;` comment is outside any string and counts.
        // (Beancount lexer's comment regex excludes the newline, so
        // the comment region ends at `\r`; either way, the predicate
        // says "yes, the formatter would fold this byte".)
        let comment_with_cr = "; comment with \"quote\rstuff\n";
        assert!(super::cr_outside_strings_present(comment_with_cr));
    }

    #[test]
    fn canonicalize_directives_directive_count_mismatch_is_reported() {
        // Drive the new DirectiveCountMismatch error variant.
        // Today's Directive variants all round-trip with matching
        // counts, so this test pins the Display rendering of the
        // variant (the user-facing message). The positive-count-
        // match path is exercised by
        // `canonicalize_directives_positive_count_check` below.
        let err = super::CanonicalizeError::DirectiveCountMismatch {
            input: 3,
            reparsed: 2,
        };
        let msg = format!("{err}");
        assert!(msg.contains("3 directive(s)"), "got: {msg}");
        assert!(msg.contains("2 survived"), "got: {msg}");
        assert!(msg.contains("rledger bug"), "got: {msg}");
    }

    /// Single source of truth for the variant → fixture mapping
    /// used by both the compile-time exhaustiveness check
    /// ([`_directive_variant_fixture_coverage`]) and the runtime
    /// semantic check
    /// ([`directive_variant_fixture_names_resolve_in_matrix`]).
    ///
    /// Each tuple is `(VariantName, fixture_name)`. The
    /// `VariantName` half is the string the runtime check uses to
    /// confirm the fixture parses to that variant; the
    /// `fixture_name` half is what the compile-time match returns
    /// for the same variant. A future `Directive::Hedge` variant
    /// only ships with canonical-form coverage if BOTH a new
    /// arm is added to the compile-time match AND a row here
    /// names a fixture that actually produces a `Hedge` on parse.
    const DIRECTIVE_VARIANT_FIXTURE_MAP: &[(&str, &str)] = &[
        ("Transaction", "transaction_with_cost_and_price"),
        ("Balance", "balance_with_arithmetic_and_tolerance"),
        ("Open", "only_directive"),
        ("Close", "close_directive"),
        ("Commodity", "commodity_directive"),
        ("Pad", "pad_directive"),
        ("Event", "event_directive"),
        ("Query", "query_directive"),
        ("Note", "note_directive"),
        ("Document", "document_directive"),
        ("Price", "price_with_thousands_separator"),
        ("Custom", "custom_with_date_value"),
    ];

    /// Lookup helper: variant tag string → fixture name. Used by
    /// the compile-time match below. Panics if the variant is not
    /// in the map (which would be an internal-consistency bug, not
    /// a user-facing case).
    const fn fixture_for_variant(tag: &str) -> &'static str {
        let mut i = 0;
        while i < DIRECTIVE_VARIANT_FIXTURE_MAP.len() {
            let (v, f) = DIRECTIVE_VARIANT_FIXTURE_MAP[i];
            // const_str equality: compare byte slices.
            let v_bytes = v.as_bytes();
            let t_bytes = tag.as_bytes();
            if v_bytes.len() == t_bytes.len() {
                let mut k = 0;
                let mut eq = true;
                while k < v_bytes.len() {
                    if v_bytes[k] != t_bytes[k] {
                        eq = false;
                        break;
                    }
                    k += 1;
                }
                if eq {
                    return f;
                }
            }
            i += 1;
        }
        panic!("DIRECTIVE_VARIANT_FIXTURE_MAP missing entry for variant tag");
    }

    /// Compile-time check that every `rustledger_core::Directive`
    /// variant has at least one source-text fixture in
    /// [`IDEMPOTENCE_MATRIX`] exercising its emit path. The
    /// function NEVER runs — its body is an exhaustive `match` over
    /// the `Directive` enum. Adding a new variant breaks
    /// compilation unless the author adds a match arm referencing
    /// `fixture_for_variant("NewVariantName")`, AND adds a row to
    /// [`DIRECTIVE_VARIANT_FIXTURE_MAP`] naming the fixture. The
    /// runtime test then confirms the fixture parses to a directive
    /// of that variant.
    ///
    /// The non-`Directive` pragma-style directives (Pushtag,
    /// Poptag, Pushmeta, Popmeta, options, includes, plugins)
    /// don't appear in the typed `Directive` enum; they're covered
    /// by separate fixtures whose names map directly into
    /// `IDEMPOTENCE_MATRIX`.
    #[allow(dead_code)]
    fn _directive_variant_fixture_coverage(d: &rustledger_core::Directive) -> &'static str {
        match d {
            rustledger_core::Directive::Transaction(_) => fixture_for_variant("Transaction"),
            rustledger_core::Directive::Balance(_) => fixture_for_variant("Balance"),
            rustledger_core::Directive::Open(_) => fixture_for_variant("Open"),
            rustledger_core::Directive::Close(_) => fixture_for_variant("Close"),
            rustledger_core::Directive::Commodity(_) => fixture_for_variant("Commodity"),
            rustledger_core::Directive::Pad(_) => fixture_for_variant("Pad"),
            rustledger_core::Directive::Event(_) => fixture_for_variant("Event"),
            rustledger_core::Directive::Query(_) => fixture_for_variant("Query"),
            rustledger_core::Directive::Note(_) => fixture_for_variant("Note"),
            rustledger_core::Directive::Document(_) => fixture_for_variant("Document"),
            rustledger_core::Directive::Price(_) => fixture_for_variant("Price"),
            rustledger_core::Directive::Custom(_) => fixture_for_variant("Custom"),
        }
    }

    #[test]
    fn directive_variant_fixture_names_resolve_in_matrix() {
        // Runtime mirror of the compile-time match above:
        //
        //   (1) every fixture name appears in IDEMPOTENCE_MATRIX;
        //   (2) parsing that fixture produces AT LEAST one
        //       directive of the variant the map row names.
        //
        // Without check (2) the compile-time match is satisfied by
        // any fixture-name string — a future contributor adding
        // a row `("Hedge", "only_comment")` would compile, the
        // lookup would resolve, and Hedge would ship with zero
        // canonical-form coverage. The semantic check rejects that
        // by parsing the named fixture and inspecting the
        // directive variant.
        use rustledger_core::Directive;
        fn matches_variant(d: &Directive, expected: &str) -> bool {
            matches!(
                (d, expected),
                (Directive::Transaction(_), "Transaction")
                    | (Directive::Balance(_), "Balance")
                    | (Directive::Open(_), "Open")
                    | (Directive::Close(_), "Close")
                    | (Directive::Commodity(_), "Commodity")
                    | (Directive::Pad(_), "Pad")
                    | (Directive::Event(_), "Event")
                    | (Directive::Query(_), "Query")
                    | (Directive::Note(_), "Note")
                    | (Directive::Document(_), "Document")
                    | (Directive::Price(_), "Price")
                    | (Directive::Custom(_), "Custom")
            )
        }
        for (variant, name) in DIRECTIVE_VARIANT_FIXTURE_MAP {
            let (_, src) = IDEMPOTENCE_MATRIX
                .iter()
                .find(|(n, _)| *n == *name)
                .unwrap_or_else(|| {
                    panic!(
                        "fixture `{name}` is named by \
                     DIRECTIVE_VARIANT_FIXTURE_MAP but missing from \
                     IDEMPOTENCE_MATRIX"
                    )
                });
            let parsed = crate::parse(src);
            let found = parsed
                .directives
                .iter()
                .any(|s| matches_variant(&s.value, variant));
            assert!(
                found,
                "fixture `{name}` is mapped to `Directive::{variant}` by \
                 DIRECTIVE_VARIANT_FIXTURE_MAP, but parsing it produced \
                 no directive of that variant (got {:?}). This silently \
                 leaves the variant without canonical-form coverage.",
                parsed
                    .directives
                    .iter()
                    .map(|s| std::mem::discriminant(&s.value))
                    .collect::<Vec<_>>()
            );
        }
    }

    /// Coverage-mirror check: every `matrix_name` half of the
    /// `MIRROR_PAIRS` table in the file-pair integration test
    /// (`crates/rustledger-parser/tests/format_compat.rs`) must
    /// exist as an entry in [`IDEMPOTENCE_MATRIX`]. The
    /// integration test asserts the symmetric half (every
    /// `file_pair_name` exists as a directory under `cases/`).
    /// Together the two checks guarantee that retiring a
    /// bug-class fixture from EITHER side forces an edit to
    /// `MIRROR_PAIRS` - which surfaces in review and prevents
    /// the silent one-sided drop the README's "two audience" split
    /// design would otherwise admit.
    ///
    /// Hand-maintained copy of the matrix half of the table.
    /// Editing `MIRROR_PAIRS` in the integration test requires
    /// editing this list too; the test below fires otherwise.
    #[test]
    fn idempotence_matrix_mirrors_format_compat_pairs() {
        const MIRROR_PAIRS_MATRIX_HALF: &[&str] = &[
            "balance_leading_unary_minus",
            "balance_leading_parenthesized_expression",
            "price_leading_unary_minus",
            "cost_spec_with_negative",
            "cost_spec_with_comma_and_date",
            "transaction_with_per_unit_plus_total_cost",
            "metadata_unary_minus",
            "metadata_arithmetic",
            "non_latin_account_name",
            "posting_with_trailing_comment",
            "multiline_note_string",
            "comment_containing_quote",
            "transaction_with_tags_and_links",
            "custom_with_date_value",
            "options_and_includes",
            "balance_assertion_with_meta",
            "crlf_input",
        ];
        let matrix_names: std::collections::BTreeSet<&str> =
            IDEMPOTENCE_MATRIX.iter().map(|(name, _)| *name).collect();
        let missing: Vec<&&str> = MIRROR_PAIRS_MATRIX_HALF
            .iter()
            .filter(|name| !matrix_names.contains(*name))
            .collect();
        assert!(
            missing.is_empty(),
            "IDEMPOTENCE_MATRIX is missing the matrix-half of MIRROR_PAIRS: {missing:?}. \
             Either re-add the entry to IDEMPOTENCE_MATRIX, or edit MIRROR_PAIRS in \
             tests/format_compat.rs to retire the pair from BOTH sides.",
        );
    }

    /// Property test: the `SourceState` classification used by the
    /// line-ending helpers must agree with the lexer's
    /// classification on every byte of a corpus of fixtures.
    ///
    /// Concretely: for every byte offset in every fixture, the
    /// state machine's `InString` periods MUST line up with the
    /// lexer's STRING token spans, and its `InComment` periods MUST
    /// line up with the union of COMMENT / SHEBANG /
    /// `EMACS_DIRECTIVE` token spans. A divergence — e.g. the lexer
    /// gains a new comment lexeme that the state machine treats as
    /// code — fails this test instead of silently mutating user
    /// bytes inside the new lexeme on a line-ending round-trip.
    #[test]
    fn source_state_classification_agrees_with_lexer() {
        use crate::logos_lexer::{Token, tokenize_lossless};

        for (name, src) in IDEMPOTENCE_MATRIX {
            // Run the lexer to get authoritative classification of
            // each token. Build a per-byte map of expected state.
            let tokens = tokenize_lossless(src);
            let mut expected = vec![SourceState::Code; src.len()];
            for (token, span) in &tokens {
                let classify = match token {
                    Token::String(_) => Some(SourceState::InString),
                    Token::Comment(_) | Token::Shebang(_) | Token::EmacsDirective(_) => {
                        Some(SourceState::InComment)
                    }
                    _ => None,
                };
                if let Some(state) = classify {
                    for byte in &mut expected[span.start..span.end] {
                        *byte = state;
                    }
                }
            }

            // Run the state-machine classifier and compare per
            // byte. We skip ONLY the exact bytes where a
            // transition fires — the lexer includes those bytes
            // inside the resulting token while the state machine
            // tags them with the PRE-transition state (the
            // 'opener' is still Code, the closing LF is still
            // InComment). Tracking the transition indices
            // explicitly (rather than skipping every `"`/`;`/`%`
            // / newline byte) means a state-machine bug at any
            // non-transition `"`/`;`/`%` byte — e.g. inside a
            // comment or string — surfaces as a real failure
            // instead of being silently masked.
            let (actual, transitions) = classify_source_bytes_with_transitions(src);

            for (i, (&want, &got)) in expected.iter().zip(actual.iter()).enumerate() {
                if transitions.contains(&i) {
                    continue;
                }
                assert_eq!(
                    want,
                    got,
                    "state-machine / lexer disagreement on fixture `{name}` \
                     at byte {i} ({:?}): lexer said {want:?}, state machine said {got:?}",
                    src.as_bytes()[i] as char
                );
            }
        }
    }

    /// Walk `s` through the same state-machine logic the
    /// line-ending helpers use, returning a per-byte classification
    /// AND the set of byte indices where a state transition
    /// fired. The transition indices are the ONLY bytes where the
    /// state machine and the lexer can legitimately disagree (the
    /// off-by-one at opener / closer / terminator); callers
    /// comparing against the lexer should skip exactly those
    /// indices and assert agreement everywhere else.
    fn classify_source_bytes_with_transitions(
        s: &str,
    ) -> (Vec<SourceState>, std::collections::HashSet<usize>) {
        let (body, bom_len) = match s.strip_prefix('\u{FEFF}') {
            Some(rest) => (rest, '\u{FEFF}'.len_utf8()),
            None => (s, 0),
        };
        let mut out: Vec<SourceState> = vec![SourceState::Code; s.len()];
        let mut transitions = std::collections::HashSet::new();
        let mut chars = body.char_indices().peekable();
        let mut state = SourceState::Code;
        let mut prev_was_backslash = false;
        while let Some((rel_i, ch)) = chars.next() {
            let i = bom_len + rel_i;
            let peek = chars.peek().map(|&(_, c)| c);
            // Classify THIS byte under the state BEFORE advancing.
            for byte in &mut out[i..i + ch.len_utf8()] {
                *byte = state;
            }
            let prev_state = state;
            let next_state = advance_source_state(ch, peek, state, &mut prev_was_backslash);
            // Record only OPENING transitions and the comment-
            // closing newline, where the state machine and lexer
            // legitimately disagree on this single byte:
            //   - Code → InString : opening `"` is Code-side but
            //     the lexer puts it inside the STRING token.
            //   - Code → InComment: opening `;` / `%` / `#!` /
            //     `#+` is Code-side but the lexer puts it inside
            //     the COMMENT / SHEBANG / EMACS_DIRECTIVE token.
            //   - InComment → Code: the `\n` ending the comment is
            //     classified InComment by the state machine but
            //     sits OUTSIDE the comment token (the lexer's
            //     `[^\n\r]*` excludes it).
            // The InString → Code transition (closing `"`) is NOT
            // a disagreement: the state machine still tags that
            // byte as InString (pre-transition), and the lexer
            // includes the closing `"` inside the STRING token.
            // Skipping it would silently mask a real bug.
            if next_state != state {
                let opening = matches!(prev_state, SourceState::Code)
                    && matches!(next_state, SourceState::InString | SourceState::InComment);
                let comment_close = matches!(prev_state, SourceState::InComment)
                    && matches!(next_state, SourceState::Code);
                if opening || comment_close {
                    transitions.insert(i);
                    // For a `#!` or `#+` opener the lexer's token
                    // span begins at the `#`, so the second byte
                    // (`!` / `+`) is also a "before the lexer's
                    // token start" byte the state machine tags as
                    // Code. Record it too.
                    if matches!(ch, '#') && matches!(peek, Some('!' | '+')) {
                        transitions.insert(i + 1);
                    }
                }
            }
            state = next_state;
        }
        (out, transitions)
    }

    #[test]
    fn canonicalize_directives_positive_count_check() {
        // Pin the success path of the count check: pass a real
        // multi-directive input through canonicalize_directives and
        // assert that the output round-trips to the SAME directive
        // count. Without this test, a regression that always
        // returned CountMismatch (e.g. `==` instead of `!=` on the
        // count comparison) would be caught only on production
        // calls, not in CI. Together with the Display test above,
        // this gives coverage of both arms of the count guard.
        use rustledger_core::format::FormatConfig;
        let cfg = FormatConfig::default();
        let src = "2024-01-01 open Assets:Cash\n2024-01-02 open Assets:Bank\n2024-01-03 close Assets:Cash\n";
        let parsed = crate::parse(src);
        assert_eq!(
            parsed.directives.len(),
            3,
            "fixture must parse to 3 directives"
        );
        let dirs: Vec<&rustledger_core::Directive> =
            parsed.directives.iter().map(|s| &s.value).collect();
        let formatted = super::canonicalize_directives(dirs.iter().copied(), &cfg)
            .expect("canonicalize_directives should succeed on this input");
        let reparsed = crate::parse(&formatted);
        assert_eq!(
            reparsed.directives.len(),
            3,
            "count check accepted but round-trip dropped directives: {formatted}"
        );
    }

    // ---- format_node_range -----------------------------------------

    /// Parse `source` via the same pipeline `format_source` uses
    /// so the resulting `SyntaxNode`'s `TextRange`s are in the
    /// same byte frame `format_node_range`'s `range` argument
    /// is expected to use (post-BOM-strip, post-CRLF-to-LF).
    /// Returns the syntax node + the normalized source text so
    /// tests can compute byte offsets by `.find()`.
    fn parse_for_range(source: &str) -> (crate::SyntaxNode, String) {
        let (stripped, _bom) = crate::bom::strip_leading(source);
        let normalized = crlf_to_lf_outside_strings(stripped).to_string();
        let sf = SourceFile::parse(&normalized);
        (sf.syntax().clone(), normalized)
    }

    fn ts(n: usize) -> rowan::TextSize {
        rowan::TextSize::try_from(n).expect("offset fits TextSize")
    }

    /// For any selection covering the whole file, the result text
    /// equals `format_node(node)`. Pins the round-trip invariant
    /// the design rests on: range formatting is the whole-file
    /// formatter restricted to a range, not a parallel canonical
    /// form.
    #[test]
    fn format_node_range_full_range_matches_format_node() {
        let source = "\
2024-01-01 open Assets:Bank USD
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
2024-01-31 close Assets:Bank
";
        let (node, src) = parse_for_range(source);
        let full = rowan::TextRange::new(ts(0), ts(src.len()));
        let (snap, formatted) =
            format_node_range(&node, full).expect("full range must include all directives");
        assert_eq!(
            snap,
            rowan::TextRange::new(ts(0), ts(src.len())),
            "snap range should be the whole file's textual span"
        );
        assert_eq!(formatted, format_node(&node));
    }

    /// A selection that hits only inter-directive whitespace
    /// (no directive intersected, no top-level comment
    /// intersected) returns `None` — the caller surfaces this
    /// as an empty `Vec<TextEdit>`.
    #[test]
    fn format_node_range_trivia_only_returns_none() {
        // The phase-2.0 Directive-Terminator Rule puts every
        // inter-directive blank line on the next directive's
        // leading trivia, so any byte index between two
        // directives is INSIDE the next directive's text_range.
        // The only way to reach a truly trivia-only selection
        // is a source that has no directives at all (file is
        // pure whitespace). That is the case worth pinning —
        // the LSP handler maps `None` to an empty
        // `Vec<TextEdit>`, which is exactly the right "nothing
        // to format" response for a whitespace-only buffer.
        let (empty, _) = parse_for_range("\n\n\n");
        let sel = rowan::TextRange::new(ts(0), ts(3));
        assert!(format_node_range(&empty, sel).is_none());
    }

    /// Selecting only the first directive's content (the
    /// transaction) snaps to that directive and the second
    /// directive is left out of both the snap and the output.
    #[test]
    fn format_node_range_single_directive() {
        let source = "\
2024-01-01 open Assets:Bank USD
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
";
        let (node, src) = parse_for_range(source);
        // Position the selection inside the `open` line. Use
        // the byte offset of the word `open` so the test is
        // robust to whitespace changes in the fixture.
        let open_byte = src.find("open").expect("fixture contains 'open'");
        let sel = rowan::TextRange::new(ts(open_byte), ts(open_byte + "open".len()));
        let (snap, formatted) = format_node_range(&node, sel).expect("intersects 1 directive");

        // Snap should start at byte 0 (the open directive's
        // text_range starts at the file's start) and end at
        // the open directive's terminating newline.
        let open_end = src.find('\n').expect("first directive has terminator") + 1;
        assert_eq!(snap.start(), ts(0));
        assert_eq!(snap.end(), ts(open_end));
        // Output is exactly the open directive's canonical form
        // + its `\n` terminator. No second-directive content.
        assert_eq!(formatted, "2024-01-01 open Assets:Bank USD\n");
    }

    /// Multi-directive selection: the author's inter-directive
    /// blank lines are preserved (a blank stays a blank; grouped
    /// stays grouped), matching whole-file formatting (#1325).
    #[test]
    fn format_node_range_multi_directive_preserves_blank_lines() {
        // #1325: range formatting preserves the author's inter-directive
        // blank lines, identically to whole-file formatting. A source
        // with a blank between the two directives keeps it...
        let spaced = "\
2024-01-01 open Assets:Bank USD

2024-01-31 close Assets:Bank
";
        let (node, src) = parse_for_range(spaced);
        let sel = rowan::TextRange::new(ts(0), ts(src.len()));
        let (snap, formatted) = format_node_range(&node, sel).expect("intersects 2 directives");
        assert_eq!(snap, rowan::TextRange::new(ts(0), ts(src.len())));
        assert_eq!(formatted, spaced, "the blank separator must be preserved");

        // ...and a grouped source (no blank) stays grouped, rather than
        // having a separator inserted.
        let grouped = "\
2024-01-01 open Assets:Bank USD
2024-01-31 close Assets:Bank
";
        let (node2, src2) = parse_for_range(grouped);
        let sel2 = rowan::TextRange::new(ts(0), ts(src2.len()));
        let (_, formatted2) = format_node_range(&node2, sel2).expect("intersects 2 directives");
        assert_eq!(formatted2, grouped, "grouped directives must stay grouped");
    }

    #[test]
    fn format_node_range_first_directive_in_snap_keeps_leading_blank() {
        // Regression (Copilot review of #1325): when the selection
        // covers only the SECOND directive, its predecessor sits outside
        // the snap, but the blank line between them is the second
        // directive's leading trivia and therefore inside the snapped
        // range. Range formatting must re-emit it, not silently delete
        // the blank line above the selection.
        let source = "2024-01-01 open Assets:Bank USD\n\n2024-01-31 close Assets:Bank\n";
        let (node, src) = parse_for_range(source);
        // Cursor inside the second (close) directive only.
        let close_byte = src.find("close").expect("fixture has 'close'");
        let cursor = rowan::TextRange::new(ts(close_byte), ts(close_byte));
        let (snap, formatted) = format_node_range(&node, cursor).expect("intersects close");
        // The leading blank is preserved in the replacement text...
        assert_eq!(formatted, "\n2024-01-31 close Assets:Bank\n");
        // ...so applying the edit leaves the blank line intact.
        let mut result = src;
        result.replace_range(
            usize::from(snap.start())..usize::from(snap.end()),
            &formatted,
        );
        assert_eq!(
            result, source,
            "range-formatting the second directive must not delete the blank above it"
        );
    }

    /// Cursor-only (zero-width) selection inside a directive
    /// snaps to that directive. The cursor convention: inside
    /// or at the directive's start byte counts as inside;
    /// boundary at the directive's end belongs to the next
    /// child.
    #[test]
    fn format_node_range_cursor_inside_directive() {
        let source = "\
2024-01-01 open Assets:Bank USD
2024-01-31 close Assets:Bank
";
        let (node, src) = parse_for_range(source);
        // Cursor on the `c` of `close` (line 2 of the fixture).
        let close_byte = src.find("close").expect("fixture has 'close'");
        let cursor = rowan::TextRange::new(ts(close_byte), ts(close_byte));
        let (snap, formatted) = format_node_range(&node, cursor).expect("intersects close");
        // Snap starts at the close directive's text_range start.
        // Per Directive-Terminator Rule the second directive
        // OWNS the leading inter-directive trivia — so snap
        // starts immediately after the first directive's
        // terminator newline.
        let close_dir_start = src
            .find("\n2024-01-31")
            .map(|n| n + 1)
            .expect("close directive starts on its own line");
        assert_eq!(snap.start(), ts(close_dir_start));
        assert_eq!(snap.end(), ts(src.len()));
        assert_eq!(formatted, "2024-01-31 close Assets:Bank\n");
    }

    /// Cursor exactly at the start of a directive snaps to
    /// that directive (start-boundary inclusion rule).
    #[test]
    fn format_node_range_cursor_at_directive_start_includes_directive() {
        let source = "\
2024-01-01 open Assets:Bank USD
2024-01-31 close Assets:Bank
";
        let (node, _src) = parse_for_range(source);
        // Cursor at byte 0 = start of first directive.
        let cursor = rowan::TextRange::new(ts(0), ts(0));
        let (_snap, formatted) = format_node_range(&node, cursor).expect("intersects open");
        // Only the OPEN should be formatted, not the close.
        assert!(formatted.starts_with("2024-01-01 open"));
        assert!(!formatted.contains("close"));
    }

    /// Selection containing a top-level standalone comment
    /// (file-leading or between-directive comment that the
    /// trivia attachment policy puts on `SOURCE_FILE`) includes
    /// the comment in both the snap and the output.
    #[test]
    fn format_node_range_includes_top_level_comments() {
        let source = "\
; header
2024-01-01 open Assets:Bank USD
";
        let (node, src) = parse_for_range(source);
        let sel = rowan::TextRange::new(ts(0), ts(src.len()));
        let (snap, formatted) = format_node_range(&node, sel).expect("intersects both");
        assert_eq!(snap, rowan::TextRange::new(ts(0), ts(src.len())));
        // Header comment, then directive on the next line. No
        // canonical blank between a file-level comment group
        // and a directive (matches format_node's policy).
        assert_eq!(formatted, "; header\n2024-01-01 open Assets:Bank USD\n");
    }

    /// A selection that lands entirely inside an `ERROR_NODE`
    /// (no Directive intersected) returns None. Matches
    /// `format_node`'s policy of skipping `ERROR_NODE` children
    /// at the top level.
    #[test]
    fn format_node_range_error_node_only_returns_none() {
        // `}}}` at top level isn't a directive — the parser
        // wraps it in an ERROR_NODE.
        let source = "}}}\n";
        let (node, src) = parse_for_range(source);
        let sel = rowan::TextRange::new(ts(0), ts(src.len()));
        assert!(format_node_range(&node, sel).is_none());
    }

    /// Past-EOF selection still works: the snap clamps to the
    /// last child that intersects within the file. (rowan's
    /// `TextRange` is bounded by usize but `format_node_range`
    /// doesn't validate `range` against file length — bytes past
    /// EOF can never intersect any child, so the rule is
    /// degenerate but well-defined.)
    #[test]
    fn format_node_range_past_eof_clamps() {
        let source = "2024-01-01 open Assets:Bank USD\n";
        let (node, src) = parse_for_range(source);
        let past_eof = rowan::TextRange::new(ts(src.len()), ts(src.len() + 1000));
        // The cursor / range is past EOF — no child intersects.
        assert!(format_node_range(&node, past_eof).is_none());
        // But a range that STRADDLES EOF still snaps to the
        // last intersecting directive.
        let straddle = rowan::TextRange::new(ts(0), ts(src.len() + 1000));
        let (snap, formatted) = format_node_range(&node, straddle).expect("intersects open");
        assert_eq!(snap, rowan::TextRange::new(ts(0), ts(src.len())));
        assert_eq!(formatted, "2024-01-01 open Assets:Bank USD\n");
    }

    /// A cursor inside a posting (sub-directive position) snaps
    /// up to the enclosing transaction — the design pins
    /// "round to top-level directive boundaries, no finer."
    #[test]
    fn format_node_range_cursor_in_posting_snaps_to_transaction() {
        let source = "\
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
";
        let (node, src) = parse_for_range(source);
        // Position the cursor on the `B` of `Bank` in the
        // first posting.
        let bank_byte = src.find("Bank").expect("fixture has Bank");
        let cursor = rowan::TextRange::new(ts(bank_byte), ts(bank_byte));
        let (snap, _formatted) = format_node_range(&node, cursor).expect("intersects transaction");
        // Snap covers the WHOLE transaction (start of file
        // through final posting's newline).
        assert_eq!(snap.start(), ts(0));
        assert_eq!(snap.end(), ts(src.len()));
    }

    /// Selection straddling an `ERROR_NODE` between two valid
    /// directives: snap range would cover the union (including
    /// `ERROR_NODE` bytes), so `format_node_range` returns
    /// `None` instead of silently deleting the error content.
    ///
    /// This is the deliberate divergence from `format_node`'s
    /// whole-file policy. `format_source(broken_source)` does
    /// drop `ERROR_NODE` content — but that path's callers
    /// (`rledger format` CLI, FFI `format.entry`) opt into
    /// content loss by invoking the canonical-form pipeline. The
    /// per-handler LSP `textDocument/rangeFormatting` path has no
    /// such opt-in, so it refuses to delete user content the
    /// parser couldn't classify. See the function's rustdoc for
    /// the per-handler asymmetry rationale.
    #[test]
    fn format_node_range_bails_when_snap_covers_error_node() {
        let source = "\
2024-01-01 open Assets:Bank USD
}}}garbage{{{
2024-01-31 close Assets:Bank
";
        let (node, src) = parse_for_range(source);
        let sel = rowan::TextRange::new(ts(0), ts(src.len()));
        assert!(
            format_node_range(&node, sel).is_none(),
            "selection covering both directives + ERROR_NODE between them must bail \
             to avoid silently deleting the garbage line — got Some output",
        );
    }

    /// Selection that intersects only the FIRST valid directive
    /// in a broken file (no `ERROR_NODE` byte in the snap range)
    /// still formats. Pins that the `ERROR_NODE` bail is precisely
    /// scoped to the snap range, not to "the file has any
    /// `ERROR_NODE` at all".
    #[test]
    fn format_node_range_formats_directive_when_snap_does_not_cover_error_node() {
        let source = "\
2024-01-01 open Assets:Bank USD
}}}garbage{{{
2024-01-31 close Assets:Bank
";
        let (node, src) = parse_for_range(source);
        // Selection covers ONLY the open directive (first line +
        // its terminator). The ERROR_NODE on line 1 sits at byte
        // offset == open_end (length of first line including \n)
        // onward, OUTSIDE the snap range.
        let open_end = src.find('\n').expect("first directive has newline") + 1;
        let sel = rowan::TextRange::new(ts(0), ts(open_end));
        let (snap, formatted) =
            format_node_range(&node, sel).expect("selection covers only the open");
        assert_eq!(snap.start(), ts(0));
        assert_eq!(snap.end(), ts(open_end));
        assert_eq!(formatted, "2024-01-01 open Assets:Bank USD\n");
    }

    /// `format_node_with_alignment(node, compute_alignment(sf))` is
    /// byte-identical to `format_node(node)`. Pins the cache
    /// contract: passing the correct alignment is a pure
    /// optimization, NOT a behavior change.
    #[test]
    fn format_node_equals_format_node_with_alignment() {
        let fixtures: &[(&str, &str)] = &[
            ("empty", ""),
            ("open only", "2024-01-01 open Assets:Bank USD\n"),
            (
                "single txn",
                "\
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
",
            ),
            (
                "multi txn varying widths",
                "\
2024-01-15 * \"A\"
  Assets:Bank  -5.00 USD
  Expenses:Food
2024-02-15 * \"B\"
  Assets:Investment:Long:Path  -123456.78 USD
  Expenses:Tax  100.00 USD
",
            ),
        ];
        for (label, source) in fixtures {
            let (node, _src) = parse_for_range(source);
            let source_file = SourceFile::cast(node.clone()).unwrap();
            let alignment = compute_alignment(&source_file, GroupingStyle::default());
            assert_eq!(
                format_node(&node),
                format_node_with_alignment(&node, alignment),
                "format_node_with_alignment must match format_node for {label}",
            );
        }
    }

    /// `format_node_range_with_alignment(node, range, compute_alignment(sf))`
    /// matches `format_node_range(node, range)` byte-identically.
    /// Same shape as the previous test, for the range path.
    #[test]
    fn format_node_range_matches_format_node_range_with_alignment() {
        let source = "\
2024-01-15 * \"A\"
  Assets:Bank  -5.00 USD
  Expenses:Food
2024-02-15 * \"B\"
  Assets:Investment:Long:Path  -123456.78 USD
  Expenses:Tax  100.00 USD
";
        let (node, src) = parse_for_range(source);
        let source_file = SourceFile::cast(node.clone()).unwrap();
        let alignment = compute_alignment(&source_file, GroupingStyle::default());
        // Pin the equivalence on three ranges: whole file,
        // cursor inside the first transaction, cursor inside the
        // second.
        let sels = [
            rowan::TextRange::new(ts(0), ts(src.len())),
            rowan::TextRange::new(ts(0), ts(10)),
            rowan::TextRange::new(ts(src.len() - 10), ts(src.len())),
        ];
        for sel in sels {
            let uncached = format_node_range(&node, sel);
            let cached = format_node_range_with_alignment(&node, sel, alignment);
            assert_eq!(
                uncached, cached,
                "format_node_range_with_alignment must match \
                 format_node_range for range {sel:?}",
            );
        }
    }

    /// The cached [`crate::ParseResult::alignment`] value matches what
    /// `format_node` would compute on the parsed tree. End-to-end
    /// regression: an LSP caller passing `parse_result.alignment()`
    /// to `format_node_with_alignment` produces the same output
    /// as the bare `format_node` (uncached path).
    #[test]
    fn parse_result_alignment_drives_identical_format_output() {
        let source = "\
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
";
        let parse_result = crate::parse(source);
        let node = parse_result.syntax_node();
        assert_eq!(
            format_node(&node),
            format_node_with_alignment(&node, parse_result.alignment()),
            "ParseResult::alignment must drive identical format output to format_node",
        );
    }

    /// `format_source_with_parsed(parse(s), s) == format_source(s)`
    /// byte-identical across a representative fixture set including
    /// CRLF and BOM-prefixed sources. This is the load-bearing
    /// equivalence for the LSP `format_document` / FFI
    /// `format.source` / WASM `ParsedLedger::format` migrations:
    /// they swap `format_source(source)` for
    /// `format_source_with_parsed(parse_result, source)` on the
    /// assumption that the two produce the same output. Without
    /// this test, a future converter or formatter change that
    /// silently diverged the two paths would break canonical-form
    /// expectations in production.
    #[test]
    fn format_source_with_parsed_matches_format_source() {
        let fixtures: &[(&str, &str)] = &[
            ("empty", ""),
            ("comment only", "; hello\n"),
            (
                "single transaction LF",
                "\
2024-01-15 * \"Coffee\"
  Assets:Bank  -5.00 USD
  Expenses:Food
",
            ),
            (
                "multi transaction varying widths LF",
                "\
2024-01-15 * \"A\"
  Assets:Bank  -5.00 USD
  Expenses:Food
2024-02-15 * \"B\"
  Assets:Investment:Long:Path  -123456.78 USD
  Expenses:Tax  100.00 USD
",
            ),
            (
                "arithmetic amounts LF",
                "\
2024-01-15 * \"Split\"
  Assets:Bank  -10.00 + 5.00 USD
  Expenses:Misc
",
            ),
            (
                "CRLF source",
                "2024-01-15 * \"Coffee\"\r\n  Assets:Bank  -5.00 USD\r\n  Expenses:Food\r\n",
            ),
            ("BOM-prefixed", "\u{FEFF}2024-01-01 open Assets:Bank USD\n"),
            // BOM + CRLF — Windows-authored ledger with a BOM
            // prefix. `format_source` BOM-strips + CRLF→LF
            // normalizes before parsing. The cache path consumes
            // a CST that's BOM-stripped but NOT CRLF-normalized.
            // Byte-identity holds because the formatter rebuilds
            // canonical output from typed values (no trivia
            // passthrough).
            (
                "BOM + CRLF combination",
                "\u{FEFF}2024-01-15 * \"Coffee\"\r\n  Assets:Bank  -5.00 USD\r\n  Expenses:Food\r\n",
            ),
            // Parse-error file — exercises the fallback. Without
            // the `errors.is_empty()` guard, the cache path would
            // emit text for ERROR_NODE-wrapped content while
            // `format_source` would drop those bytes; identity
            // would fail. The fallback delegates to
            // `format_source(source)` so identity holds.
            (
                "parse errors (exercises fallback)",
                "2024-01-15 * \"x\"\n  Assets:Bank  -5.00 USD\n}}}garbage\n",
            ),
            // Bare-`\r` (classic Mac) line terminators. The
            // `format_source` path normalizes bare-CR to LF via
            // `crlf_to_lf_outside_strings`, then parses cleanly.
            // `parse_via_cst` does NOT normalize bare-CR, so the
            // CST sees broken syntax and `parse_result.errors`
            // is non-empty — the fallback fires. Byte-identity
            // holds via the same `format_source` delegation.
            (
                "bare CR line terminators (exercises fallback)",
                "2024-01-01 open Assets:Bank USD\r2024-01-02 open Assets:Cash USD\r",
            ),
        ];
        for (label, source) in fixtures {
            let parse_result = crate::parse(source);
            let baseline = format_source(source);
            let cached = format_source_with_parsed(&parse_result, source);
            assert_eq!(
                cached, baseline,
                "format_source_with_parsed must match format_source for {label}: \
                 baseline {baseline:?}, cached {cached:?}",
            );
        }
    }

    /// Mismatched-pair safety: in debug builds, passing a
    /// length-mismatched `(parse_result, source)` pair panics via
    /// the `debug_assert_eq!`. Release builds silently emit text
    /// for the wrong buffer — pairing the two arguments is the
    /// caller's responsibility, per this function's rustdoc.
    #[cfg(debug_assertions)]
    #[test]
    #[should_panic(expected = "source` whose length doesn't match")]
    fn format_source_with_parsed_panics_on_length_mismatch() {
        let parse_result = crate::parse("2024-01-01 open Assets:Bank USD\n");
        // Different length — debug_assert fires.
        let _ = format_source_with_parsed(&parse_result, "different");
    }

    /// A grouping style imposes separators, and the alignment pre-pass measures
    /// the GROUPED width — so the currency column still lines up.
    ///
    /// They agree because `compute_alignment` and the emitters are handed the
    /// SAME `GroupingStyle`. Measuring under one style and emitting under
    /// another is the mismatch that would reproduce #1290, which is why the two
    /// entry points that could express it are crate-private and every public
    /// one either measures its own alignment or is fixed to the default style.
    #[test]
    fn grouped_formatting_aligns_and_is_idempotent() {
        let src = "\
2020-01-02 * \"mixed magnitudes\"
  Assets:Bank                        1234567.89 USD
  Assets:VeryLongAccountName:Nested       12.00 USD
  Income:Sales                      -1234579.89 USD
";
        let grouped = format_source_grouped(src, grouped_style(&all_commas()));
        assert!(
            grouped.contains("1,234,567.89") && grouped.contains("-1,234,579.89"),
            "grouping must be imposed regardless of the source form:\n{grouped}"
        );
        // Currency column uniform => every ` USD` starts at the same column.
        let cols: Vec<usize> = grouped
            .lines()
            .filter(|l| l.contains(" USD"))
            .map(|l| l.find("USD").expect("USD"))
            .collect();
        assert!(
            cols.windows(2).all(|w| w[0] == w[1]),
            "currency column must stay uniform under grouping: {cols:?}\n{grouped}"
        );
        assert_eq!(
            format_source_grouped(&grouped, grouped_style(&all_commas())),
            grouped,
            "grouped formatting must be idempotent"
        );
    }

    /// Grouping is a TOTAL rewrite, not a preserve: it converges from either
    /// direction, so a file whose numerals are inconsistent still normalizes.
    /// That is the property `preserve` would have given up.
    #[test]
    fn grouping_converges_from_either_direction() {
        let mixed = "\
2020-01-02 * \"inconsistent source\"
  Assets:A  1,234,567.89 USD
  Assets:B     -1234567.89 USD
";
        let on = format_source_grouped(mixed, grouped_style(&all_commas()));
        assert_eq!(on.matches(',').count(), 4, "both numerals grouped:\n{on}");
        let off = format_source_grouped(mixed, GroupingStyle::default());
        assert!(!off.contains(','), "both numerals bare:\n{off}");
        // And each is a fixed point of its own rule.
        assert_eq!(format_source_grouped(&on, grouped_style(&all_commas())), on);
        assert_eq!(format_source_grouped(&off, GroupingStyle::default()), off);
    }

    /// The default entry point is untouched: every existing caller, and every
    /// ledger that has not opted in, gets byte-identical output.
    #[test]
    fn default_formatting_still_strips_separators() {
        let src = "2020-01-02 balance Assets:A  1,234.50 USD\n";
        assert_eq!(
            format_source(src),
            "2020-01-02 balance Assets:A 1234.50 USD\n"
        );
        assert_eq!(
            format_source(src),
            format_source_grouped(src, GroupingStyle::default())
        );
    }

    /// Grouped output must re-parse to the SAME values — the formatter may not
    /// emit text its own lexer rejects. Groups are three digits because that is
    /// all `(\d{1,3}(,\d{3})*|\d+)` admits.
    #[test]
    fn grouped_output_reparses_to_the_same_values() {
        for n in [
            "1",
            "12",
            "123",
            "1234",
            "1234567",
            "1234567.891",
            "0.5",
            "1000000",
        ] {
            let src = format!("2020-01-02 balance Assets:A  {n} USD\n");
            let grouped = format_source_grouped(&src, grouped_style(&all_commas()));
            let reparsed = crate::parse(&grouped);
            assert!(
                reparsed.errors.is_empty(),
                "grouped `{n}` -> `{}` must re-parse: {:?}",
                grouped.trim(),
                reparsed.errors
            );
            // And round-trips to the identical value.
            let before = crate::parse(&src);
            let val = |r: &crate::ParseResult| match &r.directives[0].value {
                rustledger_core::Directive::Balance(b) => b.amount.number,
                _ => panic!("balance"),
            };
            assert_eq!(val(&before), val(&reparsed), "value changed for `{n}`");
        }
    }

    /// A ledger-wide context used by the grouping tests.
    fn all_commas() -> rustledger_core::DisplayContext {
        let mut c = rustledger_core::DisplayContext::new();
        c.set_render_commas(true);
        c
    }

    fn grouped_style(ctx: &rustledger_core::DisplayContext) -> GroupingStyle<'_> {
        GroupingStyle::from_context(ctx)
    }

    /// A commodity may opt OUT of the ledger-wide default, so a 4000:1 currency
    /// can be grouped without also grouping two-digit USD amounts — the reason
    /// a single global boolean was not enough (#1892).
    #[test]
    fn grouping_is_resolved_per_commodity() {
        let mut ctx = rustledger_core::DisplayContext::new();
        ctx.set_render_commas(true);
        ctx.set_render_commas_for("USD", false);

        let src = "\
2020-01-02 * \"two currencies\"
  Assets:Local   1234567.89 IQD
  Assets:Dollars 1234567.89 USD
";
        let out = format_source_grouped(src, GroupingStyle::from_context(&ctx));
        assert!(
            out.contains("1,234,567.89 IQD"),
            "the ledger default applies to IQD:\n{out}"
        );
        assert!(
            out.contains("1234567.89 USD") && !out.contains("1,234,567.89 USD"),
            "USD opted out and must stay bare:\n{out}"
        );
    }

    /// The inverse: grouping declared on ONE commodity while the ledger default
    /// is off. This is the shape a user with a single hyperinflated currency
    /// actually wants.
    #[test]
    fn a_single_commodity_can_opt_in() {
        let mut ctx = rustledger_core::DisplayContext::new();
        ctx.set_render_commas_for("IQD", true);

        let src = "\
2020-01-02 * \"two currencies\"
  Assets:Local   1234567.89 IQD
  Assets:Dollars 1234567.89 USD
";
        let out = format_source_grouped(src, GroupingStyle::from_context(&ctx));
        assert!(out.contains("1,234,567.89 IQD"), "IQD opted in:\n{out}");
        assert!(
            out.contains("1234567.89 USD") && !out.contains("1,234,567.89 USD"),
            "USD keeps the (off) default:\n{out}"
        );
    }

    /// A context that groups nothing must produce the no-lookup style, so the
    /// overwhelming majority of ledgers pay nothing per numeral.
    #[test]
    fn a_context_that_groups_nothing_yields_the_default_style() {
        let mut ctx = rustledger_core::DisplayContext::new();
        ctx.set_fixed_precision("USD", 2);
        ctx.set_render_commas_for("USD", false);
        assert!(!ctx.renders_any_commas());
        let src = "2020-01-02 balance Assets:A  1234.50 USD\n";
        assert_eq!(
            format_source_grouped(src, GroupingStyle::from_context(&ctx)),
            format_source(src),
            "no declared grouping must be byte-identical to the default path"
        );
    }

    /// Formatting must not change what a balance ASSERTS.
    ///
    /// The tolerance is an expression: `~ 0.005 + 0.005 USD` and
    /// `~ 0.005 * 2 USD` both mean 0.010. `balance_tolerance` kept only the
    /// first NUMBER and rewrote them as `~ 0.005 USD`, halving the tolerance.
    /// The output reparsed cleanly and asserted something else, so `rledger
    /// format` could turn a passing ledger into a failing one with nothing to
    /// show for it -- the same truncation #1944 fixed in the parser and left
    /// standing here.
    ///
    /// Asserted on the parsed TOLERANCE VALUE rather than the text, because a
    /// text comparison would pass on any spelling that happens to round-trip
    /// while still meaning something new.
    #[test]
    fn formatting_preserves_the_tolerance_value() {
        for (src, want) in [
            ("2024-01-15 balance Assets:C 1.00 ~ 0.01 USD\n", "0.01"),
            (
                "2024-01-15 balance Assets:C 1.00 ~ 0.005 + 0.005 USD\n",
                "0.010",
            ),
            (
                "2024-01-15 balance Assets:C 1.00 ~ (0.005 + 0.005) USD\n",
                "0.010",
            ),
            (
                "2024-01-15 balance Assets:C 1.00 ~ 0.005 * 2 USD\n",
                "0.010",
            ),
            // main dropped this MINUS entirely, so `~ -0.01` formatted to
            // `~ 0.01` -- a second changed value in the same function.
            ("2024-01-15 balance Assets:C 1.00 ~ -0.01 USD\n", "-0.01"),
            (
                "2024-01-15 balance Assets:C 1.00 ~ 0.02 - 0.01 USD\n",
                "0.01",
            ),
            (
                "2024-01-15 balance Assets:C 1.00 ~ -0.005 + 0.015 USD\n",
                "0.010",
            ),
        ] {
            let tolerance = |text: &str| {
                crate::parse(text)
                    .directives
                    .iter()
                    .find_map(|d| match &d.value {
                        rustledger_core::Directive::Balance(b) => Some(b.tolerance),
                        _ => None,
                    })
                    .flatten()
                    .map(|t| t.to_string())
            };
            assert_eq!(
                tolerance(src).as_deref(),
                Some(want),
                "fixture itself must parse to {want}: {src}"
            );
            let formatted = format_source(src);
            assert_eq!(
                tolerance(&formatted).as_deref(),
                Some(want),
                "formatting changed the asserted tolerance\n  {src}  -> {formatted}"
            );
            // Every one of these is already canonical, so formatting must be a
            // no-op on the TEXT too. Without this the value assertions accept
            // any spelling that happens to evaluate the same -- and one did:
            // rendering a unary minus as `~ - 0.01` keeps the value and makes
            // the directive disagree with itself, since the amount on the same
            // line renders `-1.00` tight.
            assert_eq!(formatted, src, "already-canonical input was rewritten");
        }
    }

    /// `format` must not duplicate a balance tolerance.
    ///
    /// `emit_amount_expression` ran from the first NUMBER to the first
    /// CURRENCY, which SWALLOWED the `~ tolerance` clause; `emit_balance` then
    /// emitted it again via `balance_tolerance`. So
    /// `balance Assets:A 0.00 ~ 1234.5 USD` was rewritten as
    /// `... 0.00 ~ 1234.5 USD ~ 1234.5 USD` — stable across reformats, but
    /// almost certainly not valid beancount, whose balance grammar takes at
    /// most one tolerance. `--check` reported the ORIGINAL as unformatted, so a
    /// CI gate pushed users into the corrupted form.
    ///
    /// Pre-existing on main; unrelated to grouping.
    #[test]
    fn balance_tolerance_is_emitted_exactly_once() {
        for (src, want) in [
            (
                "2020-01-04 balance Assets:A 0.00 ~ 1234.5 USD\n",
                "2020-01-04 balance Assets:A 0.00 ~ 1234.5 USD\n",
            ),
            // A source that repeats the currency collapses to the one-currency
            // beancount form rather than keeping both.
            (
                "2020-01-04 balance Assets:A 0.00 USD ~ 0.05 USD\n",
                "2020-01-04 balance Assets:A 0.00 ~ 0.05 USD\n",
            ),
            // No tolerance: unchanged.
            (
                "2020-01-04 balance Assets:A 1234.50 USD\n",
                "2020-01-04 balance Assets:A 1234.50 USD\n",
            ),
            // Arithmetic still terminates at the currency, not the tilde.
            (
                "2020-01-04 balance Assets:A (1 + 5) / 2 USD\n",
                "2020-01-04 balance Assets:A (1 + 5) / 2 USD\n",
            ),
        ] {
            let out = format_source(src);
            assert_eq!(out, want, "formatting {src:?}");
            assert_eq!(format_source(&out), out, "not idempotent for {src:?}");
            // And the output must still parse — a formatter may not emit text
            // its own parser rejects.
            assert!(
                crate::parse(&out).errors.is_empty(),
                "output must re-parse: {out}"
            );
        }
    }
}