rustledger_parser/cst/

convert.rs

//! CST -> `ParseResult` converter.
//!
//! [`parse_via_cst`] is the implementation behind the public
//! [`crate::parse`] entry point. It walks the structured CST from
//! [`crate::parse_structured`] via the typed-AST surface in
//! [`crate::cst::ast`] and produces the legacy AST-shaped
//! [`ParseResult`] that downstream consumers (loader, booking,
//! validate, query, LSP) consume.
//!
//! ## Conversion scope
//!
//! Per-directive converters: Open, Close, Commodity, Note,
//! Document, Event, Query, Price, Balance, Pad, Custom, and
//! Transaction (with its full posting / cost-spec / price-
//! annotation / metadata / trailing-comments machinery).
//!
//! State-only directives (Pushtag / Poptag / Pushmeta / Popmeta)
//! mutate `tag_stack` / `meta_stack` inherited by subsequent
//! directives; mismatched-pop and unclosed-at-EOF emit specific
//! `ParseErrorKind` variants. Arithmetic AMOUNT expressions
//! (`120 / 3 USD` ≡ `40 USD`) are evaluated; the same logic
//! powers numeric values in BALANCE and PRICE directives.
//!
//! Field-level extractors populate `ParseResult.options`,
//! `.includes`, `.plugins`, `.comments`, `.currency_occurrences`,
//! `.account_occurrences`.
//!
//! ## Error surfacing
//!
//! A single [`walk_descendants_once`] pass collects standalone
//! comments, currency occurrences, account occurrences, and inline
//! `ERROR_TOKEN` / mid-file-BOM errors. Specialized extractors run alongside for
//! `ERROR_NODE` classification, transaction body errors, unclosed
//! cost braces, indented top-level directives, and bare-currency
//! values in custom directives.

use rust_decimal::Decimal;
use rustledger_core::cost::{CostNumber, CostSpec};
use rustledger_core::directive::{PriceAnnotation, PriceKind};
use rustledger_core::{
    Account, Amount, Currency, Directive, IncompleteAmount, InternedStr, Link, MetaValue, Metadata,
    NaiveDate, Posting, Span, Spanned, Tag, naive_date,
};

use crate::ParseResult;
use crate::cst::ast::{
    self, AstNode, AstToken, BalanceDirective, CloseDirective, CommodityDirective, CustomDirective,
    DocumentDirective, EventDirective, IncludeDirective, MetaEntry, NoteDirective, OpenDirective,
    OptionDirective, PadDirective, PluginDirective, PostingFlagKind, PriceDirective,
    QueryDirective, SourceFile, Transaction as AstTransaction, TransactionFlagKind,
};

/// Parse Beancount source via the CST and produce the AST-shaped
/// [`ParseResult`]. This is the implementation behind
/// [`crate::parse`]; the public entry delegates here unconditionally.
///
/// See the module-level rustdoc for the conversion scope.
#[must_use]
pub fn parse_via_cst(source: &str) -> ParseResult {
    parse_via_cst_opts(source, /* collect_occurrences = */ true)
}

/// Like [`parse_via_cst`], but only collects `currency_occurrences` /
/// `account_occurrences` when `collect_occurrences` is true.
///
/// Those two indices are consumed **solely by the LSP** (rename / references /
/// highlight). The loader / CLI processing path never reads them, so passing
/// `false` skips the per-`ACCOUNT`/`CURRENCY` `Account::new` / `Currency::new`
/// construction and the per-token in-`ERROR_NODE` ancestor walk inside
/// `walk_descendants_once` — profiling flagged that walk as the #1
/// allocation-count site. Inline errors and top-level comments are still
/// collected unconditionally (the processing path needs them).
#[must_use]
pub fn parse_via_cst_opts(source: &str, collect_occurrences: bool) -> ParseResult {
    parse_via_cst_inner(source, collect_occurrences, /* use_green = */ true)
}

/// Test/fuzz hook: parse like [`crate::parse`] but force the **red** conversion
/// path (green transaction conversion disabled). Used by the `green_eq_red`
/// differential fuzz target to assert the green-wired path is output-equivalent.
#[doc(hidden)]
#[must_use]
pub fn parse_red_only(source: &str) -> ParseResult {
    parse_via_cst_inner(
        source, /* collect_occurrences = */ true, /* use_green = */ false,
    )
}

fn parse_via_cst_inner(source: &str, collect_occurrences: bool, use_green: bool) -> ParseResult {
    // BOM detection mirrors the legacy parser's behavior: strip a
    // leading 3-byte BOM from the source before tokenizing and
    // record its presence in the result. Spans index the original
    // source frame INCLUDING the BOM offset.
    let (stripped, has_leading_bom) = crate::bom::strip_leading(source);
    let bom_offset: u32 = if has_leading_bom { 3 } else { 0 };

    let source_file = SourceFile::parse(stripped);

    let mut directives: Vec<Spanned<Directive>> = Vec::new();
    let mut directive_nodes: Vec<crate::SyntaxNode> = Vec::new();
    let mut options: Vec<(String, String, Span)> = Vec::new();
    let mut includes: Vec<(String, Span)> = Vec::new();
    let mut plugins: Vec<(String, Option<String>, Span)> = Vec::new();
    // Single-pass descendants walk that yields inline errors,
    // top-level comments, and currency occurrences (replaces three
    // separate `descendants_with_tokens` walks at 3·O(N) → 1·O(N)).
    let DescendantsWalkResult {
        inline_errors,
        top_level_comments,
        currency_occurrences,
        account_occurrences,
        cost_brace_errors,
        link_meta_errors,
        custom_pushmeta_errors,
    } = if use_green {
        // Green-tree walk (no per-node red allocation); byte-identical to red.
        super::green::walk_descendants(
            source_file.syntax(),
            stripped,
            bom_offset,
            collect_occurrences,
        )
    } else {
        walk_descendants_once(&source_file, bom_offset, collect_occurrences)
    };

    // Fused single pass over the top-level children replaces the
    // five former per-child traversals (error-node, transaction-body,
    // indented-directive, custom-value diagnostics + section-marker
    // comments). See `walk_top_level_once`.
    let TopLevelWalkResult {
        errors: top_level_errors,
        section_marker_comments,
    } = if use_green {
        super::green::walk_top_level(source_file.syntax(), stripped, bom_offset)
    } else {
        walk_top_level_once(&source_file, stripped, bom_offset)
    };

    let mut comments: Vec<Spanned<String>> = top_level_comments;
    comments.extend(section_marker_comments);
    // Merge in source order; the two helpers' classifiers are
    // disjoint today (STAR-first vs COMMENT-kind-first) but
    // dedup-by-start keeps the invariant local.
    comments.sort_by_key(|s| s.span.start);
    comments.dedup_by_key(|s| s.span.start);
    let mut errors = top_level_errors;
    // Three per-node shape rules — unclosed cost braces, links as metadata
    // values, tags/links as custom/pushmeta values — in a FIXED order that
    // both paths below reproduce.
    //
    // Green folds them into `walk_descendants`, which already visits every
    // node, so they cost a `match` per node and nothing else. Red still runs
    // them as three standalone whole-tree `descendants()` scans, each behind a
    // byte-scan guard (`contains('{')` and friends) that skips the scan when
    // the source cannot contain the construct at all.
    //
    // The guards are what made this cost invisible for so long: they are free
    // on a ledger with no '{', '^' or '#', which is exactly the `simple`
    // profiling shape. Real ledgers have tags and cost specs, and there the
    // three red scans measured 7.46% of all instructions on `tagged` and
    // 2.15% on `investment` (cachegrind ablation, 10k txns). Green — the path
    // every caller but the parity test takes — no longer pays any of it.
    //
    // The green path needs no guards: no '{' in the source means the parser
    // built no COST_SPEC node, so the folded rule finds nothing to report.
    if use_green {
        errors.extend(cost_brace_errors);
        errors.extend(link_meta_errors);
        errors.extend(custom_pushmeta_errors);
    } else {
        if stripped.contains('{') {
            errors.extend(extract_unclosed_cost_brace_errors(
                &source_file,
                stripped,
                bom_offset,
            ));
        }
        if stripped.contains('^') {
            errors.extend(extract_link_metadata_value_errors(&source_file, bom_offset));
        }
        if stripped.contains('^') || stripped.contains('#') {
            errors.extend(extract_custom_pushmeta_taglink_errors(
                &source_file,
                bom_offset,
            ));
        }
    }
    errors.extend(inline_errors);
    let warnings = Vec::new();

    // pushtag/poptag/pushmeta/popmeta state. The legacy parser
    // maintains a stack across directives; each Transaction
    // inherits the active pushed-tag set, and EVERY directive
    // inherits the active pushed-meta set. We pair each entry
    // with the originating directive's span so unclosed-at-EOF
    // diagnostics can point at the offending push.
    let mut tag_stack: Vec<(Tag, Span)> = Vec::new();
    // Vec-of-tuples (NOT a `Metadata` map) so legacy semantics
    // are preserved: `pushmeta x: 1` then `pushmeta x: 2` should
    // shadow (peek returns 2) and `popmeta x` should pop the
    // most recent, leaving x=1 active. A HashMap would have lost
    // the shadowed entry on the second push.
    let mut meta_stack: Vec<(String, MetaValue, Span)> = Vec::new();

    for directive in source_file.directives() {
        // Helper to push a successfully-converted directive
        // alongside its CST node so the post-pass span fixup
        // can index them in parallel.
        let cst_node = directive.syntax().clone();
        // `is_directive_producing` tracks whether THIS arm is
        // expected to emit a `Spanned<Directive>` (the 12
        // directive types). The catch-all below uses it to
        // surface a `SyntaxError` when a producing converter
        // returned `None` without emitting a more specific
        // diagnostic - the silent-drop class of bug the integ
        // tests caught for `2024-01-01 open` (no account),
        // `balance Assets:X` (no amount), etc.
        let is_directive_producing = matches!(
            directive,
            ast::Directive::Open(_)
                | ast::Directive::Close(_)
                | ast::Directive::Commodity(_)
                | ast::Directive::Note(_)
                | ast::Directive::Document(_)
                | ast::Directive::Event(_)
                | ast::Directive::Query(_)
                | ast::Directive::Price(_)
                | ast::Directive::Balance(_)
                | ast::Directive::Pad(_)
                | ast::Directive::Custom(_)
                | ast::Directive::Transaction(_)
        );
        let errors_before = errors.len();
        let pushed_directive = match directive {
            ast::Directive::Open(node) => convert_open(&node, bom_offset, &mut errors),
            ast::Directive::Close(node) => convert_close(&node, bom_offset, &mut errors),
            ast::Directive::Commodity(node) => convert_commodity(&node, bom_offset, &mut errors),
            ast::Directive::Note(node) => convert_note(&node, bom_offset, &mut errors),
            ast::Directive::Document(node) => convert_document(&node, bom_offset, &mut errors),
            ast::Directive::Event(node) => convert_event(&node, bom_offset, &mut errors),
            ast::Directive::Query(node) => convert_query(&node, bom_offset, &mut errors),
            ast::Directive::Price(node) => convert_price(&node, bom_offset, &mut errors),
            ast::Directive::Balance(node) => convert_balance(&node, bom_offset, &mut errors),
            ast::Directive::Pad(node) => convert_pad(&node, bom_offset, &mut errors),
            ast::Directive::Custom(node) => convert_custom(&node, bom_offset, &mut errors),
            ast::Directive::Transaction(node) => {
                // Green-tree conversion (no red-node allocation) with a red
                // fallback for transactions it doesn't yet handle exactly. The
                // green path returns `Some` only when its output is identical to
                // red's, so the hybrid is output-equivalent to the pure-red path.
                let green = node.syntax().green();
                let base =
                    u32::from(node.syntax().text_range().start()) as usize + bom_offset as usize;
                let green_dir = if use_green {
                    super::green::convert_transaction(green, base)
                } else {
                    None
                };
                match green_dir {
                    Some(d) => Some(d),
                    None => convert_transaction(&node, bom_offset, &mut errors),
                }
            }
            ast::Directive::Option(node) => {
                if let Some(triple) = convert_option(&node, bom_offset) {
                    options.push(triple);
                }
                None
            }
            ast::Directive::Include(node) => {
                if let Some(pair) = convert_include(&node, bom_offset) {
                    includes.push(pair);
                }
                None
            }
            ast::Directive::Plugin(node) => {
                if let Some(triple) = convert_plugin(&node, bom_offset) {
                    plugins.push(triple);
                }
                None
            }
            // State-only side effects: mutate the inherited
            // tag/meta sets that apply to subsequent directives.
            ast::Directive::Pushtag(node) => {
                if let Some(tag_token) = node.tag() {
                    let span = node_span(node.syntax(), bom_offset);
                    tag_stack.push((Tag::new(tag_token.text().trim_start_matches('#')), span));
                }
                None
            }
            ast::Directive::Poptag(node) => {
                if let Some(tag_token) = node.tag() {
                    let name = tag_token.text().trim_start_matches('#');
                    if let Some(pos) = tag_stack.iter().rposition(|(t, _)| t.as_str() == name) {
                        tag_stack.remove(pos);
                    } else {
                        errors.push(crate::ParseError::new(
                            crate::ParseErrorKind::InvalidPoptag(name.to_string()),
                            node_span(node.syntax(), bom_offset),
                        ));
                    }
                }
                None
            }
            ast::Directive::Pushmeta(node) => {
                if let Some(key_token) = node.key() {
                    let key = key_token.text_without_colon().to_string();
                    let value = pushmeta_value(node.syntax());
                    let span = node_span(node.syntax(), bom_offset);
                    meta_stack.push((key, value, span));
                }
                None
            }
            ast::Directive::Popmeta(node) => {
                if let Some(key_token) = node.key() {
                    let key = key_token.text_without_colon().to_string();
                    if let Some(pos) = meta_stack.iter().rposition(|(k, _, _)| k == &key) {
                        meta_stack.remove(pos);
                    } else {
                        errors.push(crate::ParseError::new(
                            crate::ParseErrorKind::InvalidPopmeta(key),
                            node_span(node.syntax(), bom_offset),
                        ));
                    }
                }
                None
            }
        };
        if let Some(mut spanned) = pushed_directive {
            apply_inherited_state(&mut spanned.value, &tag_stack, &meta_stack);
            directives.push(spanned);
            directive_nodes.push(cst_node);
        } else if is_directive_producing && errors.len() == errors_before {
            // Producing converter silently dropped the directive
            // (typically: a required field like an account on
            // `open`, an amount on `balance`, or a source account
            // on `pad` was missing). Mirror the legacy parser's
            // top-level error-recovery path which emits a
            // `SyntaxError("unexpected input")` for the failed
            // span so downstream tooling sees the same shape.
            errors.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError("unexpected input".to_string()),
                node_span(&cst_node, bom_offset),
            ));
        }
    }

    // Unclosed pushtag/pushmeta at EOF - legacy emits one error
    // per leftover stack entry, pointing at the originating push
    // directive's span.
    for (tag, span) in &tag_stack {
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::UnclosedPushtag(tag.as_str().to_string()),
            *span,
        ));
    }
    for (key, _, span) in &meta_stack {
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::UnclosedPushmeta(key.clone()),
            *span,
        ));
    }
    errors.sort_by_key(|e| e.span.start);

    // Post-pass: align directive spans with the legacy parser's
    // convention (skip leading trivia, extend through inter-
    // directive trivia to the next directive's start).
    fixup_directive_spans(&source_file, bom_offset, &directive_nodes, &mut directives);

    // Pre-compute the file-wide formatter alignment from the
    // same `source_file` we just walked, so the formatter (and
    // every LSP handler that calls it) can skip the O(N_postings)
    // re-walk on every format request. See
    // `ParseResult::alignment` rustdoc for the cache contract;
    // the equivalence with a fresh `compute_alignment` call is
    // pinned by `parse_result_alignment_cache::*` (lib.rs tests).
    // NOT computed here: `compute_alignment` walks the tree through the red
    // ast accessors, and rowan allocates a `Box<NodeData>` per red node with
    // no recycling. Doing it eagerly charged every parse for a formatter pass
    // it usually never reads — 5.7%-12.3% of instructions by workload. See
    // `ParseResult::alignment`, which computes on first use and caches.
    let alignment = std::sync::OnceLock::new();

    // Capture the green root before we drop `source_file`. `.green()`
    // borrows (`&GreenNodeData`), so promote to an owned `GreenNode`; it is
    // reference-counted internally, cheap to clone, and `Send + Sync` — safe
    // to stash in the `Arc<ParseResult>` the LSP shares across threads.
    //
    // `to_owned()`, not `into_owned()`: rowan 0.17 changed `green()` from
    // returning `Cow<GreenNodeData>` to returning `&GreenNodeData`, so the
    // promotion now goes through `ToOwned` instead of `Cow::into_owned`.
    let syntax_root = source_file.syntax().green().to_owned();

    ParseResult {
        directives,
        options,
        includes,
        plugins,
        comments,
        errors,
        warnings,
        currency_occurrences,
        account_occurrences,
        has_leading_bom,
        syntax_root,
        alignment,
    }
}

// ---- Directive converters --------------------------------------

/// Valid booking methods per beancount v3 - must match the
/// whitelist legacy `parser::parse_open_directive` enforces. An
/// `open` directive whose explicit booking string isn't on this
/// list is rejected (directive dropped, `InvalidBookingMethod`
/// error emitted) by both the legacy parser and `convert_open`.
const VALID_BOOKING_METHODS: &[&str] = &[
    "FIFO",
    "STRICT",
    "STRICT_WITH_SIZE",
    "LIFO",
    "HIFO",
    "NONE",
    "AVERAGE",
];

/// Reject `#tag` / `^link` tokens on a directive that does not take them.
///
/// beancount allows tags and links on TRANSACTIONS, and in v3 on `note` and
/// `document` — nowhere else. rledger accepted them everywhere, silently: each
/// `convert_*` reads the fields it wants and ignores the rest, so a trailing
/// token was never objected to by anything. `2018-06-01 open Assets:A #tag`
/// loaded clean here and is a parse error there (#1949).
///
/// Deliberately scans DIRECT child tokens only. Metadata lives in `META_ENTRY`
/// child NODES, and a metadata VALUE may legitimately be a tag (`k: #x`), so a
/// descendant walk would reject valid input — the opposite mistake, and a worse
/// one.
///
/// Not called from `note` or `document`: both take tags and links in beancount
/// v3 and we already agree with it there. A blanket rule over non-transaction
/// directives would break the two cases that are currently right, which is why
/// this is a per-directive call rather than one check in the dispatcher.
///
/// REPORTS BUT DOES NOT DROP, and that is a deliberate divergence in the error
/// SET (both tools still reject the file). beancount treats this as a parser
/// syntax error, so the directive never exists and every later reference to it
/// cascades:
///
///   2018-06-01 open Assets:N #tag
///   2018-06-02 * "t"
///     Assets:N   1.00 USD
///     ...
///
///   beancount   `ParserSyntaxError` + `ValidationError`: unknown account
///   rledger     the tag error alone; the account is still opened
///
/// Keeping the directive means the user gets one error naming the real
/// problem instead of that error plus a cascade of unopened-account noise
/// pointing at innocent lines. The compat oracle cannot flag the difference,
/// because its error axis compares only WHETHER a file errs and not which
/// errors, so it is written down here rather than left to be rediscovered.
fn reject_tags_and_links(
    node: &crate::SyntaxNode,
    directive: &str,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) {
    use crate::SyntaxKind as K;
    for t in node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
    {
        let kind = t.kind();
        if !matches!(kind, K::TAG | K::LINK) {
            continue;
        }
        let what = if kind == K::TAG { "tag" } else { "link" };
        let range = t.text_range();
        let off = bom_offset as usize;
        let span = Span::new(
            usize::from(range.start()) + off,
            usize::from(range.end()) + off,
        );
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(format!(
                "the {directive} directive does not take a {what} ({}); \
                 tags and links belong to transactions, and to note and \
                 document directives",
                t.text()
            )),
            span,
        ));
    }
}

fn convert_open(
    node: &OpenDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "open", bom_offset, errors);
    let account = Account::new(node.account()?.text());
    let currencies: Vec<Currency> = node.currencies().map(|c| Currency::new(c.text())).collect();
    let booking = node.booking_method().and_then(|s| s.text_decoded());
    let span = node_span(node.syntax(), bom_offset);
    if let Some(b) = &booking
        && !VALID_BOOKING_METHODS.contains(&b.as_str())
    {
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::InvalidBookingMethod(b.clone()),
            span,
        ));
        return None;
    }
    let meta = convert_meta_entries(node.syntax());

    let open = rustledger_core::directive::Open {
        date,
        account,
        currencies,
        booking,
        meta,
    };
    Some(Spanned::new(Directive::Open(open), span))
}

fn convert_close(
    node: &CloseDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "close", bom_offset, errors);
    let account = Account::new(node.account()?.text());
    let meta = convert_meta_entries(node.syntax());

    let close = rustledger_core::directive::Close {
        date,
        account,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Close(close), span))
}

fn convert_commodity(
    node: &CommodityDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "commodity", bom_offset, errors);
    let currency = Currency::new(node.currency()?.text());
    let meta = convert_meta_entries(node.syntax());

    let commodity = rustledger_core::directive::Commodity {
        date,
        currency,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Commodity(commodity), span))
}

/// Collect the `#tag` / `^link` tokens on a directive's HEADER line.
///
/// Stops at the newline that ends the header -- but only after the header has
/// actually started. A directive node can begin with leading trivia: when a
/// blank line precedes it, the node's first token is a NEWLINE, and breaking
/// on that dropped every tag. Both `note` and `document` did this, so a
/// `document` written the way ledgers are normally formatted -- a blank line
/// between directives -- silently lost its tags long before `note` gained any
/// (#2160 review).
///
/// The formatter already had this right: its header walks guard on a
/// `seen_content` flag for the same reason (`cst/format.rs`). The converter
/// was the outlier.
///
/// The header is deemed started once the directive's STRING has been seen:
/// the comment for a `note`, the path for a `document`. Tags and links follow
/// it on the same line, and metadata lines live in `META_ENTRY` child nodes
/// rather than as direct tokens, so they cannot leak in.
fn header_tags_and_links(node: &crate::SyntaxNode) -> (Vec<Tag>, Vec<Link>) {
    let mut tags: Vec<Tag> = Vec::new();
    let mut links: Vec<Link> = Vec::new();
    let mut header_started = false;
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        match t.kind() {
            crate::SyntaxKind::STRING => header_started = true,
            crate::SyntaxKind::NEWLINE if header_started => break,
            crate::SyntaxKind::TAG => {
                tags.push(Tag::new(t.text().trim_start_matches('#')));
            }
            crate::SyntaxKind::LINK => {
                links.push(Link::new(t.text().trim_start_matches('^')));
            }
            _ => {}
        }
    }
    (tags, links)
}

fn convert_note(
    node: &NoteDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    let account = Account::new(node.account()?.text());
    let comment = node.text()?.text_decoded()?;
    // Trailing tags/links on the note header, collected the same way
    // `convert_document` does -- see `header_tags_and_links` for why "the
    // first NEWLINE" is the wrong place to stop.
    //
    // beancount v3 accepts these on a `note`. We parsed them and threw them
    // away, because `Note` had nowhere to put them (#2160) -- note this
    // directive does NOT call `reject_tags_and_links`, unlike `commodity` and
    // `event`, so they were accepted and silently dropped rather than
    // diagnosed.
    let (tags, links) = header_tags_and_links(node.syntax());
    let meta = convert_meta_entries(node.syntax());

    let note = rustledger_core::directive::Note {
        date,
        account,
        comment,
        tags,
        links,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Note(note), span))
}

fn convert_document(
    node: &DocumentDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    let account = Account::new(node.account()?.text());
    let path = node.path()?.text_decoded()?;
    // Trailing tags/links on the document header (legacy
    // `parse_document_directive` collects them in a loop after
    // the path STRING). TAG / LINK tokens only appear in the
    // header, not in the META_ENTRY children walked separately
    // below -- see `header_tags_and_links` for where the walk
    // has to start and stop.
    let (tags, links) = header_tags_and_links(node.syntax());
    let meta = convert_meta_entries(node.syntax());

    let document = rustledger_core::directive::Document {
        date,
        account,
        path,
        tags,
        links,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Document(document), span))
}

fn convert_event(
    node: &EventDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "event", bom_offset, errors);
    let event_type = node.event_type()?.text_decoded()?;
    let value = node.value()?.text_decoded()?;
    let meta = convert_meta_entries(node.syntax());

    let event = rustledger_core::directive::Event {
        date,
        event_type,
        value,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Event(event), span))
}

fn convert_query(
    node: &QueryDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    // `query` was in neither camp: it did not keep tags and links the way a
    // transaction, note or document does, and it did not refuse them the way
    // the other seven directives do. So it took them and dropped them --
    // `Query` has no field to put them in -- while the message the refusal
    // arm emits states the very rule this omission broke.
    //
    // beancount rejects them: `2024-01-01 query "n" "SELECT date" #qtag`
    // gives `syntax error, unexpected TAG, expecting end of file or EOL`
    // (3.2.3), and its `Query` namedtuple is `(meta, date, name,
    // query_string)` with nowhere to keep one either (#2194).
    reject_tags_and_links(node.syntax(), "query", bom_offset, errors);
    let name = node.name()?.text_decoded()?;
    let query = node.query()?.text_decoded()?;
    let meta = convert_meta_entries(node.syntax());

    let q = rustledger_core::directive::Query {
        date,
        name,
        query,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Query(q), span))
}

/// The span of a `balance` / `price` value that holds MORE than the single
/// signed number those directives can represent — extra `NUMBER` tokens, or a
/// misplaced `,`.
///
/// Both directives fall back to "take the first `NUMBER` token and apply a
/// leading sign" when the value is not an arithmetic expression. That fallback
/// reads only the first token, so anything after it was DISCARDED IN SILENCE:
/// `price HOOL 1,23,4.50 USD` stored `1 USD`, a thousandfold error with exit 0.
/// Postings have rejected the same shapes all along (the lexer's grouping regex
/// is strict, and a split number never forms one `NUMBER` token) — this closes
/// the same hole for the directive family (#1892 follow-up).
///
/// Deliberately NOT flagged:
/// - a `~ tolerance` clause, whose second number is legitimate — the scan stops
///   at the `TILDE`;
/// - arithmetic (`0.25 + 0.75 USD`), which is evaluated before this is
///   consulted, so a well-formed expression never reaches it.
fn malformed_directive_value(node: &crate::SyntaxNode) -> Option<crate::TextRange> {
    let mut numbers = 0usize;
    let mut saw_comma = false;
    let mut start: Option<crate::TextRange> = None;
    let mut end: Option<crate::TextRange> = None;
    for t in node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
    {
        match t.kind() {
            // Tolerance begins; its number is not part of the value.
            crate::SyntaxKind::TILDE => break,
            // The trailing currency closes the value. A `price` directive's
            // BASE currency precedes the number, so only break once a number
            // has been seen.
            crate::SyntaxKind::CURRENCY if numbers > 0 => break,
            crate::SyntaxKind::NUMBER => {
                numbers += 1;
                start.get_or_insert(t.text_range());
                end = Some(t.text_range());
            }
            crate::SyntaxKind::COMMA => {
                saw_comma = true;
                start.get_or_insert(t.text_range());
                end = Some(t.text_range());
            }
            _ => {}
        }
    }
    if !saw_comma && numbers <= 1 {
        return None;
    }
    Some(crate::TextRange::new(start?.start(), end?.end()))
}

fn convert_price(
    node: &PriceDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "price", bom_offset, errors);
    let base_currency = Currency::new(node.base_currency()?.text());
    if let Some(range) = has_top_level_tilde(node.syntax()) {
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "a price has no tolerance: `~` is only valid on a balance \
                 assertion"
                    .to_string(),
            ),
            Span::new((start + bom_offset) as usize, (end + bom_offset) as usize),
        ));
        return None;
    }
    // Same arithmetic support as `convert_balance`: a price
    // directive's value can use `+`, `-`, `*`, `/`, and parens.
    let number =
        directive_arithmetic_value(node.syntax(), ToleranceForm::Forbidden).or_else(|| {
            // The fallback keeps only the first NUMBER, so refuse a value that
            // carries more than one — see `malformed_directive_value`.
            if let Some(range) = malformed_directive_value(node.syntax()) {
                let start: u32 = range.start().into();
                let end: u32 = range.end().into();
                let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
                errors.push(crate::ParseError::new(
                    crate::ParseErrorKind::SyntaxError(
                        "malformed amount: expected one number, optionally signed, \
                     or an arithmetic expression. A thousands separator must be \
                     inside the number, as in `-1,234.00`"
                            .to_string(),
                    ),
                    span,
                ));
                return None;
            }
            let mut n = parse_decimal_token(node.number()?.text())?;
            if node_has_minus_before_number(node.syntax()) {
                // Python's rule: negating a zero yields a POSITIVE zero, so a
                // literal `-0.00` loads as `0.00` exactly as beancount parses it.
                // A bare `-n` would keep the sign bit and render `-0.00`.
                n = rustledger_core::negate_python(n);
            }
            Some(n)
        })?;
    let quote_currency = Currency::new(node.quote_currency()?.text());
    let amount = Amount::new(number, quote_currency);
    let meta = convert_meta_entries(node.syntax());

    let price = rustledger_core::directive::Price {
        date,
        currency: base_currency,
        amount,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Price(price), span))
}

fn convert_balance(
    node: &BalanceDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "balance", bom_offset, errors);
    let account = Account::new(node.account()?.text());
    // Beancount accepts arithmetic in the balance assertion's
    // value (`balance Assets:X 0.25 + 0.75 GBP` ≡ 1.00 GBP).
    // Falls back to the first NUMBER token if the expression
    // can't be evaluated, with the legacy sign-flip behavior.
    let number =
        directive_arithmetic_value(node.syntax(), ToleranceForm::Allowed).or_else(|| {
            // The fallback keeps only the first NUMBER, so refuse a value that
            // carries more than one — see `malformed_directive_value`.
            if let Some(range) = malformed_directive_value(node.syntax()) {
                let start: u32 = range.start().into();
                let end: u32 = range.end().into();
                let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
                errors.push(crate::ParseError::new(
                    crate::ParseErrorKind::SyntaxError(
                        "malformed amount: expected one number, optionally signed, \
                     or an arithmetic expression. A thousands separator must be \
                     inside the number, as in `-1,234.00`"
                            .to_string(),
                    ),
                    span,
                ));
                return None;
            }
            let mut n = parse_decimal_token(node.number()?.text())?;
            if node_has_minus_before_number(node.syntax()) {
                // Python's rule: negating a zero yields a POSITIVE zero, so a
                // literal `-0.00` loads as `0.00` exactly as beancount parses it.
                // A bare `-n` would keep the sign bit and render `-0.00`.
                n = rustledger_core::negate_python(n);
            }
            Some(n)
        })?;
    let currency = Currency::new(node.currency()?.text());
    let amount = Amount::new(number, currency);
    let tolerance = check_balance_tolerance(node.syntax(), bom_offset, errors);
    let meta = convert_meta_entries(node.syntax());

    let balance = rustledger_core::directive::Balance {
        date,
        account,
        amount,
        tolerance,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Balance(balance), span))
}

/// Diagnose a tolerance clause that says something the model cannot keep.
///
/// beancount's grammar is `NUMBER ~ NUMBER CURRENCY` — one currency, trailing —
/// and rejects any currency before the `~`. We are deliberately laxer in one
/// direction and stricter in another, on a single rule: **accept what has
/// exactly one meaning, diagnose what has none or contradicts itself.**
///
/// - `1.00 USD ~ 0.01 USD` is ACCEPTED, though beancount calls it a syntax
///   error. The currency is stated twice and agrees, so there is one reading;
///   it canonicalizes to `1.00 ~ 0.01 USD` losslessly. Rejecting it would
///   refuse a file whose meaning is not in doubt.
/// - `1.00 USD ~ 0.01 EUR` is DIAGNOSED. A tolerance denominated in a
///   different currency than the amount is not redundancy, it is an
///   assertion the model has no field for — and `rledger format` used to
///   erase the EUR from the file on its way past.
/// - `1.00 ~ 0.001 0.02 USD` is DIAGNOSED. Two juxtaposed numbers with no
///   operator have no reading at all; this took the first and dropped the
///   rest silently. (`~ 0.005 + 0.005` is arithmetic and still evaluates.)
///
/// Recorded as a deliberate divergence in `docs/reference/compatibility.md`
/// (#2193).
fn check_balance_tolerance(
    node: &crate::SyntaxNode,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Decimal> {
    // Validate AND extract in one walk. Splitting them cost a second full
    // token scan per `balance`, which measured +5.8% on a 56,000-balance
    // ledger -- small in a real ledger, where most directives are
    // transactions, but paid on every balance for nothing.
    //
    // Most balance directives carry no tolerance at all, so answer that
    // first, from an iterator rather than a Vec.
    // Answer "is there a tolerance at all" WITHOUT allocating: on a ledger of
    // balances the overwhelming majority have none, and collecting their
    // tokens first measured +4% against main, which reached the same answer
    // through a `skip_while` that left an empty Vec. The collect below now
    // happens only for the directives that actually carry a `~`.
    if !node
        .children_with_tokens()
        .any(|el| el.kind() == crate::SyntaxKind::TILDE)
    {
        return None;
    }
    let toks: Vec<crate::SyntaxToken> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        .collect();
    let tilde = toks
        .iter()
        .position(|t| t.kind() == crate::SyntaxKind::TILDE)?;
    let (head, tail) = toks.split_at(tilde);
    let tail = &tail[1..];

    let currency_of = |ts: &[crate::SyntaxToken]| {
        ts.iter()
            .find(|t| t.kind() == crate::SyntaxKind::CURRENCY)
            .cloned()
    };
    if let (Some(before), Some(after)) = (currency_of(head), currency_of(tail))
        && before.text() != after.text()
    {
        let range = after.text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(format!(
                "a balance tolerance must be in the same currency as the amount: \
                 the amount is {} and the tolerance is {}. Drop the second \
                 currency, or correct it",
                before.text(),
                after.text(),
            )),
            Span::new((start + bom_offset) as usize, (end + bom_offset) as usize),
        ));
    }

    // A second `~` is its own mistake and deserves its own sentence. Without
    // this it fell into the juxtaposed-numbers arm below, whose message says
    // the numbers are "side by side" when a tilde sits between them --
    // a diagnostic describing input the author did not write.
    if let Some(extra) = tail.iter().find(|t| t.kind() == crate::SyntaxKind::TILDE) {
        let range = extra.text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "a balance takes one tolerance: `AMOUNT ~ TOLERANCE CURRENCY`".to_string(),
            ),
            Span::new((start + bom_offset) as usize, (end + bom_offset) as usize),
        ));
        return None;
    }

    // Numbers in the tolerance region, i.e. before its trailing CURRENCY.
    // Sliced out of `tail` rather than collected: `cost_region_value` takes
    // the same `&[SyntaxToken]`, so the two Vecs this used to build (one of
    // references, one cloning every token back out) bought nothing.
    let end = tail
        .iter()
        .position(|t| t.kind() == crate::SyntaxKind::CURRENCY)
        .unwrap_or(tail.len());
    let region = &tail[..end];
    let numbers = region
        .iter()
        .filter(|t| t.kind() == crate::SyntaxKind::NUMBER)
        .count();

    // A `~` with nothing after it announces a tolerance that is not there.
    // It was accepted as though unwritten, and `rledger format` then deleted
    // the tilde -- the same erase-the-difference behavior as the mismatched
    // currency. beancount calls it a syntax error.
    if numbers == 0 {
        let range = toks[tilde].text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "a `~` must be followed by a tolerance: `AMOUNT ~ TOLERANCE CURRENCY`".to_string(),
            ),
            Span::new((start + bom_offset) as usize, (end + bom_offset) as usize),
        ));
        return None;
    }
    // Stay quiet unless the region is made only of things an expression can
    // contain. On `~ .005 + .005` the lexer already reports the real problem
    // (a number with no integer part, which we reject and beancount accepts)
    // and this would add "the second was being discarded" on top -- a claim
    // about input that never parsed, next to the diagnostic that explains it.
    let expression_shaped = region.iter().all(|t| {
        matches!(
            t.kind(),
            crate::SyntaxKind::NUMBER
                | crate::SyntaxKind::PLUS
                | crate::SyntaxKind::MINUS
                | crate::SyntaxKind::STAR
                | crate::SyntaxKind::SLASH
                | crate::SyntaxKind::L_PAREN
                | crate::SyntaxKind::R_PAREN
        )
    });
    if expression_shaped
        && numbers > 1
        && cost_region_value(region).is_none()
        && let Some(second) = region
            .iter()
            .filter(|t| t.kind() == crate::SyntaxKind::NUMBER)
            .nth(1)
    {
        let range = second.text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "a balance tolerance takes one number, or an arithmetic \
                 expression. Two numbers side by side have no reading, and the \
                 second was being discarded"
                    .to_string(),
            ),
            Span::new((start + bom_offset) as usize, (end + bom_offset) as usize),
        ));
    }

    tolerance_value(region)
}

/// Evaluate a tolerance region: an expression when it is one, else its first
/// NUMBER.
///
/// `10.00 ~ 0.005 * 2 USD` asserts a tolerance of 0.010. Taking the first
/// NUMBER instead truncated it to 0.005 and REJECTED files beancount accepts
/// -- and the E2002 message printed the truncated figure, so the diagnostic
/// advertised the bug (#1944). Same root cause as the cost-spec truncation in
/// #1939: a number-bearing position that never reached the shared evaluator.
fn tolerance_value(region: &[crate::SyntaxToken]) -> Option<Decimal> {
    if region.is_empty() {
        return None;
    }
    if let Some(value) = cost_region_value(region) {
        return Some(value);
    }
    region
        .iter()
        .find(|t| t.kind() == crate::SyntaxKind::NUMBER)
        .and_then(|t| parse_decimal_token(t.text()))
}

fn convert_pad(
    node: &PadDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    reject_tags_and_links(node.syntax(), "pad", bom_offset, errors);
    let account = Account::new(node.target_account()?.text());
    let source_account = Account::new(node.source_account()?.text());
    let meta = convert_meta_entries(node.syntax());

    let pad = rustledger_core::directive::Pad {
        date,
        account,
        source_account,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Pad(pad), span))
}

fn convert_custom(
    node: &CustomDirective,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;
    let custom_type = node.custom_type()?.text_decoded()?;
    let values = extract_custom_values(node.syntax());
    let meta = convert_meta_entries(node.syntax());

    let custom = rustledger_core::directive::Custom {
        date,
        custom_type,
        values,
        meta,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Custom(custom), span))
}

/// Walk the heterogeneous value tokens after the `custom "type"`
/// header. The legacy parser tries each value type in this order:
/// string > account > bool > amount (NUMBER+CURRENCY) > number >
/// date > currency. We replicate that priority on the flat token
/// stream, with one structural pass that pairs an immediately-
/// adjacent NUMBER+CURRENCY into an [`Amount`].
fn extract_custom_values(node: &crate::SyntaxNode) -> Vec<MetaValue> {
    let mut values = Vec::new();
    let mut seen_type_string = false;
    // Collect tokens by kind, skipping trivia. We do a two-pass:
    // first form Amount pairs (NUMBER + CURRENCY adjacent, ignoring
    // whitespace), then emit remaining tokens individually.
    let raw: Vec<rowan::SyntaxToken<crate::BeancountLanguage>> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::WHITESPACE
                    | crate::SyntaxKind::NEWLINE
                    | crate::SyntaxKind::COMMENT
            )
        })
        .collect();

    let mut i = 0;
    while i < raw.len() {
        // Skip the directive's header tokens (DATE, CUSTOM_KW, and
        // the first STRING which is the custom-type name).
        if !seen_type_string {
            if raw[i].kind() == crate::SyntaxKind::STRING {
                seen_type_string = true;
            }
            i += 1;
            continue;
        }
        // One value at a time through the shared discriminator — this is what
        // gives custom directives the same MINUS-sign, Tag/Link and
        // `NUMBER CURRENCY` → Amount handling as metadata entries.
        // ONE advance for both branches, so the loop provably terminates.
        //
        // It used to advance in two places, and only the value branch was
        // guarded. That left the `else` free to move `i` BACKWARD, which does
        // not merely spin: stepping back re-enters the branch above, pushes
        // another value, steps forward, and repeats — allocating without
        // bound. Memory runs out before any per-test timeout can fire, so on
        // CI it takes the whole runner down and surfaces as "the runner has
        // received a shutdown signal", indistinguishable from infrastructure.
        // Two full mutation runs lost the same three parser shards to it
        // (runs 30765289550 and 30768895946) before the cause was found.
        //
        // `value_tokens_to_meta` returns the index past what it consumed, so
        // the assert is the contract and the clamp is the belt: in release a
        // violation costs one wasted token rather than the process.
        let next = if let Some((value, consumed)) = value_tokens_to_meta(&raw, i) {
            // Drop the values the directive is simultaneously being REJECTED
            // for. `value_tokens_to_meta` is shared with metadata, where a
            // `#tag` is legal, but a custom directive accepts none of these --
            // matching beancount, measured across all eight value types.
            //
            // Emitting them anyway made `Custom.values` contain entries the
            // parser had already errored on, so a consumer reading `values`
            // without also reading `errors` saw data the parser had declared
            // invalid (#2190). The errors themselves are raised by
            // `custom_value_check` (bare currency) and
            // `extract_custom_pushmeta_taglink_errors` (tag/link); this keeps
            // the emitted values in step with them.
            //
            // A CURRENCY that pairs with a preceding NUMBER has already become
            // an `Amount` by here, so a surviving `Currency` is exactly the
            // bare one those checks reject. Valid values on the same line are
            // unaffected: `custom "b" "ok" #atag 42` keeps `"ok"` and `42`.
            if !matches!(
                value,
                MetaValue::Tag(_) | MetaValue::Link(_) | MetaValue::Currency(_)
            ) {
                values.push(value);
            }
            consumed
        } else {
            i + 1
        };
        debug_assert!(
            next > i,
            "value_tokens_to_meta must advance: returned {next} at {i}"
        );
        i = next.max(i + 1);
    }
    values
}

fn strip_string_quotes(raw: &str) -> Option<&str> {
    let bytes = raw.as_bytes();
    if bytes.len() < 2 || bytes[0] != b'"' || bytes[bytes.len() - 1] != b'"' {
        return None;
    }
    Some(&raw[1..raw.len() - 1])
}

fn convert_option(node: &OptionDirective, bom_offset: u32) -> Option<(String, String, Span)> {
    let key = node.key()?.text_decoded()?;
    let value = node.value()?.text_decoded()?;
    Some((
        key,
        value,
        single_line_directive_span(node.syntax(), bom_offset),
    ))
}

fn convert_include(node: &IncludeDirective, bom_offset: u32) -> Option<(String, Span)> {
    let path = node.path()?.text_decoded()?;
    Some((path, single_line_directive_span(node.syntax(), bom_offset)))
}

fn convert_plugin(
    node: &PluginDirective,
    bom_offset: u32,
) -> Option<(String, Option<String>, Span)> {
    let module = node.module()?.text_decoded()?;
    let config = node.config().and_then(|c| c.text_decoded());
    Some((
        module,
        config,
        single_line_directive_span(node.syntax(), bom_offset),
    ))
}

// ---- Transaction + Posting + sub-nodes -------------------------

fn convert_transaction(
    node: &AstTransaction,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Directive>> {
    let date = parse_directive_date(&node.date()?, errors, bom_offset)?;

    // Flag: explicit (TransactionFlag) or implied (leading STRING
    // with no flag token; defaults to '*').
    let flag = node.flag().map_or('*', |f| flag_char_from_transaction(&f));

    // Header strings, consumed straight off the iterator (no intermediate Vec,
    // no count-then-unwrap): 0 -> empty narration; 1 -> narration only;
    // 2 -> payee + narration; 3+ -> surface only the last as narration (the
    // middles are unreachable through this typed shape).
    let mut it = node.strings().filter_map(|s| s.text_decoded());
    let (payee_str, narration_str) = match (it.next(), it.next(), it.next()) {
        (None, _, _) => (None, String::new()),
        (Some(n), None, _) => (None, n),
        (Some(p), Some(n), None) => (Some(p), n),
        // 3+: `c` is the 3rd string; if more follow, `it.last()` is the actual
        // last (else it falls back to `c`). No clone in the common 0/1/2 cases.
        (Some(_), Some(_), Some(c)) => (None, it.last().unwrap_or(c)),
    };

    let payee = payee_str.map(InternedStr::from);
    let narration = InternedStr::from(narration_str);

    // Tags / links from the TRANSACTION node: the typed AST
    // accessor `tags()`/`links()` is scoped to the header region.
    // Trailing TAG / LINK tokens appearing on body lines (after
    // the header NEWLINE, OUTSIDE any POSTING / META_ENTRY child
    // node) are also part of the transaction's tag/link set per
    // Beancount semantics - `extract_transaction_body_errors`
    // already exempts them from the malformed-body diagnostic for
    // this reason. Aggregate them here so they don't silently
    // disappear.
    let mut tags: Vec<Tag> = node
        .tags()
        .map(|t| Tag::new(t.text().trim_start_matches('#')))
        .collect();
    let mut links: Vec<Link> = node
        .links()
        .map(|l| Link::new(l.text().trim_start_matches('^')))
        .collect();
    for el in node.syntax().children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            // Nodes (POSTING / META_ENTRY) own their own internal
            // tokens; we don't recurse into them.
            continue;
        };
        match t.kind() {
            crate::SyntaxKind::TAG => {
                let stripped = t.text().trim_start_matches('#');
                let new_tag = Tag::new(stripped);
                if !tags.contains(&new_tag) {
                    tags.push(new_tag);
                }
            }
            crate::SyntaxKind::LINK => {
                let stripped = t.text().trim_start_matches('^');
                let new_link = Link::new(stripped);
                if !links.contains(&new_link) {
                    links.push(new_link);
                }
            }
            _ => {}
        }
    }

    // Transaction-level metadata (META_ENTRY children directly on
    // the TRANSACTION node, NOT on POSTING children).
    let meta = convert_meta_entries(node.syntax());

    // Postings + pre-posting comments. The CST puts inter-
    // posting trivia (including `; comment` lines) as flat
    // tokens DIRECT under TRANSACTION between two POSTING
    // nodes. Walk in source order: COMMENT tokens accumulate
    // into `pending`, then attach to the next POSTING node's
    // `comments` field when we reach it. Tokens before the
    // header NEWLINE are skipped (they're transaction-header
    // content). Comments that remain in `pending` after the
    // final posting belong to the transaction itself
    // (legacy: `txn.trailing_comments = pending_comments`).
    let (postings, trailing_comments) = collect_postings_with_comments(node, bom_offset, errors);

    // Deprecated `|` separator between payee and narration: a
    // PIPE token in the header region. Legacy treats this as a
    // recoverable warning-shaped error (`DeprecatedPipeSymbol`)
    // and keeps the directive, so we do the same here.
    if header_has_pipe(node) {
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::DeprecatedPipeSymbol,
            node_span(node.syntax(), bom_offset),
        ));
    }

    let txn = rustledger_core::directive::Transaction {
        date,
        flag,
        payee,
        narration,
        tags,
        links,
        meta,
        postings,
        trailing_comments,
    };
    let span = node_span(node.syntax(), bom_offset);
    Some(Spanned::new(Directive::Transaction(txn), span))
}

/// Returns true if the TRANSACTION header (direct-child tokens
/// up to the first NEWLINE) contains a `PIPE` token. The legacy
/// parser surfaces a `DeprecatedPipeSymbol` diagnostic for this
/// shape; the CST lexer classifies `|` as `PIPE`, so we just
/// scan the header directly.
fn header_has_pipe(node: &AstTransaction) -> bool {
    for el in node.syntax().children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        if t.kind() == crate::SyntaxKind::NEWLINE {
            return false;
        }
        if t.kind() == crate::SyntaxKind::PIPE {
            return true;
        }
    }
    false
}

/// Walk a `TRANSACTION`'s children in source order, attaching any
/// inter-posting `; comment` lines that appear as flat tokens
/// between `POSTING` nodes to the NEXT posting's `comments`
/// field. Matches the legacy parser, which collects
/// `pending_comments` while reading the body and applies them to
/// the next posting it parses.
///
/// Tokens before the header-terminator NEWLINE belong to the
/// transaction header (date/flag/strings/tags/links) and are
/// skipped.
///
/// Returns `(postings, trailing_comments)`: the second element is
/// any pending comments left over AFTER the final posting, which
/// legacy assigns to `Transaction::trailing_comments`.
fn collect_postings_with_comments(
    node: &AstTransaction,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> (Vec<Spanned<Posting>>, Vec<String>) {
    let mut out = Vec::new();
    let mut pending: Vec<String> = Vec::new();
    let mut past_header = false;
    for el in node.syntax().children_with_tokens() {
        match el {
            rowan::NodeOrToken::Token(t) => {
                if !past_header {
                    if t.kind() == crate::SyntaxKind::NEWLINE {
                        past_header = true;
                    }
                    continue;
                }
                if is_comment_kind(t.kind()) {
                    pending.push(t.text().to_string());
                } else if !is_trivia_kind(t.kind())
                    && !matches!(t.kind(), crate::SyntaxKind::TAG | crate::SyntaxKind::LINK)
                {
                    // Non-trivia, non-comment token in the
                    // transaction body that's NOT inside a
                    // POSTING / META_ENTRY child node = malformed
                    // body line (caught separately by
                    // `extract_transaction_body_errors`). Treat
                    // the same as a failed POSTING: clear pending
                    // so the malformed line's preceding comments
                    // don't migrate onto the next valid posting.
                    //
                    // EXEMPT TAG / LINK: trailing tags/links on
                    // transaction body lines (after the header)
                    // are valid Beancount - they extend the
                    // transaction's tag/link set without being
                    // a new posting. Treating them as malformed
                    // would drop legitimate preceding comments
                    // that belong to the NEXT posting. The same
                    // exemption appears in
                    // `extract_transaction_body_errors`, which
                    // does the parallel "is this a malformed
                    // body line?" classification.
                    pending.clear();
                }
            }
            rowan::NodeOrToken::Node(n) => {
                if !past_header {
                    // META_ENTRY or POSTING before the header
                    // NEWLINE shouldn't happen in well-formed
                    // input; treat any child node as "past the
                    // header" if we somehow encounter one.
                    past_header = true;
                }
                if let Some(p) = ast::Posting::cast(n) {
                    if let Some(mut spanned) = convert_posting(&p, bom_offset, errors) {
                        if !pending.is_empty() {
                            spanned.value.comments = std::mem::take(&mut pending);
                        }
                        out.push(spanned);
                    } else {
                        // Failed posting consumes any pending
                        // inter-posting comments - they belonged
                        // to it. Without this clear, a malformed
                        // posting's preceding comments would
                        // migrate forward and attach to the NEXT
                        // successful posting, misattributing them
                        // visibly to the wrong account line.
                        pending.clear();
                    }
                }
                // META_ENTRY child nodes: comments collected so
                // far don't apply to them (they're transaction
                // metadata). Drop them.
            }
        }
    }
    (out, pending)
}

fn flag_char_from_transaction(flag: &ast::TransactionFlag) -> char {
    match flag.classify() {
        TransactionFlagKind::Star | TransactionFlagKind::Txn => '*',
        TransactionFlagKind::Pending => '!',
        TransactionFlagKind::Hash => '#',
        TransactionFlagKind::Letter | TransactionFlagKind::CurrencyLetter => {
            flag.text().chars().next().unwrap_or('*')
        }
    }
}

fn convert_posting(
    node: &ast::Posting,
    bom_offset: u32,
    errors: &mut Vec<crate::ParseError>,
) -> Option<Spanned<Posting>> {
    let account = Account::new(node.account()?.text());

    let flag = node.flag().map(|f| flag_char_from_posting(&f));

    // A well-formed posting has AT MOST one `AMOUNT` child node
    // (the units). The CST builder will accept input like
    // `Expenses:Food  5 USD + 3 USD` and produce TWO sibling
    // `AMOUNT` nodes joined by a flat PLUS token, because the
    // grammar doesn't enforce that PLUS between two complete
    // amounts is invalid. `Posting::amount()` returns only the
    // first via `first_child`, so without this guard the second
    // amount (and the joining `+`) would be silently dropped and
    // the user's transaction would balance against the wrong
    // number. Emit a `SyntaxError` pointing at the trailing
    // siblings and keep the first amount.
    let mut amount_children = node
        .syntax()
        .children()
        .filter(|n| ast::Amount::can_cast(n.kind()));
    let first_amount = amount_children.next();
    let first_amount_end: Option<u32> = first_amount.as_ref().map(|n| n.text_range().end().into());
    let mut sibling_start: Option<u32> = None;
    let mut sibling_end: u32 = 0;
    for extra in amount_children {
        let range = extra.text_range();
        let start_u32: u32 = range.start().into();
        let end_u32: u32 = range.end().into();
        if sibling_start.is_none() {
            sibling_start = Some(start_u32);
        }
        sibling_end = end_u32;
    }
    if let Some(start_u32) = sibling_start {
        // Extend the span back to the end of the FIRST AMOUNT so
        // the diagnostic underline covers any joining operator
        // (`+`, `*`, whitespace) between the kept amount and the
        // orphans. Without this, a user sees only `3 USD` in
        // `5 USD + 3 USD` highlighted - and may not realize the
        // `+ 3 USD` together is what needs to be removed.
        let underline_start = first_amount_end.unwrap_or(start_u32);
        let span = Span::new(
            (underline_start + bom_offset) as usize,
            (sibling_end + bom_offset) as usize,
        );
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "unexpected trailing tokens after posting amount".to_string(),
            ),
            span,
        ));
    }
    // The mirror of the guard above, for tokens dropped BEFORE the amount
    // rather than after it.
    //
    // `starts_amount` only opens an `AMOUNT` at `NUMBER`, `CURRENCY`,
    // `L_PAREN`, or a sign directly followed by one of those. Anything else
    // becomes a flat `POSTING` child, and nothing downstream reads flat
    // children — so `-,123.00 USD` parsed as `POSTING(MINUS COMMA
    // AMOUNT(NUMBER CURRENCY))` and booked as **+123.00**, silently losing the
    // sign. A stray comma is the way to hit this in practice, because a
    // thousands separator only belongs INSIDE a `NUMBER` token (the lexer's
    // grouping regex is strict) and a misplaced one splits the amount in two.
    //
    // Report rather than repair: `,123` and `-,123` have no agreed meaning, so
    // guessing one would be inventing data. See issue #1892's discussion.
    if let Some(range) = orphaned_amount_prefix(node.syntax()) {
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
        errors.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "unexpected token before posting amount: a `+`/`-` must be \
                 followed by a number, and a thousands separator must be \
                 inside one (as in `-1,234.00`)"
                    .to_string(),
            ),
            span,
        ));
    }

    let units = first_amount
        .and_then(ast::Amount::cast)
        .and_then(|amt| convert_amount_to_incomplete(&amt, errors, bom_offset));
    let cost = node.cost_spec().map(|cs| convert_cost_spec(&cs));
    let price = node
        .price_annotation()
        .map(|pa| convert_price_annotation(&pa, errors, bom_offset));
    let meta = convert_meta_entries(node.syntax());

    // Trailing comments on the posting line: COMMENT direct-
    // child tokens BEFORE the terminator NEWLINE. The legacy
    // parser collects same-line `;` content into
    // `posting.trailing_comments`.
    let trailing_comments: Vec<String> = node
        .syntax()
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .take_while(|t| t.kind() != crate::SyntaxKind::NEWLINE)
        .filter(|t| is_comment_kind(t.kind()))
        .map(|t| t.text().to_string())
        .collect();

    let posting = Posting {
        account,
        units,
        cost: cost.map(Box::new),
        price: price.map(Box::new),
        flag,
        meta,
        comments: Vec::new(),
        trailing_comments,
    };
    let span = posting_span(node.syntax(), bom_offset);
    Some(Spanned::new(posting, span))
}

fn flag_char_from_posting(flag: &ast::PostingFlag) -> char {
    match flag.classify() {
        PostingFlagKind::Star => '*',
        PostingFlagKind::Pending => '!',
        PostingFlagKind::Hash => '#',
        PostingFlagKind::Letter | PostingFlagKind::CurrencyLetter => {
            flag.text().chars().next().unwrap_or('*')
        }
    }
}

/// Convert an AMOUNT node into an [`IncompleteAmount`]. Returns
/// `None` if neither a number nor a currency is present (which
/// shouldn't happen for a well-formed AMOUNT, but matches the
/// lossless CST contract). Sign is folded into the number.
///
/// **Arithmetic limitation**: when the AMOUNT contains an
/// arithmetic expression (`100+5 USD`), only the FIRST `NUMBER`
/// is used. A proper expression evaluator is deferred - none of
/// the directive types we currently handle outside of postings
/// use AMOUNT shapes that the legacy parser would have evaluated
/// differently.
fn convert_amount_to_incomplete(
    amt: &ast::Amount,
    errors: &mut Vec<crate::ParseError>,
    bom_offset: u32,
) -> Option<IncompleteAmount> {
    // Arithmetic AMOUNT expressions (`120 / 3 USD`, `(1+2) USD`):
    // run the recursive-descent evaluator on the flat token
    // stream. Fast-path plain `NUMBER CURRENCY` shapes to keep
    // the common case allocation-free.
    let number = if amt.is_arithmetic() {
        let evaluated = evaluate_amount_expression(amt);
        if evaluated.is_none() {
            // `is_arithmetic` was true but the evaluator gave up
            // (decimal overflow, division by zero, malformed
            // expression, unbalanced parens). Without this
            // emission the amount silently degrades to
            // `CurrencyOnly` and the user only sees a downstream
            // "transaction doesn't balance" - masking the actual
            // root cause. Pin the span to the AMOUNT node so the
            // diagnostic underlines the offending expression.
            let range = amt.syntax().text_range();
            let start: u32 = range.start().into();
            let end: u32 = range.end().into();
            let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
            errors.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError(
                    "invalid arithmetic expression in amount (overflow, division by zero, or malformed)"
                        .to_string(),
                ),
                span,
            ));
        }
        evaluated
    } else {
        amt.number().and_then(|n| {
            let parsed = parse_decimal_token(n.text());
            if parsed.is_none() {
                // Symmetry with the arithmetic-failure path: when
                // a plain NUMBER token in an AMOUNT can't be
                // turned into a Decimal (e.g., 30+ digits - the
                // lexer's NUMBER regex has no max length but
                // `rust_decimal`'s 28-digit ceiling rejects it),
                // surface a diagnostic instead of silently
                // degrading to `CurrencyOnly`. Without this the
                // user only sees "transaction doesn't balance"
                // and never learns the parser dropped a number.
                let range = n.syntax().text_range();
                let start: u32 = range.start().into();
                let end: u32 = range.end().into();
                let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
                errors.push(crate::ParseError::new(
                    crate::ParseErrorKind::SyntaxError(
                        "invalid number in amount (likely exceeds 28-digit Decimal precision)"
                            .to_string(),
                    ),
                    span,
                ));
            }
            let mut value = parsed?;
            if let Some(sign) = amt.sign()
                && sign.is_minus()
            {
                // See the sign note above — `negate_python` keeps a zero
                // unsigned, matching beancount's parse of `-0.00`.
                value = rustledger_core::negate_python(value);
            }
            Some(value)
        })
    };
    let currency = amt.currency().map(|c| Currency::new(c.text()));
    match (number, currency) {
        (Some(n), Some(c)) => Some(IncompleteAmount::Complete(Amount::new(n, c))),
        (Some(n), None) => Some(IncompleteAmount::NumberOnly(n)),
        (None, Some(c)) => Some(IncompleteAmount::CurrencyOnly(c)),
        (None, None) => None,
    }
}

/// Evaluate the arithmetic expression inside an `AMOUNT` node and
/// return the resulting decimal. Returns `None` when evaluation
/// fails (division by zero, decimal overflow, malformed parens,
/// missing operand).
///
/// AMOUNT children are flat tokens (no expression sub-tree): a
/// sequence of `NUMBER`, `PLUS`, `MINUS`, `STAR`, `SLASH`,
/// `L_PAREN`, `R_PAREN`, and a trailing `CURRENCY` at depth 0
/// that's the amount's currency rather than part of the
/// expression. The currency is stripped first; the rest goes
/// through recursive descent mirroring legacy
/// `parser::parse_expr` / `parse_term` / `parse_primary`.
///
/// Operator precedence and unary handling match Python beancount:
/// `*` and `/` bind tighter than `+` and `-`; a leading or post-
/// operator `-` is unary negation.
fn evaluate_amount_expression(amt: &ast::Amount) -> Option<Decimal> {
    let tokens = amount_expression_tokens(amt);
    let mut cursor = 0usize;
    let value = parse_arith_expr(&tokens, &mut cursor)?;
    // Trailing tokens after a successful parse mean the expression
    // is malformed (`1+2 3 USD`); refuse rather than silently
    // dropping them.
    if cursor != tokens.len() {
        return None;
    }
    Some(value)
}

/// Evaluate the arithmetic expression that appears as the
/// numeric value of a `BALANCE` / `PRICE` directive, returning
/// the resulting decimal or `None` if not arithmetic (single
/// NUMBER, callers fall back to `parse_decimal_token`).
///
/// Unlike `AMOUNT`, these directives don't wrap their value in
/// a dedicated node - the tokens are flat under the directive
/// node. The relevant region is from the FIRST `NUMBER` token up
/// to (but not including) the FIRST `CURRENCY` token at paren-
/// depth 0 (the amount currency). For BALANCE, this correctly
/// stops before any trailing `~ NUMBER [CURRENCY]` tolerance
/// region too.
///
/// Returns `Some` only when the slice contains at least one
/// arithmetic operator (`+`, `-`, `*`, `/`) or parens - for a
/// bare single `NUMBER`, returns `None` so the caller can use
/// the existing fast path (which preserves the legacy sign-flip
/// behavior).
/// Does this directive's header carry a top-level `~`?
///
/// Only `balance` has a tolerance. A `price` with one used to be accepted
/// silently -- `price USD 1.10 ~ 0.01 EUR` evaluated to 1.10 and dropped the
/// rest on the floor, because `~` is not an arithmetic operator, so the
/// expression scan took the fast path and fell back to the first NUMBER. A
/// value the author wrote and the parser discarded without a word is worse
/// than a diagnosed error.
fn has_top_level_tilde(node: &crate::SyntaxNode) -> Option<crate::TextRange> {
    let mut depth: i32 = 0;
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            // Skip child nodes rather than stopping at one. A price's only
            // child node is a META_ENTRY, whose own tokens are not direct
            // children of this node, so nothing here can see a `~` written
            // inside metadata -- but bailing out at the first node would mean
            // a header that ever gains one stops being checked at all.
            continue;
        };
        match t.kind() {
            crate::SyntaxKind::L_PAREN => depth += 1,
            crate::SyntaxKind::R_PAREN => depth -= 1,
            crate::SyntaxKind::TILDE if depth == 0 => return Some(t.text_range()),
            _ => {}
        }
    }
    None
}

/// Whether a `~` may end this directive's amount expression.
///
/// `balance` takes a tolerance; `price` does not. The distinction matters
/// because stopping at `~` unconditionally makes a price with a tolerance
/// evaluate its amount and DISCARD the tolerance silently, which is worse
/// than the malformed-value error it used to get.
#[derive(Clone, Copy, PartialEq, Eq)]
enum ToleranceForm {
    /// `balance ACCOUNT AMOUNT ~ TOLERANCE CURRENCY`
    Allowed,
    /// No tolerance in this directive's grammar.
    Forbidden,
}

fn directive_arithmetic_value(
    node: &crate::SyntaxNode,
    tolerance: ToleranceForm,
) -> Option<Decimal> {
    let raw: Vec<crate::SyntaxToken> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        // Skip the directive header (DATE, keyword, ACCOUNT / base CURRENCY)
        // and stop at the first token that can BEGIN a value.
        //
        // This used to skip to the first `NUMBER`, which also swallowed a
        // leading `(`: `(1 + 5) / 2.1 USD` became `1 + 5 ) / 2.1`, failed to
        // parse, and fell through to "take the first NUMBER" — so the
        // directive silently asserted against **1**. A leading sign was
        // likewise dropped and re-applied by the caller. Neither header token
        // can be a `NUMBER`, `L_PAREN` or a sign, so this cannot over-skip.
        .skip_while(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::NUMBER
                    | crate::SyntaxKind::L_PAREN
                    | crate::SyntaxKind::MINUS
                    | crate::SyntaxKind::PLUS
            )
        })
        .collect();
    let mut depth: i32 = 0;
    let mut first_currency_idx: Option<usize> = None;
    let mut tilde_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,
            crate::SyntaxKind::CURRENCY if depth == 0 && first_currency_idx.is_none() => {
                first_currency_idx = Some(i);
            }
            // `~` ends the amount and begins the tolerance, which is read
            // separately. Without this the tolerance was swallowed INTO the
            // expression: `0.25 + 0.75 ~ 0.01 USD` became `0.25 + 0.75 ~ 0.01`,
            // which is not arithmetic, so the whole value was reported
            // malformed. Only a currency-less amount reached that state --
            // `0.25 + 0.75 USD ~ 0.01 USD` stopped at the first CURRENCY and
            // parsed -- which is why the plain-number form (`100.00 ~ 0.01
            // USD`) worked and only its arithmetic sibling did not (#2191).
            crate::SyntaxKind::TILDE
                if tolerance == ToleranceForm::Allowed && depth == 0 && tilde_idx.is_none() =>
            {
                tilde_idx = Some(i);
            }
            _ => {}
        }
    }
    let end = match (first_currency_idx, tilde_idx) {
        (Some(c), Some(t)) => c.min(t),
        (Some(i), None) | (None, Some(i)) => i,
        (None, None) => raw.len(),
    };
    let tokens: Vec<crate::SyntaxToken> = raw.into_iter().take(end).collect();
    // Fast-path: zero or one token = no arithmetic.
    let has_op = tokens.iter().any(|t| {
        matches!(
            t.kind(),
            crate::SyntaxKind::PLUS
                | crate::SyntaxKind::MINUS
                | crate::SyntaxKind::STAR
                | crate::SyntaxKind::SLASH
                | crate::SyntaxKind::L_PAREN
        )
    });
    if !has_op {
        return None;
    }
    let mut cursor = 0usize;
    let value = parse_arith_expr(&tokens, &mut cursor)?;
    if cursor != tokens.len() {
        return None;
    }
    Some(value)
}

/// Collect AMOUNT's expression tokens - every non-trivia direct-
/// child token EXCEPT the trailing `CURRENCY` at paren-depth 0
/// (which is the amount's currency, not part of the expression).
/// Parens at any depth are preserved so `parse_arith_primary` can
/// recurse through them.
fn amount_expression_tokens(amt: &ast::Amount) -> Vec<crate::SyntaxToken> {
    let raw: Vec<crate::SyntaxToken> = amt
        .syntax()
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| !is_trivia_kind(t.kind()))
        .collect();
    // Find the index of the LAST `CURRENCY` at depth 0 - same
    // disambiguator as `Amount::currency()`. Tokens before that
    // index form the arithmetic expression.
    let mut depth: i32 = 0;
    let mut trailing_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,
            crate::SyntaxKind::CURRENCY if depth == 0 => trailing_currency_idx = Some(i),
            _ => {}
        }
    }
    let end = trailing_currency_idx.unwrap_or(raw.len());
    raw.into_iter().take(end).collect()
}

/// `expr := term (('+' | '-') term)*` - left-associative.
fn parse_arith_expr<T: TokenView>(tokens: &[T], cursor: &mut usize) -> Option<Decimal> {
    let mut result = parse_arith_term(tokens, cursor)?;
    while let Some(op) = tokens.get(*cursor).map(TokenView::kind) {
        match op {
            crate::SyntaxKind::PLUS => {
                *cursor += 1;
                let rhs = parse_arith_term(tokens, cursor)?;
                result = result.checked_add(rhs)?;
            }
            crate::SyntaxKind::MINUS => {
                *cursor += 1;
                let rhs = parse_arith_term(tokens, cursor)?;
                result = result.checked_sub(rhs)?;
            }
            _ => break,
        }
    }
    Some(result)
}

/// `term := primary (('*' | '/') primary)*` - left-associative.
fn parse_arith_term<T: TokenView>(tokens: &[T], cursor: &mut usize) -> Option<Decimal> {
    let mut result = parse_arith_primary(tokens, cursor)?;
    while let Some(op) = tokens.get(*cursor).map(TokenView::kind) {
        match op {
            crate::SyntaxKind::STAR => {
                *cursor += 1;
                let rhs = parse_arith_primary(tokens, cursor)?;
                result = result.checked_mul(rhs)?;
            }
            crate::SyntaxKind::SLASH => {
                *cursor += 1;
                let rhs = parse_arith_primary(tokens, cursor)?;
                if rhs.is_zero() {
                    return None;
                }
                result = result.checked_div(rhs)?;
            }
            _ => break,
        }
    }
    Some(result)
}

/// `primary := '(' expr ')' | '-' primary | '+' primary | NUMBER`.
fn parse_arith_primary<T: TokenView>(tokens: &[T], cursor: &mut usize) -> Option<Decimal> {
    let t = tokens.get(*cursor)?;
    match t.kind() {
        crate::SyntaxKind::L_PAREN => {
            *cursor += 1;
            let inner = parse_arith_expr(tokens, cursor)?;
            // Mandatory closer; bail (returning None) on unbalance
            // - `Amount::currency()` already refuses to surface a
            // currency for unbalanced parens, so the amount as a
            // whole degrades cleanly to `NumberOnly`/`None`.
            let close = tokens.get(*cursor)?;
            if close.kind() != crate::SyntaxKind::R_PAREN {
                return None;
            }
            *cursor += 1;
            Some(inner)
        }
        crate::SyntaxKind::MINUS => {
            *cursor += 1;
            let inner = parse_arith_primary(tokens, cursor)?;
            Some(-inner)
        }
        crate::SyntaxKind::PLUS => {
            *cursor += 1;
            parse_arith_primary(tokens, cursor)
        }
        crate::SyntaxKind::NUMBER => {
            let value = parse_decimal_token(t.text())?;
            *cursor += 1;
            Some(value)
        }
        _ => None,
    }
}

/// The slice of a cost-spec segment holding its NUMBER expression: from the
/// first token that can begin a value up to the first `CURRENCY` at paren
/// depth 0.
///
/// Depth matters because a currency cannot appear inside the parens of a
/// numeric expression, but stopping at the first `CURRENCY` unconditionally
/// would be wrong the moment one ever could.
fn cost_number_region<T: TokenView>(seg: &[T]) -> &[T] {
    use crate::SyntaxKind as K;
    let start = seg
        .iter()
        .position(|t| matches!(t.kind(), K::NUMBER | K::L_PAREN | K::MINUS | K::PLUS))
        .unwrap_or(seg.len());
    let mut depth = 0i32;
    let mut end = seg.len();
    for (i, t) in seg.iter().enumerate().skip(start) {
        match t.kind() {
            K::L_PAREN => depth += 1,
            K::R_PAREN => depth -= 1,
            K::CURRENCY if depth == 0 => {
                end = i;
                break;
            }
            // A comma ends the number region too (`{10 USD, 2014-02-25}`);
            // without this a malformed spec could drag the date into the
            // expression and fail the whole parse instead of just the number.
            K::COMMA if depth == 0 => {
                end = i;
                break;
            }
            _ => {}
        }
    }
    &seg[start..end]
}

/// Evaluate a cost-spec number region, but ONLY when it is genuinely
/// arithmetic.
///
/// Returns `None` for an UNSIGNED bare `NUMBER` so the caller keeps its
/// existing single-token latch, leaving the overwhelmingly common case on the
/// allocation-free path.
///
/// A SIGNED number is not in that set: `MINUS`/`PLUS` counts as an operator, so
/// `{-200.00 USD}` routes through the evaluator. That is deliberate and is
/// itself a fix — the latch only ever read `NUMBER` tokens and never applied a
/// leading sign, so a negative cost was booked as POSITIVE. An earlier draft of
/// this comment claimed the latch "carries legacy sign handling"; it does not,
/// and the two `TotalsAndSigns` corpus fixtures prove it.
///
/// Why this exists at all: the price path has always evaluated expressions
/// (`convert_amount_to_incomplete` -> `evaluate_amount_expression`) while the
/// cost path latched the first `NUMBER` token and silently dropped the rest, so
/// `{10.00 * 3 USD}` booked a cost of 10.00. Same computation, two
/// implementations, nothing asserting agreement — the exact drift shape
/// CLAUDE.md's Canonical-Function Discipline describes. The fix is to share the
/// evaluator, not to teach this path its own arithmetic (#1939).
fn cost_region_value<T: TokenView>(seg: &[T]) -> Option<Decimal> {
    use crate::SyntaxKind as K;
    // Trivia first. `cost_spec_from_tokens` is handed EVERY child token,
    // whitespace included, and the evaluator refuses any token it does not
    // recognize — so without this the region is `10.00 WS * WS 3` and every
    // expression silently falls back to the latched first number, i.e. the bug
    // this is meant to fix, still there but now with more code. The other two
    // evaluators (`directive_arithmetic_value`, `amount_expression_tokens`)
    // both filter trivia for the same reason.
    let kept: Vec<&T> = seg.iter().filter(|t| !is_trivia_kind(t.kind())).collect();
    let region = cost_number_region(&kept);
    if !region.iter().any(|t| {
        matches!(
            t.kind(),
            K::PLUS | K::MINUS | K::STAR | K::SLASH | K::L_PAREN
        )
    }) {
        return None;
    }
    let mut cursor = 0usize;
    let value = parse_arith_expr(region, &mut cursor)?;
    // Trailing tokens mean the expression is malformed; refuse rather than
    // silently dropping them, matching `evaluate_amount_expression`.
    if cursor != region.len() {
        return None;
    }
    Some(value)
}

fn convert_cost_spec(cs: &ast::CostSpec) -> CostSpec {
    cost_spec_from_tokens(
        cs.syntax()
            .children_with_tokens()
            .filter_map(rowan::NodeOrToken::into_token),
    )
}

/// Minimal view of a lexed token — kind + text — implemented by both tree
/// walkers (red `SyntaxToken`, green `&GreenTokenData`) so the token-level
/// semantic helpers ([`cost_spec_from_tokens`], [`meta_value_from_tokens`])
/// are SHARED rather than hand-mirrored. Every historical green/red fuzz
/// divergence (#1704, #1713, the `{*}` merge flag) landed in a hand-mirrored
/// copy of these semantics; with one implementation the class is gone.
pub(super) trait TokenView {
    /// The token's [`crate::SyntaxKind`].
    fn kind(&self) -> crate::SyntaxKind;
    /// The token's source text.
    fn text(&self) -> &str;
}

impl<T: TokenView> TokenView for &T {
    fn kind(&self) -> crate::SyntaxKind {
        (*self).kind()
    }
    fn text(&self) -> &str {
        (*self).text()
    }
}

impl TokenView for rowan::SyntaxToken<crate::BeancountLanguage> {
    // `Self::kind`/`Self::text` resolve to the INHERENT `SyntaxToken`
    // methods (inherent associated functions take precedence over trait
    // methods in path resolution) — this is delegation, not recursion.
    // The `Self::` form is clippy's own preference here (`use_self`).
    fn kind(&self) -> crate::SyntaxKind {
        Self::kind(self)
    }
    fn text(&self) -> &str {
        Self::text(self)
    }
}

impl TokenView for &rowan::GreenTokenData {
    fn kind(&self) -> crate::SyntaxKind {
        <crate::BeancountLanguage as rowan::Language>::kind_from_raw((*self).kind())
    }
    fn text(&self) -> &str {
        (*self).text()
    }
}
/// The `{*}` merge-flag state machine.
///
/// ONE implementation, driven by both token walkers: [`cost_spec_from_tokens`]
/// feeds it inside its single pass, and [`super::ast::CostSpec::is_merge`]
/// feeds it while walking the red tree. The rule was hand-mirrored in those two
/// places until the 2026-08-01 mutation run showed EVERY mutant in this machine
/// surviving on the `convert.rs` side: the only tests exercising the rule went
/// through the `ast.rs` copy, so the canonical could have been broken outright
/// without one test failing. That is the drift this module's `TokenView` doc
/// says it exists to prevent, in the one place it had not been applied.
///
/// The rule: the flag is decided by the first non-whitespace, non-opener token
/// after an opener (`*` means merge, anything else means not). A `*` elsewhere
/// is the multiplication operator, as in `{500 * 2 USD}`, and a pass that
/// re-arms on later openers flips the flag on malformed input where it must not.
#[derive(Default)]
pub(in crate::cst) struct MergeFlag {
    past_opener: bool,
    decided: bool,
    merge: bool,
}

impl MergeFlag {
    /// Feed the next token kind, in source order. Ignores everything once the
    /// flag is decided.
    pub(in crate::cst) const fn feed(&mut self, kind: crate::SyntaxKind) {
        use crate::SyntaxKind as K;
        if self.decided {
            return;
        }
        match kind {
            K::L_BRACE | K::L_DOUBLE_BRACE | K::L_BRACE_HASH => self.past_opener = true,
            K::WHITESPACE => {}
            K::STAR if self.past_opener => {
                self.merge = true;
                self.decided = true;
            }
            _ if self.past_opener => self.decided = true,
            _ => {}
        }
    }

    /// Whether the tokens fed so far describe a `{*}` merge cost.
    pub(in crate::cst) const fn is_merge(&self) -> bool {
        self.merge
    }
}

/// Convert the direct child tokens of a `COST_SPEC` node into a [`CostSpec`]
/// (forms `{N CCY}`, `{{T CCY}}`, `{N # T CCY}`, `{*}` merge, plus optional
/// date + label). The single source of truth for BOTH walkers (red
/// `convert_cost_spec`, green `convert_cost_spec`).
///
/// Cost numbers are plain `NUMBER` tokens (no arithmetic evaluation); an
/// unparsable one yields `number: None` with no diagnostic, so this needs no
/// bail and always returns a `CostSpec`.
///
/// There are TWO number semantics and both must be carried (#1713):
/// - the compound `{a # b}` path retries past UNPARSABLE number tokens on
///   each side of the hash (`is_none()` guards re-arm when the parse fails);
/// - the plain path uses the first NUMBER *token*, parsed or not (a latch).
///
/// A single latched tracker satisfies only the plain path: on
/// `{<garbage-number> 2 # ...}` the compound side must retry to `2` while
/// the latch keeps `None` (the historical `fuzz_green_eq_red` divergence).
///
/// Compound `{a # b}` (beancount `compound_amount`): per-unit AND a lump
/// total on top; the cost totals `N*a + b`. Surfaced as written — units may
/// be interpolated later, so the combined total cannot be computed here;
/// booking derives it (#1700). An omitted side is zero, which is
/// arithmetically exact (`{# b}` ≡ `{{b}}`, `{a #}` ≡ `{a}`). The hash can
/// arrive fused with the brace as a single `L_BRACE_HASH` opener.
///
/// `is_total` = any `{{` present anywhere. The `{*}` merge flag is decided
/// by the first non-whitespace, non-opener token after an opener (`*` →
/// merge, anything else → not — a STAR elsewhere is the multiplication
/// operator, e.g. `{500 * 2 USD}`); a scan-everything pass that re-arms on
/// later openers flips the flag on malformed inputs where it must not
/// (another historical divergence).
pub(super) fn cost_spec_from_tokens(tokens: impl Iterator<Item = impl TokenView>) -> CostSpec {
    use crate::SyntaxKind as K;
    // Materialized because the arithmetic evaluator needs random access
    // (backtracking over a `(`...`)` group). Cost specs are a handful of
    // tokens, so the allocation is not on any hot path worth defending.
    let toks: Vec<_> = tokens.collect();
    let mut is_total = false;
    let mut first_number: Option<Decimal> = None; // latched (plain path)
    let mut seen_number = false;
    let mut pre_hash: Option<Decimal> = None; // retried (compound path)
    let mut past_hash = false;
    let mut post_hash_total: Option<Decimal> = None; // retried (compound path)
    let mut currency: Option<Currency> = None;
    let mut date: Option<NaiveDate> = None;
    let mut date_seen = false;
    let mut label: Option<String> = None;
    let mut label_seen = false;
    let mut merge_flag = MergeFlag::default();
    for t in &toks {
        let kind = t.kind();
        // Runs alongside the value machine below; see `MergeFlag`.
        merge_flag.feed(kind);
        match kind {
            K::L_DOUBLE_BRACE => is_total = true,
            K::NUMBER => {
                if past_hash {
                    if post_hash_total.is_none() {
                        post_hash_total = parse_decimal_token(t.text());
                    }
                } else {
                    if pre_hash.is_none() {
                        pre_hash = parse_decimal_token(t.text());
                    }
                    if !seen_number {
                        seen_number = true;
                        first_number = parse_decimal_token(t.text());
                    }
                }
            }
            K::HASH | K::L_BRACE_HASH => past_hash = true,
            K::CURRENCY if currency.is_none() => currency = Some(Currency::new(t.text())),
            K::DATE if !date_seen => {
                date_seen = true;
                date = parse_date_token(t.text());
            }
            K::STRING if !label_seen => {
                label_seen = true;
                label = decode_string_token(t.text());
            }
            _ => {}
        }
    }
    // ARITHMETIC OVERRIDE. The latches above take the first NUMBER of each
    // region, which is right for `{10.00 USD}` and wrong for both
    // `{10.00 * 3 USD}` and `{-200.00 USD}` — the latch reads NUMBER tokens
    // only, so it truncated the first and dropped the sign of the second.
    // `cost_region_value` returns None for an unsigned bare number, so that
    // case keeps the latch untouched. See #1939.
    let hash_at = toks
        .iter()
        .position(|t| matches!(t.kind(), K::HASH | K::L_BRACE_HASH));
    match hash_at {
        Some(i) => {
            if let Some(v) = cost_region_value(&toks[..i]) {
                pre_hash = Some(v);
            }
            if let Some(v) = cost_region_value(&toks[i + 1..]) {
                post_hash_total = Some(v);
            }
        }
        None => {
            if let Some(v) = cost_region_value(&toks) {
                first_number = Some(v);
            }
        }
    }

    // A malformed component list means we do not know the cost, so do not
    // report one (#2008). `{, 100.0 USD, , }` and `{45.23 USD / 2015-07-16 /
    // "blabla"}` both have a number our recovery can scrape out, and scraping
    // it is how `rledger check` came to print `E3001 does not balance:
    // residual 980.10 USD` — an arithmetic complaint about a typo, whose
    // number was produced by this function rather than by anything the author
    // wrote. Dropping to "unknown" hands the posting to interpolation, which
    // already knows how to solve for a single unknown per currency group and
    // to reject more than one.
    //
    // The currency is kept: `{, 100.0 USD, , }` names USD unambiguously, so
    // "an unknown cost in USD" is the honest reading. Same rule as `{ # USD}`
    // above — never invent a number the author did not write.
    let shape_ok =
        super::cost_spec_shape::first_cost_spec_defect(toks.iter().map(|t| (t.kind(), ())))
            .is_none();

    let number = if !shape_ok {
        None
    } else if past_hash {
        match (pre_hash, post_hash_total) {
            // `{ # USD}` — no number on EITHER side of the `#`. There is no
            // cost number here at all, so say so; `unwrap_or_default()` used to
            // invent `Compound { per_unit: 0, total: 0 }`, a perfectly
            // determinable zero cost. That is why #2008 case 5 loaded clean:
            // interpolation counts a cost spec with no determinable number as
            // one unknown for its currency and enforces "at most one per
            // currency group", but an invented zero is not an unknown, so the
            // rule never saw it.
            //
            // Reported as `None`, exactly like `{USD}`, because that is what
            // the two shapes have in common: a currency and no number.
            //
            // Scoped to BOTH sides missing. `{100 # USD}` and `{# 500 USD}`
            // still default the absent side to zero — beancount treats it as
            // MISSING and would solve for it, which is a different and larger
            // change. The corpus says that is not urgent: `{ # CCY}` appears in
            // exactly one file (the #2008 fixture), and the one-sided forms
            // appear only in other parser-lima conformance fixtures. Widening
            // this without an oracle to check against is how a compat fix
            // starts breaking real ledgers.
            (None, None) => None,
            (per_unit, total) => Some(CostNumber::Compound {
                per_unit: per_unit.unwrap_or_default(),
                total: total.unwrap_or_default(),
            }),
        }
    } else {
        match (first_number, is_total) {
            (Some(v), true) => Some(CostNumber::Total { value: v }),
            (Some(v), false) => Some(CostNumber::PerUnit { value: v }),
            (None, _) => None,
        }
    };
    CostSpec {
        number,
        currency,
        date,
        label,
        merge: merge_flag.is_merge(),
    }
}

fn convert_price_annotation(
    pa: &ast::PriceAnnotation,
    errors: &mut Vec<crate::ParseError>,
    bom_offset: u32,
) -> PriceAnnotation {
    let kind = if pa.is_total() {
        PriceKind::Total
    } else {
        PriceKind::Unit
    };
    let amount = pa
        .amount()
        .and_then(|a| convert_amount_to_incomplete(&a, errors, bom_offset));
    PriceAnnotation { kind, amount }
}

// ---- Metadata extraction ---------------------------------------

/// Extract the [`Metadata`] map from the directive node's
/// `META_ENTRY` sub-line children. Matches the legacy parser's
/// behavior: each entry's key (with trailing `:` stripped) maps
/// to a typed [`MetaValue`] derived from the value tokens.
fn convert_meta_entries(node: &crate::SyntaxNode) -> Metadata {
    let mut meta = Metadata::default();
    for entry in node.children().filter_map(MetaEntry::cast) {
        let Some(key_token) = entry.key() else {
            continue;
        };
        let key = key_token.text_without_colon().to_string();
        let value = meta_value_from_entry(&entry);
        meta.insert(key, value);
    }
    meta
}

/// The span of flat `POSTING` tokens sitting between the account and the
/// amount, if any — tokens the conversion would otherwise discard in silence.
///
/// CANONICAL: both conversion paths consult this. The green path bails to red
/// when it returns `Some` (red owns the diagnostic), so the two cannot disagree
/// about which postings are well formed — the property `fuzz_green_eq_red`
/// checks.
///
/// Deliberately narrow: only a stray sign or comma counts. Those are the
/// tokens that silently CHANGE A VALUE — a dropped `-` flips the sign, and a
/// misplaced `,` splits a number. Any other junk between the account and the
/// amount (a `✨` inside an account name, a mangled transaction header) is
/// already reported as "unexpected input" by error recovery, and reporting it
/// again here would double up on the same span with a message about thousands
/// separators that does not describe the actual problem.
///
/// Trivia and the posting flag are not orphans, and neither is anything at or
/// after the first `AMOUNT`: trailing junk is already reported separately.
pub(super) const fn is_orphanable_amount_prefix(kind: crate::SyntaxKind) -> bool {
    matches!(
        kind,
        crate::SyntaxKind::MINUS | crate::SyntaxKind::PLUS | crate::SyntaxKind::COMMA
    )
}

pub(super) fn orphaned_amount_prefix(node: &crate::SyntaxNode) -> Option<crate::TextRange> {
    let mut seen_account = false;
    let mut start: Option<crate::TextRange> = None;
    let mut end: Option<crate::TextRange> = None;
    for el in node.children_with_tokens() {
        match el {
            rowan::NodeOrToken::Node(n) => {
                // Only the AMOUNT ends the prefix — deliberately NOT any
                // structured node. A `COST_SPEC` or `PRICE_ANNOTATION` can
                // legitimately precede the units, and a stray sign or comma
                // sitting between one of those and the amount is just as
                // orphaned as one right after the account. Their own commas
                // live INSIDE their nodes, so this cannot false-positive on
                // `{100.00 USD, 2020-01-01}`.
                if ast::Amount::can_cast(n.kind()) {
                    break;
                }
            }
            rowan::NodeOrToken::Token(t) => {
                let kind = t.kind();
                if kind == crate::SyntaxKind::ACCOUNT {
                    seen_account = true;
                    continue;
                }
                // NEWLINE FIRST: `is_trivia_kind` counts it as trivia, so
                // testing that earlier would skip it and keep scanning past the
                // end of the posting line — and would leave the check below
                // unreachable. The green mirror stops at the newline, so
                // getting this order wrong is also how the two paths drift.
                if kind == crate::SyntaxKind::NEWLINE {
                    break;
                }
                if !seen_account || is_trivia_kind(kind) || is_comment_kind(kind) {
                    continue;
                }
                if !is_orphanable_amount_prefix(kind) {
                    continue;
                }
                start.get_or_insert(t.text_range());
                end = Some(t.text_range());
            }
        }
    }
    let (s, e) = (start?, end?);
    Some(crate::TextRange::new(s.start(), e.end()))
}

/// Returns true if a node's flat direct-child tokens contain a
/// `MINUS` BEFORE the first `NUMBER`. Used to detect signed
/// numeric values in directives like Balance / Price whose typed-
/// AST accessors return the unsigned NUMBER token only.
fn node_has_minus_before_number(node: &crate::SyntaxNode) -> bool {
    for el in node.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        match t.kind() {
            crate::SyntaxKind::MINUS => return true,
            crate::SyntaxKind::NUMBER => return false,
            _ => {}
        }
    }
    false
}

/// Discriminate one *value group* of raw metadata/custom value tokens
/// (`tokens[start..]`, trivia already filtered) into a [`MetaValue`], returning
/// the value and the index just past it. This is the single source of truth for
/// the raw token-walk extractors ([`pushmeta_value`] and
/// [`extract_custom_values`]) so they can't drift — they previously did:
/// `extract_custom_values` dropped the leading `MINUS` (so `custom "x" -50.00`
/// emitted `+50.00`) and dropped `Tag`/`Link` entirely, while `pushmeta_value`
/// skipped the `NUMBER CURRENCY` → `Amount` lookahead.
///
/// Discrimination mirrors the typed [`meta_value_from_entry`] (the `META_ENTRY`
/// sibling, which additionally escape-decodes strings via the typed AST). A
/// leading `MINUS` negates the following `NUMBER`; an adjacent `CURRENCY` makes
/// it an `Amount`. Returns `None` for a non-value token (the caller advances).
fn value_tokens_to_meta(
    tokens: &[rowan::SyntaxToken<crate::BeancountLanguage>],
    start: usize,
) -> Option<(MetaValue, usize)> {
    let mut i = start;
    let mut negate = false;
    if tokens.get(i).map(rowan::SyntaxToken::kind) == Some(crate::SyntaxKind::MINUS) {
        negate = true;
        i += 1;
    }
    let t = tokens.get(i)?;
    match t.kind() {
        crate::SyntaxKind::STRING => {
            let s = strip_string_quotes(t.text())?;
            Some((MetaValue::String(s.to_string()), i + 1))
        }
        crate::SyntaxKind::NUMBER => {
            let mut decimal = parse_decimal_token(t.text())?;
            if negate {
                decimal = -decimal;
            }
            // Adjacent `CURRENCY` → `Amount` (negate applies to the amount too).
            if let Some(next) = tokens.get(i + 1)
                && next.kind() == crate::SyntaxKind::CURRENCY
            {
                return Some((
                    MetaValue::Amount(Amount::new(decimal, Currency::new(next.text()))),
                    i + 2,
                ));
            }
            Some((number_meta_value(t.text(), decimal), i + 1))
        }
        crate::SyntaxKind::DATE => Some((MetaValue::Date(parse_date_token(t.text())?), i + 1)),
        crate::SyntaxKind::ACCOUNT => Some((MetaValue::Account(Account::new(t.text())), i + 1)),
        crate::SyntaxKind::CURRENCY => Some((MetaValue::Currency(Currency::new(t.text())), i + 1)),
        crate::SyntaxKind::BOOL_TRUE => Some((MetaValue::Bool(true), i + 1)),
        crate::SyntaxKind::BOOL_FALSE => Some((MetaValue::Bool(false), i + 1)),
        crate::SyntaxKind::TAG => Some((
            MetaValue::Tag(Tag::new(t.text().trim_start_matches('#'))),
            i + 1,
        )),
        crate::SyntaxKind::LINK => Some((
            MetaValue::Link(Link::new(t.text().trim_start_matches('^'))),
            i + 1,
        )),
        _ => None,
    }
}

/// Discriminate the value tokens under a `META_ENTRY` into a typed
/// [`MetaValue`] — thin wrapper over the shared [`meta_value_from_tokens`].
/// The raw-token sibling is [`value_tokens_to_meta`]; keep the two in sync.
fn meta_value_from_entry(entry: &MetaEntry) -> MetaValue {
    meta_value_from_tokens(
        entry
            .syntax()
            .children_with_tokens()
            .filter_map(rowan::NodeOrToken::into_token),
    )
}

/// Derive the typed [`MetaValue`] from a `META_ENTRY` node's direct child
/// tokens. The single source of truth for BOTH walkers (red
/// [`meta_value_from_entry`], green `meta_value`).
///
/// Matches the legacy parser's preference order: string > number/amount >
/// date > account > currency > bool > tag/link > none, where each candidate
/// is the FIRST token of its kind and a type that's present-but-unparsable
/// (a malformed string, an over-precision number, a bad date) falls through
/// to the next.
///
/// A `MINUS` token after the `META_KEY` and before the first `NUMBER`
/// negates the number (legacy `parse_signed_number`, e.g. `precision: -1`);
/// a first `CURRENCY` anywhere alongside the number makes it an `Amount`
/// (legacy priority where `parse_amount` runs before `parse_signed_number`).
pub(super) fn meta_value_from_tokens(tokens: impl Iterator<Item = impl TokenView>) -> MetaValue {
    use crate::SyntaxKind as K;
    // Materialized so the arithmetic evaluator can look at the value region as
    // a slice. `key: 2 * 3` is 6 in beancount and was 2 here — the same
    // first-NUMBER truncation as the cost spec in #1939, in a third place
    // (#1944).
    let toks: Vec<_> = tokens.collect();
    let mut string_t: Option<String> = None;
    let mut number_t: Option<String> = None;
    let mut currency_t: Option<String> = None;
    let mut date_t: Option<String> = None;
    let mut account_t: Option<String> = None;
    let mut bool_v: Option<bool> = None;
    let mut tag_link: Option<MetaValue> = None;
    let mut past_key = false;
    let mut minus = false;
    let mut minus_decided = false;

    for t in &toks {
        let kind = t.kind();
        // First-of-kind value tokens (the `first_token` accessor semantics).
        match kind {
            K::STRING if string_t.is_none() => string_t = Some(t.text().to_string()),
            K::NUMBER if number_t.is_none() => number_t = Some(t.text().to_string()),
            K::CURRENCY if currency_t.is_none() => currency_t = Some(t.text().to_string()),
            K::DATE if date_t.is_none() => date_t = Some(t.text().to_string()),
            K::ACCOUNT if account_t.is_none() => account_t = Some(t.text().to_string()),
            K::BOOL_TRUE if bool_v.is_none() => bool_v = Some(true),
            K::BOOL_FALSE if bool_v.is_none() => bool_v = Some(false),
            K::TAG if tag_link.is_none() => {
                tag_link = Some(MetaValue::Tag(Tag::new(t.text().trim_start_matches('#'))));
            }
            K::LINK if tag_link.is_none() => {
                tag_link = Some(MetaValue::Link(Link::new(t.text().trim_start_matches('^'))));
            }
            _ => {}
        }
        if past_key && !minus_decided {
            match kind {
                K::MINUS => {
                    minus = true;
                    minus_decided = true;
                }
                K::NUMBER => minus_decided = true,
                _ => {}
            }
        }
        // Gates the sign machine so a MINUS in or before the key position is
        // not read as a value's sign. No input reachable through the parser
        // exercises it: the key is always the first meaningful token, and the
        // malformed shapes that would put a MINUS ahead of it (`- key: 42`,
        // `-key: 42`, `key- : 42`) yield no metadata entry at all. Kept as a
        // guard on the token contract rather than deleted, since this helper
        // is shared with the green walker and takes whatever tokens it is
        // handed. Flipping this comparison survives mutation testing for the
        // same reason -- recorded so the next reader does not hunt for a test.
        if kind == K::META_KEY {
            past_key = true;
        }
    }

    if let Some(s) = string_t
        && let Some(decoded) = decode_string_token(&s)
    {
        return MetaValue::String(decoded);
    }
    // Arithmetic override, mirroring the cost-spec fix. Only when the value
    // region really is an expression; a bare number keeps the latch above and
    // its `Int` vs `Number` discrimination, which is archived (cache v11) and
    // must not shift for ordinary metadata.
    //
    // The sign is NOT reapplied on this path: a leading MINUS is part of the
    // expression the evaluator already consumed, so `minus` would double it.
    let value_region: Vec<&_> = toks
        .iter()
        .skip_while(|t| t.kind() != K::META_KEY)
        .filter(|t| !is_trivia_kind(t.kind()) && t.kind() != K::META_KEY)
        .collect();
    if let Some(dec) = cost_region_value(&value_region) {
        if let Some(c) = currency_t {
            return MetaValue::Amount(Amount::new(dec, Currency::new(&c)));
        }
        // Render the RESULT to decide Int vs Number, so `2 * 3` is Int(6) —
        // what beancount reports — rather than inheriting the first operand's
        // spelling.
        return number_meta_value(&dec.to_string(), dec);
    }
    if let Some(nt) = number_t
        && let Some(mut dec) = parse_decimal_token(&nt)
    {
        if minus {
            dec = -dec;
        }
        if let Some(c) = currency_t {
            return MetaValue::Amount(Amount::new(dec, Currency::new(&c)));
        }
        return number_meta_value(&nt, dec);
    }
    if let Some(dt) = date_t
        && let Some(date) = parse_date_token(&dt)
    {
        return MetaValue::Date(date);
    }
    if let Some(a) = account_t {
        return MetaValue::Account(Account::new(&a));
    }
    if let Some(c) = currency_t {
        return MetaValue::Currency(Currency::new(&c));
    }
    if let Some(b) = bool_v {
        return MetaValue::Bool(b);
    }
    if let Some(tl) = tag_link {
        return tl;
    }
    MetaValue::None
}

// ---- Inherited state (pushtag/poptag/pushmeta/popmeta) ---------

/// Merge active pushed-tag and pushed-meta state into a freshly
/// converted directive's value. Mirrors the legacy parser's
/// `apply_pushed_tags` + `apply_pushed_meta`: tags apply ONLY to
/// `Transaction`; meta applies to every directive's `meta` field.
///
/// The meta stack is a `Vec` (not a map) to preserve shadow/pop
/// semantics - `pushmeta x: 1; pushmeta x: 2; popmeta x` should
/// leave `x = 1` active, which a map-replacing-on-insert can't
/// express. Iterating in push order and inserting into the
/// directive's meta means later entries naturally win, matching
/// "topmost-shadow wins" behavior.
fn apply_inherited_state(
    value: &mut Directive,
    tag_stack: &[(Tag, Span)],
    meta_stack: &[(String, MetaValue, Span)],
) {
    if let Directive::Transaction(txn) = value {
        for (tag, _) in tag_stack {
            if !txn.tags.contains(tag) {
                txn.tags.push(tag.clone());
            }
        }
    }
    if meta_stack.is_empty() {
        return;
    }
    let meta = match value {
        Directive::Transaction(d) => &mut d.meta,
        Directive::Balance(d) => &mut d.meta,
        Directive::Open(d) => &mut d.meta,
        Directive::Close(d) => &mut d.meta,
        Directive::Commodity(d) => &mut d.meta,
        Directive::Pad(d) => &mut d.meta,
        Directive::Event(d) => &mut d.meta,
        Directive::Query(d) => &mut d.meta,
        Directive::Note(d) => &mut d.meta,
        Directive::Document(d) => &mut d.meta,
        Directive::Price(d) => &mut d.meta,
        Directive::Custom(d) => &mut d.meta,
    };
    for (k, v, _) in meta_stack {
        meta.insert(k.clone(), v.clone());
    }
}

/// Extract the value tokens after the `META_KEY` of a Pushmeta
/// directive into a typed [`MetaValue`]. Walks the directive's
/// direct-child tokens (the directive isn't a `META_ENTRY` so the
/// typed-AST accessors aren't reusable).
fn pushmeta_value(node: &crate::SyntaxNode) -> MetaValue {
    // The first value token after the key wins. `value_tokens_to_meta` returns
    // `None` for the key/colon (and any non-value token), so the loop walks to
    // the first real value — sharing MINUS-sign, Tag/Link and
    // `NUMBER CURRENCY` → Amount handling with metadata/custom values.
    let raw: Vec<rowan::SyntaxToken<crate::BeancountLanguage>> = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .filter(|t| {
            !matches!(
                t.kind(),
                crate::SyntaxKind::WHITESPACE
                    | crate::SyntaxKind::NEWLINE
                    | crate::SyntaxKind::COMMENT
            )
        })
        .collect();

    let mut i = 0;
    while i < raw.len() {
        if let Some((value, _)) = value_tokens_to_meta(&raw, i) {
            return value;
        }
        i += 1;
    }
    MetaValue::None
}

// ---- ParseResult.comments --------------------------------------

/// Comment-like syntax kinds that the legacy parser surfaces as
/// `ParseResult.comments` entries when they appear at the top
/// level (outside any directive's content).
pub(super) const fn is_comment_kind(kind: crate::SyntaxKind) -> bool {
    matches!(
        kind,
        crate::SyntaxKind::COMMENT
            | crate::SyntaxKind::PERCENT_COMMENT
            | crate::SyntaxKind::SHEBANG
            | crate::SyntaxKind::EMACS_DIRECTIVE
    )
}

/// Output of the fused top-level pass [`walk_top_level_once`].
pub(super) struct TopLevelWalkResult {
    pub(super) errors: Vec<crate::ParseError>,
    pub(super) section_marker_comments: Vec<Spanned<String>>,
}

/// Single walk over `source_file`'s direct children that runs
/// every per-directive diagnostic in one pass, replacing five
/// separate `source_file.syntax().children()` traversals
/// (`extract_error_node_errors`, `extract_transaction_body_errors`,
/// `extract_indented_directive_errors`, `extract_custom_value_errors`,
/// `extract_section_marker_comments`). Each former pass re-walked
/// the top-level child list and materialized a fresh red node per
/// directive; on a large ledger that is 5·O(N) red-node churn for
/// work that is naturally per-child. The checks are independent and
/// all diagnostics are span-sorted by the caller, so fusing them is
/// order-preserving.
fn walk_top_level_once(
    source_file: &SourceFile,
    stripped: &str,
    bom_offset: u32,
) -> TopLevelWalkResult {
    let mut errors: Vec<crate::ParseError> = Vec::new();
    let mut section_marker_comments: Vec<Spanned<String>> = Vec::new();
    for child in source_file.syntax().children() {
        let kind = child.kind();
        // Applies to every recognized directive node (incl. CUSTOM).
        if ast::Directive::can_cast(kind) {
            indented_directive_check(&child, stripped, bom_offset, &mut errors);
        }
        match kind {
            crate::SyntaxKind::CUSTOM_DIRECTIVE => {
                custom_value_check(&child, bom_offset, &mut errors);
            }
            crate::SyntaxKind::TRANSACTION => {
                transaction_header_check(&child, stripped, bom_offset, &mut errors);
                transaction_body_check(&child, bom_offset, &mut errors);
            }
            crate::SyntaxKind::ERROR_NODE => {
                error_node_check(&child, stripped, bom_offset, &mut errors);
                section_marker_check(&child, bom_offset, &mut section_marker_comments);
            }
            _ => {}
        }
    }
    TopLevelWalkResult {
        errors,
        section_marker_comments,
    }
}

/// A `^link` is not a valid metadata VALUE (#1954).
///
/// beancount's grammar has no production for it — `ref: ^inv-1` fails with
/// `syntax error, unexpected LINK, expecting end of file or EOL` — while a
/// `#tag` in the same position is perfectly valid there. We accepted both.
///
/// Deliberately asymmetric, and that asymmetry is the whole point: rejecting
/// TAG here as well would break input beancount accepts, which is the mistake
/// #1953 avoided in the mirror-image case. Tags and links lex as sibling kinds
/// and are handled as a pair almost everywhere in this file, so the pairing is
/// the natural thing to reach for and the wrong thing to do.
///
/// Scoped to `META_ENTRY` nodes. A link on a TRANSACTION (`* "x" ^lnk`) is a
/// direct child of the transaction, not of a metadata entry, so it is
/// untouched — as it must be, that being the one place links belong.
///
/// NOT extended to `custom` / `pushmeta` values, which run through
/// `value_tokens_to_meta` rather than this path. beancount rejects BOTH a tag
/// and a link there, so it is a different rule needing its own evidence;
/// filed separately rather than folded in here.
fn extract_link_metadata_value_errors(
    source_file: &SourceFile,
    bom_offset: u32,
) -> Vec<crate::ParseError> {
    let mut out = Vec::new();
    for entry in source_file.syntax().descendants() {
        if entry.kind() != crate::SyntaxKind::META_ENTRY {
            continue;
        }
        for el in entry.children_with_tokens() {
            let rowan::NodeOrToken::Token(t) = el else {
                continue;
            };
            if t.kind() != crate::SyntaxKind::LINK {
                continue;
            }
            let range = t.text_range();
            let off = bom_offset as usize;
            out.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError(format!(
                    "a link ({}) is not a valid metadata value; beancount \
                     accepts a tag here but not a link",
                    t.text()
                )),
                Span::new(
                    usize::from(range.start()) + off,
                    usize::from(range.end()) + off,
                ),
            ));
        }
    }
    out
}

/// Walk every `COST_SPEC` node in the tree and emit a
/// `SyntaxError("unclosed cost specification: missing '}'")` for
/// any spec whose opener (`{`, `{{`, or `{#`) doesn't have a
/// matching closer at the spec's depth-0. Mirrors the legacy
/// parser's deferred-error emission at `parser.rs:705-707` so a
/// `10 AAPL {150 USD\n` posting or an EOF-truncated cost block
/// surfaces a diagnostic instead of silently producing a half-
/// built cost spec.
/// Tags and links are not valid `custom` or `pushmeta` VALUES (#1958).
///
/// Two different rules, and conflating them is the trap here:
///
/// | position          | `#tag`   | `^link`  |
/// |-------------------|----------|----------|
/// | metadata value    | valid    | rejected |
/// | `pushmeta k: ...` | valid    | rejected |
/// | `custom "t" ...`  | rejected | rejected |
///
/// `pushmeta` follows the METADATA rule - it pushes a metadata key/value, so a
/// tag is fine there and only a link is not. `custom` is stricter than both and
/// takes neither. Checked against beancount per position, each with a control
/// on the same directive.
///
/// This is why the check cannot live inside `value_tokens_to_meta`, which both
/// callers share: one rule in the shared helper would either let a tag through
/// in `custom` or wrongly reject one in `pushmeta`. It is the same shape as
/// #1953 (`note`/`document` DO take both) and #1954 (a tag is valid where a
/// link is not) - three times now the right answer has split a pair that lexes
/// and reads as one.
fn extract_custom_pushmeta_taglink_errors(
    source_file: &SourceFile,
    bom_offset: u32,
) -> Vec<crate::ParseError> {
    use crate::SyntaxKind as K;
    let mut out = Vec::new();
    for node in source_file.syntax().descendants() {
        let (reject_tag, what) = match node.kind() {
            K::CUSTOM_DIRECTIVE => (true, "custom"),
            K::PUSHMETA_DIRECTIVE => (false, "pushmeta"),
            _ => continue,
        };
        for el in node.children_with_tokens() {
            let rowan::NodeOrToken::Token(t) = el else {
                continue;
            };
            let kind = t.kind();
            let bad = match kind {
                K::LINK => true,
                K::TAG => reject_tag,
                _ => false,
            };
            if !bad {
                continue;
            }
            let noun = if kind == K::TAG { "tag" } else { "link" };
            let range = t.text_range();
            let off = bom_offset as usize;
            out.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError(format!(
                    "a {noun} ({}) is not a valid {what} value",
                    t.text()
                )),
                Span::new(
                    usize::from(range.start()) + off,
                    usize::from(range.end()) + off,
                ),
            ));
        }
    }
    out
}

fn extract_unclosed_cost_brace_errors(
    source_file: &SourceFile,
    stripped: &str,
    bom_offset: u32,
) -> Vec<crate::ParseError> {
    let mut out = Vec::new();
    for cs in source_file.syntax().descendants() {
        if cs.kind() != crate::SyntaxKind::COST_SPEC {
            continue;
        }
        let mut has_opener = false;
        let mut has_closer = false;
        for el in cs.children_with_tokens() {
            let rowan::NodeOrToken::Token(t) = el else {
                continue;
            };
            match t.kind() {
                crate::SyntaxKind::L_BRACE
                | crate::SyntaxKind::L_DOUBLE_BRACE
                | crate::SyntaxKind::L_BRACE_HASH => has_opener = true,
                crate::SyntaxKind::R_BRACE | crate::SyntaxKind::R_DOUBLE_BRACE => has_closer = true,
                _ => {}
            }
        }
        if has_opener && !has_closer {
            out.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError(
                    "unclosed cost specification: missing '}'".to_string(),
                ),
                node_span(&cs, bom_offset),
            ));
            // An unclosed spec has no meaningful component list; reporting a
            // shape defect on top would just be noise about the truncation.
            continue;
        }

        // Component-list shape (#2008 cases 1 and 2). Fused into this walk
        // rather than given its own: `descendants()` allocates a red node per
        // node, which is why this scan already sits behind a `contains('{')`
        // guard, and a second identical pass would double a cost the codebase
        // deliberately profiled down.
        let tokens = cs
            .children_with_tokens()
            .filter_map(rowan::NodeOrToken::into_token)
            .map(|t| {
                let r = t.text_range();
                (t.kind(), usize::from(r.start())..usize::from(r.end()))
            });
        if let Some((defect, range)) = super::cost_spec_shape::first_cost_spec_defect(tokens) {
            // `get` rather than indexing: a non-char-boundary range must not
            // panic the parser.
            let message = match stripped.get(range.clone()) {
                Some(text) => super::cost_spec_shape::cost_defect_message(defect, text),
                None => format!(
                    "malformed cost specification at bytes {}..{} ({defect:?})",
                    range.start, range.end
                ),
            };
            out.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError(message),
                Span::new(
                    range.start + bom_offset as usize,
                    range.end + bom_offset as usize,
                ),
            ));
        }
    }
    out
}

/// Walk every top-level directive in `source_file` and emit a
/// `SyntaxError("top-level directive must start at column 0")`
/// for any whose content (first non-trivia token) starts at a
/// non-zero column. Per the Beancount language spec, top-level
/// directives are required to begin at column 0; indentation is
/// reserved for postings and metadata inside a transaction body.
///
/// The CST grammar happily accepts an indented `open` / `balance`
/// / etc., which is why this surfaces at converter level instead
/// of as a lex/parse error.
fn indented_directive_check(
    child: &crate::SyntaxNode,
    stripped: &str,
    bom_offset: u32,
    out: &mut Vec<crate::ParseError>,
) {
    // Caller dispatches: `child` is a recognized directive node.
    // Find the directive's content start - the first non-
    // trivia token. Leading WHITESPACE / NEWLINE / COMMENT
    // can land inside the directive node per the Directive-
    // Terminator Rule's inter-directive trivia attachment.
    let Some(content) = child
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .find(|t| !is_trivia_kind(t.kind()))
    else {
        return;
    };
    let content_start: usize = u32::from(content.text_range().start()) as usize;
    // Column = offset since the last NEWLINE in the source,
    // or since byte 0 if this is the first line. >0 means
    // the directive's first content token has leading WS on
    // its own line - that's the indent error.
    // Find the line start by scanning the BYTES before `content_start`, not by
    // slicing the `str`. On malformed/error-recovered input a token's start
    // offset can land inside a multi-byte UTF-8 char, and
    // `stripped[..content_start]` would then panic ("not a char boundary").
    // Byte slicing is boundary-agnostic, and a newline (`\n`) is always a single
    // ASCII byte, so the found position is a valid offset. `.get(..)` also guards
    // a (theoretical) out-of-bounds offset. Regression: fuzz_regressions.rs.
    let line_start = stripped
        .as_bytes()
        .get(..content_start)
        .and_then(|bytes| bytes.iter().rposition(|&b| b == b'\n'))
        .map_or(0, |nl| nl + 1);
    if content_start > line_start {
        let end: u32 = content.text_range().end().into();
        let span = Span::new(
            (line_start as u32 + bom_offset) as usize,
            (end + bom_offset) as usize,
        );
        out.push(crate::ParseError::new(
            crate::ParseErrorKind::SyntaxError(
                "top-level directive must start at column 0".to_string(),
            ),
            span,
        ));
    }
}

/// Walk each `CUSTOM` directive and emit a `SyntaxError` for
/// every bare `CURRENCY` token in the value position (a CURRENCY
/// not paired with a preceding NUMBER as an Amount).
///
/// Per the Beancount language spec, custom-directive values are
/// limited to string / date / decimal / amount / boolean -
/// `bean-check` rejects a bare currency literal with a syntax
/// error. Rustledger's `extract_custom_values` has historically
/// been more lenient, accepting ACCOUNT / TAG / LINK in value
/// position too; we keep that extension (it's covered by the
/// existing `test_parse_custom_directive` integration test) but
/// surface a diagnostic for the bare-CURRENCY case so the
/// compat metric reflects bean-check's exit-code rejection on
/// shapes like `custom "x" 10 USD "y" NZD …`.
fn custom_value_check(
    child: &crate::SyntaxNode,
    bom_offset: u32,
    out: &mut Vec<crate::ParseError>,
) {
    // Caller dispatches: `child` is a CUSTOM_DIRECTIVE.
    {
        // Collect non-trivia tokens, then skip past the
        // directive's header: DATE, CUSTOM_KW, and the first
        // STRING (the custom-type name). Everything after that
        // is values.
        let raw: Vec<crate::SyntaxToken> = child
            .children_with_tokens()
            .filter_map(rowan::NodeOrToken::into_token)
            .filter(|t| !is_trivia_kind(t.kind()))
            .collect();
        let mut seen_type_string = false;
        let mut i = 0;
        while i < raw.len() {
            let t = &raw[i];
            if !seen_type_string {
                if t.kind() == crate::SyntaxKind::STRING {
                    seen_type_string = true;
                }
                i += 1;
                continue;
            }
            if t.kind() == crate::SyntaxKind::CURRENCY {
                // Only flag BARE CURRENCY - one that doesn't
                // follow a NUMBER (Amount-pairing). The Amount
                // pairing is handled by `extract_custom_values`
                // via i+1 lookahead, so a CURRENCY that's NOT
                // preceded by a NUMBER at i-1 is bare.
                let preceded_by_number = i > 0 && raw[i - 1].kind() == crate::SyntaxKind::NUMBER;
                if !preceded_by_number {
                    let range = t.text_range();
                    let start: u32 = range.start().into();
                    let end: u32 = range.end().into();
                    let span =
                        Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
                    out.push(crate::ParseError::new(
                        crate::ParseErrorKind::SyntaxError(
                            "bare currency literal is not a valid custom directive value"
                                .to_string(),
                        ),
                        span,
                    ));
                }
            }
            i += 1;
        }
    }
}

/// Walk a `TRANSACTION` body and emit a `SyntaxError` for any body
/// line that contains flat catch-all tokens (e.g., an
/// unrecognized identifier where a posting was expected).
/// Matches the legacy parser, which fails its inner posting
/// parser on such lines and recovers by skipping to the next
/// NEWLINE while emitting a `SyntaxError`.
/// The `unexpected input` diagnostic for one catch-all transaction-body line.
///
/// Mirrors `green::unexpected_body_input`; shared by the newline- and
/// EOF-terminated sites so the span rule cannot drift between them.
fn unexpected_body_input(line_start: u32, end: u32, bom_offset: u32) -> crate::ParseError {
    crate::ParseError::new(
        crate::ParseErrorKind::SyntaxError("unexpected input".to_string()),
        Span::new(
            (line_start + bom_offset) as usize,
            (end + bom_offset) as usize,
        ),
    )
}

/// Reject transaction headers beancount's grammar refuses (#2008 cases 3, 4,
/// 6, 7). The rule itself lives in [`super::txn_header`] so this and its green
/// mirror (`green::tl_transaction_header_check`) cannot drift; here we only
/// enumerate the header tokens.
///
/// Token enumeration goes through `Transaction::header_tokens`, the same
/// accessor `flag()` / `strings()` / `tags()` use, so "what counts as the
/// header" has one definition.
fn transaction_header_check(
    child: &crate::SyntaxNode,
    stripped: &str,
    bom_offset: u32,
    out: &mut Vec<crate::ParseError>,
) {
    let Some(txn) = ast::Transaction::cast(child.clone()) else {
        return;
    };
    let tokens = txn.header_tokens().map(|t| {
        let r = t.text_range();
        (t.kind(), usize::from(r.start())..usize::from(r.end()))
    });
    if let Some((defect, range)) = super::txn_header::first_header_defect(tokens) {
        out.push(header_defect_error(defect, &range, stripped, bom_offset));
    }
}

/// Shared by both walkers so one defect cannot be reported two ways.
pub(super) fn header_defect_error(
    defect: super::txn_header::HeaderDefect,
    range: &std::ops::Range<usize>,
    stripped: &str,
    bom_offset: u32,
) -> crate::ParseError {
    // `get` rather than indexing: a range that is not a char boundary would
    // panic, and a parser must not panic on malformed input. Token ranges are
    // always char boundaries, so this is unreachable today — but an empty
    // slice would render as `unexpected "" in transaction header`, which names
    // nothing. Fall back to the byte range instead, so even the unreachable
    // branch produces something a reader can act on.
    let message = match stripped.get(range.clone()) {
        Some(text) => super::txn_header::defect_message(defect, text),
        None => format!(
            "malformed transaction header at bytes {}..{} ({defect:?})",
            range.start, range.end
        ),
    };
    crate::ParseError::new(
        crate::ParseErrorKind::SyntaxError(message),
        Span::new(
            range.start + bom_offset as usize,
            range.end + bom_offset as usize,
        ),
    )
}

fn transaction_body_check(
    child: &crate::SyntaxNode,
    bom_offset: u32,
    out: &mut Vec<crate::ParseError>,
) {
    // Caller dispatches: `child` is a TRANSACTION.
    {
        // Skip past the header NEWLINE, then look for catch-all
        // tokens (non-trivia, non-comment) appearing on lines
        // OUTSIDE POSTING / META_ENTRY child nodes.
        // Track whether we've SEEN at least one non-trivia
        // header token (DATE / flag / STRING / etc.); only AFTER
        // that does the next NEWLINE count as the header
        // terminator. Otherwise leading-trivia NEWLINEs from the
        // Directive-Terminator Rule would falsely trip
        // past_header on the very first iteration.
        let mut past_header = false;
        let mut saw_header_content = false;
        let mut line_start: Option<u32> = None;
        let mut line_has_content = false;
        for el in child.children_with_tokens() {
            match el {
                rowan::NodeOrToken::Token(t) => {
                    if !past_header {
                        if t.kind() == crate::SyntaxKind::NEWLINE {
                            if saw_header_content {
                                past_header = true;
                            }
                        } else if !is_trivia_kind(t.kind()) {
                            saw_header_content = true;
                        }
                        continue;
                    }
                    let range = t.text_range();
                    let start: u32 = range.start().into();
                    let end: u32 = range.end().into();
                    if line_start.is_none() {
                        line_start = Some(start);
                    }
                    if t.kind() == crate::SyntaxKind::NEWLINE {
                        if line_has_content && let Some(ls) = line_start {
                            out.push(unexpected_body_input(ls, end, bom_offset));
                        }
                        line_start = None;
                        line_has_content = false;
                    } else if !is_trivia_kind(t.kind())
                        && !is_comment_kind(t.kind())
                        && !matches!(t.kind(), crate::SyntaxKind::TAG | crate::SyntaxKind::LINK)
                    {
                        // TAG / LINK on body lines is valid
                        // Beancount syntax (tags/links after the
                        // first line continue the transaction's
                        // tag/link list). Don't flag as
                        // unexpected-input.
                        line_has_content = true;
                    }
                }
                rowan::NodeOrToken::Node(_) => {
                    // POSTING / META_ENTRY: not catch-all. Reset.
                    line_start = None;
                    line_has_content = false;
                    if !past_header {
                        past_header = true;
                    }
                }
            }
        }
        // EOF terminates the final body line, same as a NEWLINE. Mirrors
        // `green::tl_transaction_body_check` (#1884).
        if past_header
            && line_has_content
            && let Some(ls) = line_start
        {
            let end: u32 = child.text_range().end().into();
            out.push(unexpected_body_input(ls, end, bom_offset));
        }
    }
}

/// Walk an `ERROR_NODE` and emit a
/// `ParseError` for each line that is NEITHER a section marker
/// (`*`-starting) NOR a column-0 comment. The variant emitted
/// mirrors the legacy parser's error-recovery classifier
/// (`parser.rs:2186-2249`): BOM-in-line → `BomInDirectiveBody`
/// (with `BOM_REMOVAL_HINT`); Unicode-character account →
/// `InvalidAccount`; otherwise → `SyntaxError("unexpected
/// input")`. `stripped` is the post-BOM-strip source so token
/// `text_range` indices into it correctly.
/// Emit the recovery diagnostics for one `ERROR_NODE` line.
///
/// Mirrors `green::emit_error_node_line`, and exists for the same reason:
/// this runs from BOTH the newline-terminated and the EOF-terminated site, and
/// two inline copies is how the span rule or the secondary BOM diagnostic
/// drifts between them.
fn emit_error_node_line(
    first_non_trivia: Option<crate::SyntaxKind>,
    line_start: Option<u32>,
    end: u32,
    bom_offset: u32,
    stripped: &str,
    out: &mut Vec<crate::ParseError>,
) {
    let is_section = matches!(first_non_trivia, Some(crate::SyntaxKind::STAR));
    let is_comment = matches!(first_non_trivia, Some(k) if is_comment_kind(k));
    if is_section || is_comment || first_non_trivia.is_none() {
        return;
    }
    let Some(ls) = line_start else { return };
    // Legacy span INCLUDES the terminator NEWLINE (skip_to_newline consumes it
    // before span_from is called); at EOF the terminator is the node end.
    let span = Span::new((ls + bom_offset) as usize, (end + bom_offset) as usize);
    let line_text = stripped.get(ls as usize..end as usize).unwrap_or("");
    let primary = classify_recovery_error(line_text, span);
    let primary_is_bom = matches!(primary.kind, crate::ParseErrorKind::BomInDirectiveBody);
    out.push(primary);
    // Additive secondary `BomInDirectiveBody` when a different primary
    // diagnostic already fired AND the line ALSO contains a BOM byte. Matches
    // legacy `parser.rs:2258-2263`: without it, a Windows-exported line with
    // both problems surfaces only the actionable root cause and the user has no
    // clue the invisible BOM byte is also corrupting the line.
    if !primary_is_bom && line_text.contains(crate::bom::BOM_CHAR) {
        out.push(
            crate::ParseError::new(crate::ParseErrorKind::BomInDirectiveBody, span)
                .with_hint(crate::diagnostics::BOM_REMOVAL_HINT),
        );
    }
}

fn error_node_check(
    child: &crate::SyntaxNode,
    stripped: &str,
    bom_offset: u32,
    out: &mut Vec<crate::ParseError>,
) {
    // Caller dispatches: `child` is an ERROR_NODE.
    {
        let mut line_start: Option<u32> = None;
        let mut first_non_trivia: Option<crate::SyntaxKind> = None;
        for el in child.children_with_tokens() {
            let rowan::NodeOrToken::Token(t) = el else {
                continue;
            };
            let range = t.text_range();
            let start: u32 = range.start().into();
            let end: u32 = range.end().into();
            if line_start.is_none() {
                line_start = Some(start);
            }
            if t.kind() == crate::SyntaxKind::NEWLINE {
                emit_error_node_line(first_non_trivia, line_start, end, bom_offset, stripped, out);
                line_start = None;
                first_non_trivia = None;
                continue;
            }
            if first_non_trivia.is_none() && !is_trivia_kind(t.kind()) {
                first_non_trivia = Some(t.kind());
            }
        }
        // EOF terminates the final line exactly as a NEWLINE would. Mirrors
        // `green::tl_error_node_check`; without it the two paths disagree on
        // any input lacking a trailing newline, and a malformed last line
        // produced no diagnostic at all (#1884).
        let end: u32 = child.text_range().end().into();
        emit_error_node_line(first_non_trivia, line_start, end, bom_offset, stripped, out);
    }
}

/// Pick the most specific `ParseError` variant for an
/// error-recovery line, mirroring the legacy parser's classifier
/// at `parser.rs:2186-2249`:
/// 1. A Unicode-character account (`Assets:Café:…`) → primary
///    `InvalidAccount` - it's the actionable root cause.
/// 2. A mid-file BOM byte (`U+FEFF`) → `BomInDirectiveBody` with
///    `BOM_REMOVAL_HINT` so miette surfaces the remediation step.
/// 3. Anything else → `SyntaxError("unexpected input")`.
///
/// Order matters: a Windows-exported file with a Unicode account
/// AND an internal BOM gets the Unicode-account diagnostic
/// (the BOM is usually a side effect, not the root cause).
pub(super) fn classify_recovery_error(line_text: &str, span: Span) -> crate::ParseError {
    if let Some(account) = crate::diagnostics::find_unicode_account(line_text) {
        return crate::ParseError::new(
            crate::ParseErrorKind::InvalidAccount(account.to_string()),
            span,
        );
    }
    if line_text.contains(crate::bom::BOM_CHAR) {
        return crate::ParseError::new(crate::ParseErrorKind::BomInDirectiveBody, span)
            .with_hint(crate::diagnostics::BOM_REMOVAL_HINT);
    }
    crate::ParseError::new(
        crate::ParseErrorKind::SyntaxError("unexpected input".to_string()),
        span,
    )
}

/// Walk every descendant token and emit a `ParseError` for each
/// `ERROR_TOKEN` (or BOM-containing token) that lands inside an
/// otherwise-valid directive node - i.e., NOT inside an
/// `ERROR_NODE` ancestor. Catches lexer-reject bytes the
/// outer recovery path misses:
/// - `.` in `.50 USD` (leading-decimal in posting amount) →
///   `SyntaxError`.
/// - Mid-file U+FEFF byte inside a recognized directive (e.g.,
///   `open Assets:Bank \u{FEFF}USD`) → `BomInDirectiveBody` with
///   `BOM_REMOVAL_HINT`.
///
/// The leading `SyntaxKind::BOM` token is skipped (the
/// legitimate strict-byte-0 BOM is already tracked by
/// `has_leading_bom`). `ERROR_NODE` descendants are skipped -
/// `extract_error_node_errors` / `classify_recovery_error`
/// already cover those.
/// Result of the fused descendants-walk visitor that powers
/// `walk_descendants_once`.
pub(super) struct DescendantsWalkResult {
    pub(super) inline_errors: Vec<crate::ParseError>,
    pub(super) top_level_comments: Vec<Spanned<String>>,
    pub(super) currency_occurrences: Vec<Spanned<Currency>>,
    pub(super) account_occurrences: Vec<Spanned<rustledger_core::Account>>,
    /// The three per-node shape rules, kept in SEPARATE vecs rather than
    /// merged into `inline_errors`.
    ///
    /// Two reasons. They are emitted at a different point in the error order
    /// than the inline errors (see `parse_via_cst_inner`), and each vec stays
    /// grouped the way its former standalone pass emitted it — document order
    /// within a rule, rules in a fixed sequence. Merging them would interleave
    /// the three by position, which no test pins today but which is
    /// observable in every diagnostic list rledger prints.
    ///
    /// Populated only by the GREEN walker. The red walker leaves them empty
    /// and `parse_via_cst_inner` calls the standalone `extract_*` functions
    /// for that path instead — see the call site for why the fold is
    /// green-only.
    pub(super) cost_brace_errors: Vec<crate::ParseError>,
    pub(super) link_meta_errors: Vec<crate::ParseError>,
    pub(super) custom_pushmeta_errors: Vec<crate::ParseError>,
}

/// Fused single-pass visitor over `source_file`'s descendants -
/// replaces three separate walks (`extract_inline_token_errors`,
/// `extract_top_level_comments`, `extract_currency_occurrences`)
/// with one traversal. Each walk had its own per-token cost; the
/// LSP runs them on every keystroke, so collapsing 3·O(N) → 1·O(N)
/// matters at editor-edge latencies. The state of each former
/// walk is maintained inline below.
fn walk_descendants_once(
    source_file: &SourceFile,
    bom_offset: u32,
    collect_occurrences: bool,
) -> DescendantsWalkResult {
    let mut inline_errors: Vec<crate::ParseError> = Vec::new();
    let mut top_level_comments: Vec<Spanned<String>> = Vec::new();
    let mut currency_occurrences: Vec<Spanned<Currency>> = Vec::new();
    let mut account_occurrences: Vec<Spanned<rustledger_core::Account>> = Vec::new();

    // `extract_top_level_comments` state: column-0 tracking.
    let mut preceded_by_ws = false;

    for el in source_file.syntax().descendants_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            // `extract_top_level_comments` used the Node arm to
            // reset preceded_by_ws when entering a recognized
            // directive. Keep that behavior - directive leading
            // trivia still gets column-0-classified correctly.
            if let rowan::NodeOrToken::Node(n) = el
                && ast::Directive::can_cast(n.kind())
            {
                preceded_by_ws = false;
            }
            continue;
        };

        // ---- `extract_top_level_comments` state machine -------
        match t.kind() {
            crate::SyntaxKind::NEWLINE => preceded_by_ws = false,
            crate::SyntaxKind::WHITESPACE => preceded_by_ws = true,
            k if is_comment_kind(k) => {
                if !preceded_by_ws {
                    let range = t.text_range();
                    let start: u32 = range.start().into();
                    let end: u32 = range.end().into();
                    let span =
                        Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
                    top_level_comments.push(Spanned::new(t.text().to_string(), span));
                }
            }
            _ => {
                preceded_by_ws = false;
            }
        }

        // ---- `extract_inline_token_errors` + currency walks ---
        if t.kind() == crate::SyntaxKind::BOM {
            continue;
        }
        // ERROR_NODE-ancestor check is only consulted for tokens
        // whose downstream emission depends on it (CURRENCY, BOM-
        // text-containing, ERROR_TOKEN). For well-formed source
        // most tokens fall into none of those buckets - gating
        // the per-token `parent_ancestors` walk on relevance
        // saves an O(depth) probe per WHITESPACE/NEWLINE/comment
        // token, which dominates token counts in real ledgers.
        let kind = t.kind();
        let has_bom = t.text().contains(crate::bom::BOM_CHAR);
        let is_error_token = kind == crate::SyntaxKind::ERROR_TOKEN;
        // CURRENCY/ACCOUNT need the in-ERROR_NODE probe only to decide whether
        // to record an occurrence; skip it entirely when not collecting.
        let needs_in_error_check = (collect_occurrences
            && matches!(
                kind,
                crate::SyntaxKind::CURRENCY | crate::SyntaxKind::ACCOUNT
            ))
            || has_bom
            || is_error_token;
        if !needs_in_error_check {
            continue;
        }
        let in_error_node = t
            .parent_ancestors()
            .any(|a| a.kind() == crate::SyntaxKind::ERROR_NODE);

        // CURRENCY occurrences: only outside ERROR_NODE, and only when the
        // caller wants them (LSP path).
        if collect_occurrences && kind == crate::SyntaxKind::CURRENCY && !in_error_node {
            let range = t.text_range();
            let start: u32 = range.start().into();
            let end: u32 = range.end().into();
            let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
            currency_occurrences.push(Spanned::new(Currency::new(t.text()), span));
        }

        // ACCOUNT occurrences: only outside ERROR_NODE. The same
        // rationale as CURRENCY applies - the lexer classifies an
        // `ACCOUNT` token by its character shape independent of
        // whether the surrounding directive parses cleanly, and
        // source-position-aware tooling (LSP rename / references /
        // document-highlight) wants the token as the user typed it
        // even during a mid-edit broken state.
        if collect_occurrences && kind == crate::SyntaxKind::ACCOUNT && !in_error_node {
            let range = t.text_range();
            let start: u32 = range.start().into();
            let end: u32 = range.end().into();
            let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
            account_occurrences.push(Spanned::new(rustledger_core::Account::new(t.text()), span));
        }

        // Inline errors: BOM byte in a recognized directive
        // (-> BomInDirectiveBody + hint) or ERROR_TOKEN inside a
        // recognized directive (-> SyntaxError). Both skip when
        // already inside an ERROR_NODE (handled by the recovery
        // classifier).
        if (!has_bom && !is_error_token) || in_error_node {
            continue;
        }
        let range = t.text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
        if has_bom {
            inline_errors.push(
                crate::ParseError::new(crate::ParseErrorKind::BomInDirectiveBody, span)
                    .with_hint(crate::diagnostics::BOM_REMOVAL_HINT),
            );
        } else {
            inline_errors.push(crate::ParseError::new(
                crate::ParseErrorKind::SyntaxError("unexpected input".to_string()),
                span,
            ));
        }
    }

    DescendantsWalkResult {
        inline_errors,
        top_level_comments,
        currency_occurrences,
        account_occurrences,
        // Red keeps the three shape rules as standalone `extract_*` passes;
        // `parse_via_cst_inner` calls them for this path. Folding them in here
        // too would need each token's IMMEDIATE parent kind, and this walk is
        // a flat `descendants_with_tokens()` — recovering the parent means
        // `t.parent()`, which allocates the very red `NodeData` the green fold
        // exists to avoid.
        cost_brace_errors: Vec::new(),
        link_meta_errors: Vec::new(),
        custom_pushmeta_errors: Vec::new(),
    }
}

/// Emit empty-string comments for org-mode section-marker
/// lines (`* Heading`, `** Subheading`) inside `ERROR_NODE`
/// children. The legacy parser's `parse_entry` matches
/// `Token::Star` and emits `Comment(String::new(), line_span)`;
/// the structured CST wraps these lines in `ERROR_NODE`s so we
/// have to walk them and synthesize the same shape.
fn section_marker_check(
    child: &crate::SyntaxNode,
    bom_offset: u32,
    out: &mut Vec<Spanned<String>>,
) {
    // Caller dispatches: `child` is an ERROR_NODE.
    // Walk tokens line-by-line. A line starts at the start
    // of the first token after a NEWLINE (or at the node's
    // start) and ends at the next NEWLINE (inclusive).
    let mut line_start: Option<u32> = None;
    let mut first_non_trivia: Option<crate::SyntaxKind> = None;
    for el in child.children_with_tokens() {
        let rowan::NodeOrToken::Token(t) = el else {
            continue;
        };
        let range = t.text_range();
        let start: u32 = range.start().into();
        let end: u32 = range.end().into();
        if line_start.is_none() {
            line_start = Some(start);
        }
        if t.kind() == crate::SyntaxKind::NEWLINE {
            if first_non_trivia == Some(crate::SyntaxKind::STAR)
                && let Some(ls) = line_start
            {
                let span = Span::new((ls + bom_offset) as usize, (end + bom_offset) as usize);
                out.push(Spanned::new(String::new(), span));
            }
            line_start = None;
            first_non_trivia = None;
            continue;
        }
        if first_non_trivia.is_none() && !is_trivia_kind(t.kind()) {
            first_non_trivia = Some(t.kind());
        }
    }
    // EOF terminates the final line. Mirrors `green::tl_section_marker_check`.
    if first_non_trivia == Some(crate::SyntaxKind::STAR)
        && let Some(ls) = line_start
    {
        let end: u32 = child.text_range().end().into();
        let span = Span::new((ls + bom_offset) as usize, (end + bom_offset) as usize);
        out.push(Spanned::new(String::new(), span));
    }
}

// `extract_top_level_comments` and `extract_currency_occurrences`
// are folded into `walk_descendants_once` above - see the
// comments there for the column-0 / ERROR_NODE-exclusion rules.

// ---- Token parsing helpers -------------------------------------

/// Parse a date token, accepting the same shapes as the legacy
/// parser: canonical `YYYY-MM-DD`, slash-separated `YYYY/M/D`,
/// and single-digit month/day. Returns `None` when the token
/// can't be turned into a real calendar date (invalid month,
/// invalid day for the given month, etc.).
pub(super) fn parse_date_token(text: &str) -> Option<NaiveDate> {
    // Fast path: canonical "YYYY-MM-DD".
    if text.len() == 10
        && text.as_bytes()[4] == b'-'
        && text.as_bytes()[7] == b'-'
        && let (Ok(y), Ok(m), Ok(d)) = (
            text[0..4].parse::<i32>(),
            text[5..7].parse::<u32>(),
            text[8..10].parse::<u32>(),
        )
    {
        return naive_date(y, m, d);
    }
    // Slow path: share legacy's normalizer so single-digit
    // month/day (`2024-1-15`, `2024-01-5`) and slash separators
    // are accepted everywhere the legacy parser accepts them.
    crate::diagnostics::normalize_date_str(text)
        .parse::<NaiveDate>()
        .ok()
}

/// Parse a directive's `DATE` token. On success returns the
/// `NaiveDate`; on a token whose calendar values don't form a
/// real date (`2024-13-01`, Feb 29 in a non-leap year) emits
/// `InvalidDateValue` with the legacy parser's human-readable
/// message and returns `None`. This mirrors
/// `parser.rs:181-182` so the CST and legacy parsers surface the
/// same diagnostics for malformed dates in directive position.
fn parse_directive_date(
    date_tok: &ast::Date,
    errors: &mut Vec<crate::ParseError>,
    bom_offset: u32,
) -> Option<NaiveDate> {
    let text = date_tok.text();
    if let Some(d) = parse_date_token(text) {
        return Some(d);
    }
    let range = date_tok.syntax().text_range();
    let start: u32 = range.start().into();
    let end: u32 = range.end().into();
    let span = Span::new((start + bom_offset) as usize, (end + bom_offset) as usize);
    errors.push(crate::ParseError::new(
        crate::ParseErrorKind::InvalidDateValue(crate::diagnostics::describe_invalid_date(text)),
        span,
    ));
    None
}

/// Decode a `STRING` token's text (with surrounding quotes) into its semantic
/// value: quotes stripped, escapes decoded (`\"`→`"`, `\\`→`\`, `\n`/`\t`/`\r`,
/// unknown escape drops the backslash). `None` if not a well-formed quoted
/// string. Text-based so both the red ([`ast::StringLit::text_decoded`]) and
/// green conversion paths share one source of truth.
pub(super) fn decode_string_token(text: &str) -> Option<String> {
    let bytes = text.as_bytes();
    if bytes.len() < 2 || bytes[0] != b'"' || bytes[bytes.len() - 1] != b'"' {
        return None;
    }
    let raw = &text[1..text.len() - 1];
    if !raw.contains('\\') {
        return Some(raw.to_string());
    }
    let mut out = String::with_capacity(raw.len());
    let mut chars = raw.chars();
    while let Some(c) = chars.next() {
        if c != '\\' {
            out.push(c);
            continue;
        }
        match chars.next() {
            Some('"') => out.push('"'),
            Some('\\') => out.push('\\'),
            Some('n') => out.push('\n'),
            Some('t') => out.push('\t'),
            Some('r') => out.push('\r'),
            Some(other) => out.push(other),
            None => {}
        }
    }
    Some(out)
}

/// Parse a numeric token. Tolerates leading sign and thousands-
/// separator commas (legacy parser drops them).
pub(super) fn parse_decimal_token(text: &str) -> Option<Decimal> {
    use std::str::FromStr;
    let cleaned: String;
    let s = if text.contains(',') {
        cleaned = text.replace(',', "");
        cleaned.as_str()
    } else {
        text
    };
    Decimal::from_str(s).ok()
}

/// Choose `Int` vs `Number` for a numeric metadata literal.
///
/// Beancount represents integer metadata (`key: 42`) as an int and decimal
/// metadata (`key: 42.0`) as a `Decimal`. `text` is the original (unsigned)
/// NUMBER token: a literal with no `.` or exponent that fits in `i64` becomes
/// `Int`; decimals, exponents, and out-of-range integers stay `Number`. `value`
/// is the parsed (and sign-applied) magnitude. (Thousands-separator commas are
/// irrelevant to integer-ness, so they aren't stripped before the check.)
pub(super) fn number_meta_value(text: &str, value: Decimal) -> MetaValue {
    use rust_decimal::prelude::ToPrimitive;
    if !text.contains('.')
        && !text.contains('e')
        && !text.contains('E')
        && let Some(i) = value.to_i64()
    {
        return MetaValue::Int(i);
    }
    MetaValue::Number(value)
}

// ---- Span helpers ----------------------------------------------

/// Convert a CST node's [`rowan::TextRange`] (relative to the
/// post-BOM source frame) into a [`Span`] in the original-source
/// frame.
fn node_span(node: &crate::SyntaxNode, bom_offset: u32) -> Span {
    let range = node.text_range();
    let start: u32 = range.start().into();
    let end: u32 = range.end().into();
    Span::new((start + bom_offset) as usize, (end + bom_offset) as usize)
}

/// Trivia kinds that don't count toward a span's start/end when
/// matching the legacy parser's span convention.
///
/// Covers WHITESPACE / NEWLINE plus EVERY comment-trivia kind
/// (`COMMENT`, `PERCENT_COMMENT`, `SHEBANG`, `EMACS_DIRECTIVE`)
/// so files with ledger-style `%` comments or org-mode
/// `#!`/`#+` lines have the same span/header-tracking behavior
/// as files with only `;` comments. Mirrors
/// `SyntaxKind::is_trivia()` minus `BOM` - a mid-file BOM byte
/// is an error to surface (`extract_inline_token_errors` /
/// `classify_recovery_error`), not trivia to silently skip.
pub(super) 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
    )
}

/// Span policy for `Posting`: the legacy parser ends the posting
/// span at the position just before the line's terminating
/// NEWLINE. The CST node's range INCLUDES the terminator
/// NEWLINE; trim it by using the NEWLINE token's start position.
/// We look at the FIRST direct-child NEWLINE token because
/// posting-attached metadata sub-lines (which have their own
/// inner NEWLINEs) come after the line terminator and shouldn't
/// extend the posting-line span.
fn posting_span(node: &crate::SyntaxNode, bom_offset: u32) -> Span {
    let range = node.text_range();
    let start: u32 = range.start().into();
    let end_raw: u32 = range.end().into();
    // Postings have no inter-directive leading trivia: their
    // first direct-child NEWLINE IS the terminator.
    let end = node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
        .find(|t| t.kind() == crate::SyntaxKind::NEWLINE)
        .map_or(end_raw, |t| u32::from(t.text_range().start()));
    Span::new((start + bom_offset) as usize, (end + bom_offset) as usize)
}

/// Span policy for non-Directive single-line constructs that
/// participate in inter-directive trivia attachment (Option,
/// Include, Plugin). Unlike Posting these may have leading
/// trivia (blank-line NEWLINEs, comments) inside the node from
/// the Directive-Terminator Rule. Start at the first non-trivia
/// content token; end at the first NEWLINE after that.
fn single_line_directive_span(node: &crate::SyntaxNode, bom_offset: u32) -> Span {
    let range = node.text_range();
    let start_raw: u32 = range.start().into();
    let end_raw: u32 = range.end().into();
    let mut content_start: Option<u32> = None;
    let mut terminator: Option<u32> = None;
    for t in node
        .children_with_tokens()
        .filter_map(rowan::NodeOrToken::into_token)
    {
        if content_start.is_none() {
            if !is_trivia_kind(t.kind()) {
                content_start = Some(u32::from(t.text_range().start()));
            }
        } else if t.kind() == crate::SyntaxKind::NEWLINE {
            terminator = Some(u32::from(t.text_range().start()));
            break;
        }
    }
    let start = content_start.unwrap_or(start_raw);
    let end = terminator.unwrap_or(end_raw);
    Span::new((start + bom_offset) as usize, (end + bom_offset) as usize)
}

/// Span policy for top-level directives: legacy directives start
/// at the first content character (skipping leading trivia from
/// the Directive-Terminator Rule) and extend through any
/// inter-directive trivia up to where the NEXT directive begins.
/// Computed in a post-pass since each directive's end depends on
/// the next one's start.
fn fixup_directive_spans(
    source_file: &SourceFile,
    bom_offset: u32,
    converted_nodes: &[crate::SyntaxNode],
    directives: &mut [Spanned<Directive>],
) {
    debug_assert_eq!(
        converted_nodes.len(),
        directives.len(),
        "converted_nodes and directives must be parallel arrays"
    );

    // Walk EVERY top-level Directive-castable child (including
    // pushtag/poptag/pushmeta/popmeta that we filter out of the
    // ParseResult) so the "next directive's start" boundary used
    // for span end-fixup matches the legacy parser: there, each
    // visible directive's span ends at the next /input/
    // directive's start, regardless of whether that next
    // directive is preserved.
    let all_starts: Vec<(usize, usize)> = source_file
        .syntax()
        .children()
        .filter(|n| ast::Directive::can_cast(n.kind()))
        .map(|n| {
            let raw_start: u32 = n.text_range().start().into();
            let content_start = n
                .descendants_with_tokens()
                .filter_map(rowan::NodeOrToken::into_token)
                .find(|t| !is_trivia_kind(t.kind()))
                .map_or_else(
                    || (raw_start + bom_offset) as usize,
                    |t| (u32::from(t.text_range().start()) + bom_offset) as usize,
                );
            ((raw_start + bom_offset) as usize, content_start)
        })
        .collect();

    // `all_starts` is built from `children()`, i.e. siblings in document
    // order, and sibling text ranges are disjoint and increasing — so it is
    // sorted ascending on `raw_start` and every key is unique. The lookup
    // below binary-searches on that.
    debug_assert!(
        all_starts.windows(2).all(|w| w[0].0 < w[1].0),
        "all_starts must be strictly ascending by raw_start for the binary \
         search below; sibling text ranges are disjoint and increasing, so a \
         failure here means the enumeration is no longer document-ordered",
    );

    let source_end: usize =
        (u32::from(source_file.syntax().text_range().end()) + bom_offset) as usize;

    // For each converted directive, find its position in the all
    // list by raw_start (which is unique per CST node), then use
    // the NEXT all_starts content_start as its span end.
    //
    // INVARIANT: every node in `converted_nodes` was yielded by
    // `source_file.directives()`, which is the same iteration
    // `all_starts` filters from. So `position` always succeeds in
    // well-formed input. Falling back to the node's own
    // `text_range` rather than panicking keeps the parser usable
    // when a future change to the typed-AST surface ever de-syncs
    // those two enumerations - a `panic!()` reachable from user
    // input is a `#![forbid(unsafe_code)]`-class regression for an
    // LSP/WASM consumer.
    for (i, spanned) in directives.iter_mut().enumerate() {
        let node = &converted_nodes[i];
        let raw_start: usize = (u32::from(node.text_range().start()) + bom_offset) as usize;
        let node_end: usize = (u32::from(node.text_range().end()) + bom_offset) as usize;
        // Binary search, NOT a linear `position()` scan. This loop runs once
        // per directive over an `all_starts` that has one entry per
        // directive, so a linear probe made span fixup O(N^2) in the
        // directive count. Measured on the `simple` profiling shape: 10x the
        // transactions cost 21.8x the instructions, and cachegrind put ~40%
        // of a 20k-transaction run inside this one scan and its slice-iterator
        // internals. It is invisible on small inputs, which is why it sat
        // here — at 2k transactions the same scan is under 2%.
        //
        // Semantics are unchanged: keys are unique (see the debug_assert
        // above), so where `position` found the sole match, `binary_search`
        // finds the same one, and a miss still falls through to the
        // defensive branch below rather than panicking.
        if let Ok(pos) = all_starts.binary_search_by_key(&raw_start, |(rs, _)| *rs) {
            let start = all_starts[pos].1;
            let end = all_starts
                .get(pos + 1)
                .map_or(source_end, |(_, content)| *content);
            spanned.span = Span::new(start, end);
        } else {
            // Defensive fallback: match the success-path
            // convention by also trimming leading trivia. Without
            // this trim the fallback span would underline blank
            // lines / column-0 comments above the directive when
            // LSP/miette renders the diagnostic, even though the
            // directive itself starts further down.
            let content_start = node
                .descendants_with_tokens()
                .filter_map(rowan::NodeOrToken::into_token)
                .find(|t| !is_trivia_kind(t.kind()))
                .map_or(raw_start, |t| {
                    (u32::from(t.text_range().start()) + bom_offset) as usize
                });
            spanned.span = Span::new(content_start, node_end);
        }
    }
}

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

    /// A directive's span ends at the next *input* directive's content start,
    /// INCLUDING directives filtered out of the `ParseResult`
    /// (`pushtag`/`poptag`/`pushmeta`/`popmeta`).
    ///
    /// This is the case that makes `all_starts` longer than `directives`, so
    /// it is the one that breaks if the lookup ever stops keying on the CST
    /// node's own start — e.g. "optimizing" the binary search into an index
    /// into `directives`, which would end the `open` span at the transaction
    /// instead of at the `pushtag` between them.
    #[test]
    fn spans_end_at_the_next_input_directive_even_when_it_is_filtered_out() {
        let src = "2024-01-01 open Assets:Bank USD\n\
                   pushtag #trip\n\
                   2024-01-02 * \"a\"\n  Assets:Bank 1 USD\n  Assets:Other\n\
                   poptag #trip\n\
                   2024-01-03 close Assets:Bank\n";
        let parsed = crate::parse(src);

        let at = |needle: &str| src.find(needle).expect("fixture contains it");
        let spans: Vec<(usize, usize)> = parsed
            .directives
            .iter()
            .map(|d| (d.span.start, d.span.end))
            .collect();

        assert_eq!(
            spans,
            vec![
                // `open` stops at `pushtag`, NOT at the transaction.
                (0, at("pushtag")),
                // the transaction stops at `poptag`, NOT at `close`.
                (at("2024-01-02"), at("poptag")),
                // the last directive runs to end of source.
                (at("2024-01-03"), src.len()),
            ],
            "pushtag/poptag are filtered from `directives` but still bound the \
             preceding directive's span",
        );
    }

    /// Match a `SyntaxError` by message prefix rather than by `Debug` output.
    /// The `Debug` rendering of `ParseErrorKind` can change without any
    /// semantic change, and an assertion that reads it would fail for no
    /// reason; the variant plus the message prefix is the real contract.
    fn has_syntax_error(result: &ParseResult, prefix: &str) -> bool {
        result.errors.iter().any(
            |e| matches!(&e.kind, crate::ParseErrorKind::SyntaxError(m) if m.starts_with(prefix)),
        )
    }

    fn assert_directive_count(result: &ParseResult, expected: usize) {
        assert_eq!(
            result.directives.len(),
            expected,
            "directive count mismatch: {:#?}",
            result.directives
        );
    }

    #[test]
    fn open_directive_basic() {
        let src = "2024-01-15 open Assets:Cash\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open, got {:?}", result.directives[0].value);
        };
        assert_eq!(open.date, naive_date(2024, 1, 15).unwrap());
        assert_eq!(open.account.as_str(), "Assets:Cash");
        assert!(open.currencies.is_empty());
        assert!(open.booking.is_none());
        assert!(open.meta.is_empty());
    }

    #[test]
    fn open_directive_with_currencies_and_booking() {
        let src = "2024-01-15 open Assets:Brokerage USD,EUR \"STRICT\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        let currencies: Vec<&str> = open.currencies.iter().map(Currency::as_str).collect();
        assert_eq!(currencies, vec!["USD", "EUR"]);
        assert_eq!(open.booking.as_deref(), Some("STRICT"));
    }

    #[test]
    fn open_directive_with_metadata() {
        let src = "2024-01-15 open Assets:Cash\n  note: \"main checking\"\n  number: 42\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        assert_eq!(
            open.meta.get("note"),
            Some(&MetaValue::String("main checking".to_string()))
        );
        assert_eq!(
            open.meta.get("number"),
            // Unquoted integer metadata is now `Int`, not `Number`.
            Some(&MetaValue::Int(42))
        );
    }

    #[test]
    fn close_directive_basic() {
        let src = "2024-12-31 close Assets:Cash\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Close(close) = &result.directives[0].value else {
            panic!("expected Close, got {:?}", result.directives[0].value);
        };
        assert_eq!(close.date, naive_date(2024, 12, 31).unwrap());
        assert_eq!(close.account.as_str(), "Assets:Cash");
    }

    #[test]
    fn commodity_directive_basic() {
        let src = "2024-01-01 commodity HOOL\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Commodity(c) = &result.directives[0].value else {
            panic!("expected Commodity");
        };
        assert_eq!(c.currency.as_str(), "HOOL");
    }

    #[test]
    fn bom_offset_is_included_in_spans() {
        let src = "\u{FEFF}2024-01-15 open Assets:Cash\n";
        let result = parse_via_cst(src);
        assert!(result.has_leading_bom);
        let span = result.directives[0].span;
        assert_eq!(span.start, 3, "span should include BOM offset");
    }

    #[test]
    fn note_directive_basic() {
        let src = "2024-01-15 note Assets:Cash \"deposit received\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Note(note) = &result.directives[0].value else {
            panic!("expected Note");
        };
        assert_eq!(note.date, naive_date(2024, 1, 15).unwrap());
        assert_eq!(note.account.as_str(), "Assets:Cash");
        assert_eq!(note.comment, "deposit received");
    }

    #[test]
    fn document_directive_basic() {
        let src = "2024-01-15 document Assets:Cash \"/path/to/file.pdf\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Document(d) = &result.directives[0].value else {
            panic!("expected Document");
        };
        assert_eq!(d.account.as_str(), "Assets:Cash");
        assert_eq!(d.path, "/path/to/file.pdf");
        // tags/links currently unimplemented - pin as empty.
        assert!(d.tags.is_empty());
        assert!(d.links.is_empty());
    }

    #[test]
    fn event_directive_basic() {
        let src = "2024-01-15 event \"location\" \"Berlin\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Event(e) = &result.directives[0].value else {
            panic!("expected Event");
        };
        assert_eq!(e.event_type, "location");
        assert_eq!(e.value, "Berlin");
    }

    #[test]
    fn query_directive_basic() {
        let src = "2024-01-15 query \"income\" \"SELECT account, sum(position)\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Query(q) = &result.directives[0].value else {
            panic!("expected Query");
        };
        assert_eq!(q.name, "income");
        assert_eq!(q.query, "SELECT account, sum(position)");
    }

    #[test]
    fn price_directive_basic() {
        let src = "2024-01-15 price USD 1.10 EUR\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Price(p) = &result.directives[0].value else {
            panic!("expected Price");
        };
        assert_eq!(p.currency.as_str(), "USD");
        assert_eq!(p.amount.number, Decimal::new(110, 2));
        assert_eq!(p.amount.currency.as_str(), "EUR");
    }

    #[test]
    fn balance_directive_basic() {
        let src = "2024-06-30 balance Assets:Cash 100.00 USD\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Balance(b) = &result.directives[0].value else {
            panic!("expected Balance");
        };
        assert_eq!(b.account.as_str(), "Assets:Cash");
        assert_eq!(b.amount.number, Decimal::new(10000, 2));
        assert_eq!(b.amount.currency.as_str(), "USD");
        assert!(b.tolerance.is_none());
    }

    #[test]
    fn balance_directive_with_explicit_tolerance() {
        let src = "2024-06-30 balance Assets:Cash 100.00 ~ 0.05 USD\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Balance(b) = &result.directives[0].value else {
            panic!("expected Balance");
        };
        assert_eq!(b.amount.number, Decimal::new(10000, 2));
        assert_eq!(b.tolerance, Some(Decimal::new(5, 2)));
    }

    #[test]
    fn pad_directive_basic() {
        let src = "2024-01-01 pad Assets:Cash Equity:Opening-Balances\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Pad(p) = &result.directives[0].value else {
            panic!("expected Pad");
        };
        assert_eq!(p.account.as_str(), "Assets:Cash");
        assert_eq!(p.source_account.as_str(), "Equity:Opening-Balances");
    }

    #[test]
    fn custom_directive_basic() {
        let src = "2024-01-01 custom \"budget\" \"food\" 500 USD\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Custom(c) = &result.directives[0].value else {
            panic!("expected Custom");
        };
        assert_eq!(c.custom_type, "budget");
        assert_eq!(c.values.len(), 2);
        assert_eq!(c.values[0], MetaValue::String("food".to_string()));
        // 500 USD becomes an Amount (NUMBER + CURRENCY adjacent).
        let MetaValue::Amount(amt) = &c.values[1] else {
            panic!("expected Amount, got {:?}", c.values[1]);
        };
        assert_eq!(amt.number, Decimal::from(500));
        assert_eq!(amt.currency.as_str(), "USD");
    }

    #[test]
    fn custom_directive_heterogeneous_values() {
        let src = "2024-01-01 custom \"test\" Assets:Cash TRUE 42 2024-06-15\n";
        let result = parse_via_cst(src);
        let Directive::Custom(c) = &result.directives[0].value else {
            panic!("expected Custom");
        };
        assert_eq!(c.values.len(), 4);
        assert!(matches!(c.values[0], MetaValue::Account(_)));
        assert_eq!(c.values[1], MetaValue::Bool(true));
        assert_eq!(c.values[2], MetaValue::Int(42));
        assert!(matches!(c.values[3], MetaValue::Date(_)));
    }

    #[test]
    fn number_meta_value_int_vs_decimal_discriminator() {
        use rust_decimal_macros::dec;
        // Integer literals -> Int (the token text is unsigned; `value` carries
        // the sign, e.g. `precision: -1` parses the token "1" with value -1).
        assert_eq!(number_meta_value("42", dec!(42)), MetaValue::Int(42));
        assert_eq!(number_meta_value("0", dec!(0)), MetaValue::Int(0));
        assert_eq!(number_meta_value("1", dec!(-1)), MetaValue::Int(-1));
        // Decimal point -> Number.
        assert_eq!(
            number_meta_value("42.0", dec!(42.0)),
            MetaValue::Number(dec!(42.0))
        );
        // Exponent -> Number. The lexer doesn't currently emit exponent NUMBER
        // tokens, so this isn't reachable from real input today; it pins the
        // helper's `e`/`E` guard against a future lexer that does.
        assert_eq!(
            number_meta_value("1e3", dec!(1000)),
            MetaValue::Number(dec!(1000))
        );
        // i64 overflow stays Number (within Decimal's range).
        let huge = "99999999999999999999999999";
        let huge_dec = Decimal::from_str_exact(huge).unwrap();
        assert_eq!(
            number_meta_value(huge, huge_dec),
            MetaValue::Number(huge_dec)
        );
    }

    #[test]
    fn option_directive_populates_options_field() {
        let src = "option \"title\" \"My Ledger\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 0);
        assert_eq!(result.options.len(), 1);
        assert_eq!(result.options[0].0, "title");
        assert_eq!(result.options[0].1, "My Ledger");
    }

    #[test]
    fn include_directive_populates_includes_field() {
        let src = "include \"shared.beancount\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 0);
        assert_eq!(result.includes.len(), 1);
        assert_eq!(result.includes[0].0, "shared.beancount");
    }

    #[test]
    fn plugin_directive_with_config() {
        let src = "plugin \"my.plugin\" \"cfg\"\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 0);
        assert_eq!(result.plugins.len(), 1);
        assert_eq!(result.plugins[0].0, "my.plugin");
        assert_eq!(result.plugins[0].1.as_deref(), Some("cfg"));
    }

    #[test]
    fn plugin_directive_without_config() {
        let src = "plugin \"my.plugin\"\n";
        let result = parse_via_cst(src);
        assert_eq!(result.plugins.len(), 1);
        assert_eq!(result.plugins[0].0, "my.plugin");
        assert!(result.plugins[0].1.is_none());
    }

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

    #[test]
    fn transaction_basic_two_postings() {
        let src = "2024-01-15 * \"Coffee Shop\" \"Morning coffee\"\n  \
                   Expenses:Food:Coffee  5.00 USD\n  \
                   Assets:Cash\n";
        let result = parse_via_cst(src);
        assert_directive_count(&result, 1);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(t.date, naive_date(2024, 1, 15).unwrap());
        assert_eq!(t.flag, '*');
        assert_eq!(
            t.payee.as_ref().map(InternedStr::as_str),
            Some("Coffee Shop")
        );
        assert_eq!(t.narration.as_str(), "Morning coffee");
        assert_eq!(t.postings.len(), 2);

        let p0 = &t.postings[0].value;
        assert_eq!(p0.account.as_str(), "Expenses:Food:Coffee");
        let Some(IncompleteAmount::Complete(amt)) = &p0.units else {
            panic!("expected complete units, got {:?}", p0.units);
        };
        assert_eq!(amt.number, Decimal::new(500, 2));
        assert_eq!(amt.currency.as_str(), "USD");

        let p1 = &t.postings[1].value;
        assert_eq!(p1.account.as_str(), "Assets:Cash");
        assert!(p1.units.is_none(), "auto-posting has no units");
    }

    #[test]
    fn transaction_narration_only_no_payee() {
        let src = "2024-01-15 ! \"Pending\"\n  Assets:Cash  -5 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(t.flag, '!');
        assert!(t.payee.is_none());
        assert_eq!(t.narration.as_str(), "Pending");
    }

    #[test]
    fn transaction_three_plus_header_strings_surface_last_as_narration() {
        // A header with 3+ strings is ambiguous (the grammar caps payee+narration
        // at two); the lossless CST still keeps all of them, so the typed surface
        // drops the payee and surfaces only the LAST string as narration. Locks
        // the `(Some, Some, Some(c)) => it.last().unwrap_or(c)` arm.
        let src = "2024-01-15 * \"a\" \"b\" \"c\"\n  Assets:Cash  -5 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert!(t.payee.is_none(), "3+ strings drop the payee");
        assert_eq!(t.narration.as_str(), "c", "last string becomes narration");
    }

    #[test]
    fn transaction_implied_flag_via_leading_string() {
        let src = "2024-01-15 \"Implied\"\n  Assets:Cash  -5 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(t.flag, '*', "implied flag defaults to *");
    }

    #[test]
    fn transaction_with_tags_and_links() {
        let src = "2024-01-15 * \"Coffee\" #daily ^trip1\n  Assets:Cash  -5 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(t.tags.len(), 1);
        assert_eq!(t.tags[0].as_str(), "daily");
        assert_eq!(t.links.len(), 1);
        assert_eq!(t.links[0].as_str(), "trip1");
    }

    #[test]
    fn transaction_with_signed_amount() {
        let src = "2024-01-15 * \"x\"\n  Assets:Cash  -5.00 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete units");
        };
        assert_eq!(amt.number, Decimal::new(-500, 2));
    }

    #[test]
    fn transaction_with_posting_flag() {
        let src = "2024-01-15 * \"x\"\n  ! Assets:Cash  -5 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(t.postings[0].value.flag, Some('!'));
    }

    #[test]
    fn transaction_with_cost_spec_per_unit() {
        let src = "2024-01-15 * \"buy\"\n  \
                   Assets:Inv  10 HOOL {500.00 USD}\n  \
                   Assets:Cash\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let cost = t.postings[0].value.cost.as_ref().expect("cost spec");
        assert!(!cost.merge);
        let Some(CostNumber::PerUnit { value }) = &cost.number else {
            panic!("expected PerUnit");
        };
        assert_eq!(*value, Decimal::new(50000, 2));
        assert_eq!(cost.currency.as_ref().unwrap().as_str(), "USD");
    }

    #[test]
    fn transaction_with_cost_spec_total() {
        let src = "2024-01-15 * \"buy\"\n  \
                   Assets:Inv  10 HOOL {{5000 USD}}\n  \
                   Assets:Cash\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let cost = t.postings[0].value.cost.as_ref().expect("cost spec");
        let Some(CostNumber::Total { value }) = &cost.number else {
            panic!("expected Total");
        };
        assert_eq!(*value, Decimal::from(5000));
    }

    #[test]
    fn transaction_with_price_annotation_unit() {
        let src = "2024-01-15 * \"buy\"\n  \
                   Assets:Inv  10 HOOL @ 510 USD\n  \
                   Assets:Cash\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let price = t.postings[0]
            .value
            .price
            .as_ref()
            .expect("price annotation");
        assert!(price.is_unit());
        let Some(IncompleteAmount::Complete(amt)) = &price.amount else {
            panic!("expected complete price amount");
        };
        assert_eq!(amt.number, Decimal::from(510));
        assert_eq!(amt.currency.as_str(), "USD");
    }

    #[test]
    fn transaction_with_price_annotation_total() {
        let src = "2024-01-15 * \"buy\"\n  \
                   Assets:Inv  10 HOOL @@ 5100 USD\n  \
                   Assets:Cash\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let price = t.postings[0]
            .value
            .price
            .as_ref()
            .expect("price annotation");
        assert!(!price.is_unit(), "@@ is total form");
    }

    // ---- regression tests for review findings (#1281) ----------

    #[test]
    fn document_directive_preserves_tags_and_links() {
        // Finding 1: convert_document was filling tags/links empty
        // unconditionally. Legacy parse_document_directive collects
        // trailing `#tag` / `^link` tokens after the path STRING.
        let src = "2024-06-01 document Assets:Bank \"stmt.pdf\" #quarter1 ^scan42 #urgent\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Document(doc) = &result.directives[0].value else {
            panic!("expected Document");
        };
        let tags: Vec<&str> = doc.tags.iter().map(Tag::as_str).collect();
        let links: Vec<&str> = doc.links.iter().map(Link::as_str).collect();
        assert_eq!(tags, vec!["quarter1", "urgent"]);
        assert_eq!(links, vec!["scan42"]);
    }

    #[test]
    fn open_directive_rejects_invalid_booking_method() {
        // Finding 2: convert_open accepted any booking string; legacy
        // validates against [FIFO, STRICT, STRICT_WITH_SIZE, LIFO,
        // HIFO, NONE, AVERAGE] and on mismatch drops the directive
        // AND emits InvalidBookingMethod.
        let src = "2024-01-01 open Assets:Bank USD \"GARBAGE\"\n";
        let result = parse_via_cst(src);
        assert_eq!(result.directives.len(), 0, "directive should be dropped");
        assert_eq!(result.errors.len(), 1);
        let err = &result.errors[0];
        assert!(
            matches!(
                &err.kind,
                crate::ParseErrorKind::InvalidBookingMethod(s) if s == "GARBAGE"
            ),
            "expected InvalidBookingMethod, got {:?}",
            err.kind,
        );
    }

    #[test]
    fn open_directive_accepts_all_valid_booking_methods() {
        for method in VALID_BOOKING_METHODS {
            let src = format!("2024-01-01 open Assets:Bank USD \"{method}\"\n");
            let result = parse_via_cst(&src);
            assert!(
                result.errors.is_empty(),
                "{method} rejected: {:?}",
                result.errors
            );
            let Directive::Open(open) = &result.directives[0].value else {
                panic!("{method}: expected Open");
            };
            assert_eq!(open.booking.as_deref(), Some(*method));
        }
    }

    #[test]
    fn unclosed_pushtag_at_eof_emits_diagnostic() {
        // Finding 3: legacy emits one UnclosedPushtag per leftover
        // tag at EOF, pointing at the originating push directive.
        let src = "pushtag #active\n2024-01-01 open Assets:Bank USD\n";
        let result = parse_via_cst(src);
        let unclosed: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::UnclosedPushtag(t) => Some(t.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(unclosed, vec!["active".to_string()]);
    }

    #[test]
    fn unclosed_pushmeta_at_eof_emits_diagnostic() {
        // Finding 4: same as pushtag, for pushmeta.
        let src = "pushmeta location: \"NYC\"\n2024-01-01 open Assets:Bank USD\n";
        let result = parse_via_cst(src);
        let unclosed: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::UnclosedPushmeta(k) => Some(k.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(unclosed, vec!["location".to_string()]);
    }

    #[test]
    fn invalid_poptag_on_mismatch_emits_diagnostic() {
        // Finding 5: poptag for a tag never pushed should error,
        // not silently no-op.
        let src = "pushtag #foo\npoptag #bar\npoptag #foo\n";
        let result = parse_via_cst(src);
        let mismatches: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::InvalidPoptag(t) => Some(t.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(mismatches, vec!["bar".to_string()]);
        // and the matching #foo poptag should leave NO unclosed
        // diagnostic - i.e. the stack is empty after the matched pop.
        let leftover: Vec<_> = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::UnclosedPushtag(_)))
            .collect();
        assert!(leftover.is_empty(), "unexpected leftover: {leftover:?}");
    }

    #[test]
    fn invalid_popmeta_on_mismatch_emits_diagnostic() {
        // Finding 6: popmeta for a key never pushed should error,
        // not silently no-op. Also checks Vec-stack shadow semantics:
        // pushmeta x: 1; pushmeta x: 2; popmeta x leaves x=1 active.
        let src = "pushmeta location: \"NYC\"\npopmeta nope:\npopmeta location:\n";
        let result = parse_via_cst(src);
        let mismatches: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::InvalidPopmeta(k) => Some(k.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(mismatches, vec!["nope".to_string()]);
        let leftover: Vec<_> = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::UnclosedPushmeta(_)))
            .collect();
        assert!(leftover.is_empty(), "unexpected leftover: {leftover:?}");
    }

    #[test]
    fn pushmeta_shadow_pop_restores_prior_value() {
        // Vec-stack semantics (the reason meta_stack isn't a HashMap):
        // shadow-pop must restore the prior value, not delete the key.
        let src = "pushmeta loc: \"NYC\"\n\
                   pushmeta loc: \"LDN\"\n\
                   popmeta loc:\n\
                   2024-01-01 open Assets:Bank USD\n\
                   popmeta loc:\n";
        let result = parse_via_cst(src);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        assert_eq!(
            open.meta.get("loc"),
            Some(&MetaValue::String("NYC".to_string())),
            "shadow pop should restore NYC, got {:?}",
            open.meta.get("loc"),
        );
    }

    #[test]
    fn error_recovery_classifies_bom_in_directive_body() {
        // Finding 7: error-recovery path should distinguish BOM-in-
        // line from a generic SyntaxError so users see the
        // BOM-removal hint instead of "unexpected input".
        let src = "garbage\u{FEFF}content\n";
        let result = parse_via_cst(src);
        let bom_errors: Vec<_> = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::BomInDirectiveBody))
            .collect();
        assert_eq!(bom_errors.len(), 1, "errors: {:?}", result.errors);
        assert!(
            bom_errors[0].hint.is_some(),
            "BomInDirectiveBody should carry BOM_REMOVAL_HINT",
        );
    }

    #[test]
    fn error_recovery_emits_both_invalid_account_and_bom_for_dual_line() {
        // Round-2 finding: legacy `parser.rs:2258-2263` emits a
        // SECONDARY `BomInDirectiveBody` whenever the line ALSO
        // contains a BOM byte and the primary diagnostic isn't
        // BOM itself. Without this, a Windows-exported file with
        // a Unicode account AND an internal BOM loses the BOM
        // hint entirely.
        let src = "garbage Assets:Café\u{FEFF}content\n";
        let result = parse_via_cst(src);
        let invalid_account_count = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::InvalidAccount(_)))
            .count();
        let bom_count = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::BomInDirectiveBody))
            .count();
        assert_eq!(
            invalid_account_count, 1,
            "expected one InvalidAccount: {:?}",
            result.errors
        );
        assert_eq!(
            bom_count, 1,
            "expected secondary BomInDirectiveBody: {:?}",
            result.errors
        );
        // The secondary BOM diagnostic must carry the hint so
        // miette renders the remediation step.
        let bom_err = result
            .errors
            .iter()
            .find(|e| matches!(e.kind, crate::ParseErrorKind::BomInDirectiveBody))
            .unwrap();
        assert!(bom_err.hint.is_some());
    }

    #[test]
    fn error_recovery_classifies_unicode_account() {
        // Finding 7: a Unicode-character account name (Assets:Café)
        // should surface as InvalidAccount, not generic SyntaxError.
        // We embed it in a malformed line so the parser routes to
        // the error-recovery path.
        let src = "garbage Assets:Café content\n";
        let result = parse_via_cst(src);
        let unicode_errors: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::InvalidAccount(s) => Some(s.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(unicode_errors, vec!["Assets:Café".to_string()]);
    }

    #[test]
    fn transaction_with_pipe_emits_deprecated_pipe_symbol() {
        // Finding 7 (transaction path): legacy emits
        // DeprecatedPipeSymbol when a `|` separates payee/narration.
        let src = "2024-01-15 * \"Acme\" | \"invoice\"\n  Assets:Cash  -5 USD\n  Expenses:X\n";
        let result = parse_via_cst(src);
        let pipe_count = result
            .errors
            .iter()
            .filter(|e| matches!(e.kind, crate::ParseErrorKind::DeprecatedPipeSymbol))
            .count();
        assert_eq!(pipe_count, 1, "errors: {:?}", result.errors);
        // and the transaction itself is kept (legacy behavior).
        assert_eq!(result.directives.len(), 1);
    }

    #[test]
    fn transaction_trailing_comments_after_final_posting() {
        // Finding 8: comments that appear AFTER the last posting
        // but inside the transaction body belong to
        // Transaction::trailing_comments, not lost.
        let src = "2024-01-15 * \"x\"\n  \
                   Assets:Cash  -5 USD\n  \
                   Expenses:X\n  \
                   ; trailing one\n  \
                   ; trailing two\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(
            t.trailing_comments.len(),
            2,
            "got: {:?}",
            t.trailing_comments
        );
        assert!(t.trailing_comments[0].contains("trailing one"));
        assert!(t.trailing_comments[1].contains("trailing two"));
    }

    // ---- arithmetic AMOUNT evaluation (phase 3.7 flip blocker) -

    #[test]
    fn posting_amount_evaluates_division() {
        // Regression for `test_arithmetic_expressions_consistency`:
        // `120 / 3 USD` must evaluate to 40 USD so the transaction
        // balances. Without this the CST flip breaks every ledger
        // using arithmetic split syntax.
        let src = "2024-01-15 * \"split\"\n  \
                   Expenses:Food   120 / 3 USD\n  \
                   Assets:Bank    -40 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete amount on posting 0");
        };
        assert_eq!(amt.number, Decimal::from(40));
        assert_eq!(amt.currency.as_str(), "USD");
    }

    #[test]
    fn posting_amount_evaluates_addition_and_multiplication_precedence() {
        // `2 + 3 * 4 USD` = 14 USD (standard precedence).
        let src = "2024-01-15 * \"x\"\n  \
                   Expenses:X   2 + 3 * 4 USD\n  \
                   Assets:Y   -14 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete amount");
        };
        assert_eq!(amt.number, Decimal::from(14));
    }

    #[test]
    fn posting_amount_evaluates_parens_override_precedence() {
        // `(2 + 3) * 4 USD` = 20 USD.
        let src = "2024-01-15 * \"x\"\n  \
                   Expenses:X   (2 + 3) * 4 USD\n  \
                   Assets:Y   -20 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete amount");
        };
        assert_eq!(amt.number, Decimal::from(20));
    }

    #[test]
    fn posting_amount_evaluates_subtraction_left_associative() {
        // `10 - 3 - 2 USD` = 5 USD (left-associative, not 9).
        let src = "2024-01-15 * \"x\"\n  \
                   Expenses:X   10 - 3 - 2 USD\n  \
                   Assets:Y   -5 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete amount");
        };
        assert_eq!(amt.number, Decimal::from(5));
    }

    #[test]
    fn posting_amount_division_by_zero_drops_number() {
        // `5 / 0 USD` - legacy returns parse error; we return None
        // from the evaluator, which degrades to CurrencyOnly here.
        // The transaction won't balance and downstream validation
        // surfaces that as the user-facing error.
        let src = "2024-01-15 * \"x\"\n  \
                   Expenses:X   5 / 0 USD\n  \
                   Assets:Y\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        // Either the units degrade to CurrencyOnly (number lost)
        // or to None - both are acceptable since the input is
        // semantically invalid. The strict assertion is that we
        // DON'T silently return 5 (the first NUMBER) as the value.
        match &t.postings[0].value.units {
            None | Some(IncompleteAmount::CurrencyOnly(_)) => {}
            other => panic!("div-by-zero leaked: {other:?}"),
        }
    }

    // ---- round-8 final compat regressions (#1282 flip) ---------

    #[test]
    fn indented_top_level_directive_emits_error() {
        // A top-level directive that starts at column N>0 is a
        // syntax error per the Beancount spec; the CST grammar
        // accepts it silently, so the converter has to surface
        // the diagnostic at directive-content-start position.
        let src = "2020-07-28 open Assets:Foo\n  2020-07-28 open Assets:Bar\n";
        let result = parse_via_cst(src);
        let indent_errs = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("column 0"),
                _ => false,
            })
            .count();
        assert_eq!(
            indent_errs, 1,
            "expected one column-0 diagnostic, got: {:?}",
            result.errors
        );
    }

    #[test]
    fn indented_directive_after_blank_line_still_emits_error() {
        // Same as above but with a blank line between the
        // first directive and the indented one - the blank line
        // shouldn't mask the indentation error.
        let src = "2020-07-28 open Assets:Foo\n\n  2020-07-28 open Assets:Bar\n";
        let result = parse_via_cst(src);
        let indent_errs = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("column 0"),
                _ => false,
            })
            .count();
        assert_eq!(indent_errs, 1, "errors: {:?}", result.errors);
    }

    #[test]
    fn top_level_directive_at_column_0_no_diagnostic() {
        // Sanity: well-formed top-level directives must NOT
        // trigger the indent diagnostic.
        let src = "2020-07-28 open Assets:Foo\n2020-07-28 open Assets:Bar\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn custom_directive_with_bare_currency_emits_error() {
        // `bean-check` rejects bare currency literals in custom
        // value position; the CST converter mirrors that.
        let src = "2025-01-01 custom \"x\" 10 USD \"y\" NZD\n";
        let result = parse_via_cst(src);
        let bare_curr_errs = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("bare currency"),
                _ => false,
            })
            .count();
        assert_eq!(
            bare_curr_errs, 1,
            "expected one bare-currency diagnostic, got: {:?}",
            result.errors
        );
    }

    #[test]
    fn custom_directive_with_amount_no_error() {
        // Sanity: `10 USD` (NUMBER + CURRENCY paired as Amount)
        // is a valid custom value and must NOT trigger the
        // bare-currency diagnostic.
        let src = "2025-01-01 custom \"x\" 10 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    // ---- round-7 compat regressions (#1282 flip) ---------------

    #[test]
    fn balance_assertion_evaluates_arithmetic_value() {
        // PR #1282 compat regression: rledger emitted a balance
        // failure for `Assets:X  0.25+ 0.75 GBP` because only
        // the first NUMBER (0.25) was used as the assertion
        // target. CST converters for BALANCE/PRICE now evaluate
        // arithmetic the same way posting AMOUNTs do.
        let src = "2024-01-01 open Assets:X GBP\n\
                   2024-01-01 open Equity:Open GBP\n\
                   2024-01-02 * \"deposit\"\n  \
                   Assets:X         1.00 GBP\n  \
                   Equity:Open     -1.00 GBP\n\
                   2024-01-03 balance Assets:X  0.25 + 0.75 GBP\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let bal = result
            .directives
            .iter()
            .find_map(|d| match &d.value {
                Directive::Balance(b) => Some(b),
                _ => None,
            })
            .expect("expected a Balance directive");
        assert_eq!(bal.amount.number, Decimal::from(1));
        assert_eq!(bal.amount.currency.as_str(), "GBP");
    }

    #[test]
    fn price_directive_evaluates_arithmetic_value() {
        let src = "2024-01-01 price USD  1/2 EUR\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Price(p) = &result.directives[0].value else {
            panic!("expected Price");
        };
        assert_eq!(p.amount.number, Decimal::new(5, 1));
        assert_eq!(p.amount.currency.as_str(), "EUR");
    }

    // ---- round-5 architecture review (#1281) -------------------

    #[test]
    fn body_line_tag_does_not_drop_following_postings_comment() {
        // F2-bis: trailing TAG / LINK tokens on transaction body
        // lines are valid Beancount (extend the transaction's
        // tag/link set). Before the exemption was added, the
        // `pending.clear()` over-fired on the TAG and silently
        // dropped the preceding comment that semantically
        // belonged to the next posting.
        let src = "2024-01-01 * \"x\"\n  \
                   Assets:A   100 USD\n  \
                   ; comment-for-B\n  \
                   #late-tag\n  \
                   Assets:B   -100 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        // The trailing tag joins the transaction's tag set.
        assert!(
            t.tags.iter().any(|tag| tag.as_str() == "late-tag"),
            "expected #late-tag in tags: {:?}",
            t.tags,
        );
        // And the comment survives - attached to the next posting.
        let b = t.postings.last().expect("at least one posting");
        assert_eq!(b.value.account.as_str(), "Assets:B");
        assert!(
            b.value.comments.iter().any(|c| c.contains("comment-for-B")),
            "expected comment-for-B to survive on Assets:B: {:?}",
            b.value.comments,
        );
    }

    #[test]
    fn oversized_number_in_amount_emits_diagnostic() {
        // F5-bis: the non-arithmetic NUMBER path is now symmetric
        // with the arithmetic-evaluation path. A NUMBER whose
        // text the lexer accepts but `Decimal::from_str` rejects
        // (e.g., 30+ digits, exceeding the 28-digit precision
        // ceiling) used to silently degrade to `CurrencyOnly`.
        let huge = "1".to_string() + &"2345678901234567890".repeat(2); // 39 digits
        let src = format!("2024-01-15 * \"big\"\n  Expenses:X   {huge} USD\n  Assets:Y\n");
        let result = parse_via_cst(&src);
        let invalid_num = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("invalid number"),
                _ => false,
            })
            .count();
        assert_eq!(
            invalid_num, 1,
            "expected one invalid-number diagnostic, got: {:?}",
            result.errors
        );
    }

    // ---- round-4 architecture review (#1281) -------------------

    #[test]
    fn posting_with_two_amount_siblings_emits_error_and_keeps_first() {
        // F1: a posting like `Expenses:Food  5 USD + 3 USD` builds
        // two sibling AMOUNT nodes in the CST. `Posting::amount()`
        // only returns the first. Without an explicit guard the
        // second AMOUNT plus the joining `+` would be silently
        // dropped - the user's transaction would balance against
        // 5 USD instead of the intended 8 USD with no diagnostic.
        let src = "2024-01-15 * \"ambig\"\n  \
                   Expenses:Food   5 USD + 3 USD\n  \
                   Assets:Bank\n";
        let result = parse_via_cst(src);
        let trailing_count = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("trailing tokens"),
                _ => false,
            })
            .count();
        assert_eq!(
            trailing_count, 1,
            "expected one trailing-tokens diagnostic, got: {:?}",
            result.errors
        );
        // The first AMOUNT is still surfaced so partial recovery
        // works for downstream tooling.
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        let Some(IncompleteAmount::Complete(amt)) = &t.postings[0].value.units else {
            panic!("expected complete units from the first AMOUNT");
        };
        assert_eq!(amt.number, Decimal::from(5));
    }

    #[test]
    fn comments_dont_leak_across_failed_posting() {
        // F2: when convert_posting returns None, the queue of
        // pending pre-posting comments must be CLEARED so they
        // don't migrate forward and attach to the next valid
        // posting. Without the clear, comments labelled for the
        // failed posting would silently re-attach to the wrong
        // account, visibly misleading the user.
        let src = "2024-01-15 * \"test\"\n  \
                   Assets:A   100 USD\n  \
                   ; comment-for-bad\n  \
                   ; another-comment\n  \
                   bogus_token_line_no_account\n  \
                   ; comment-for-good\n  \
                   Assets:B   -100 USD\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        // Assets:B is the LAST successful posting; the only
        // comment that should attach to it is the one that
        // immediately precedes it (`; comment-for-good`). The
        // pre-failed-posting comments belong to the failed
        // posting and should be DROPPED with it.
        let b = t.postings.last().expect("at least one posting");
        assert_eq!(b.value.account.as_str(), "Assets:B");
        assert!(
            !b.value
                .comments
                .iter()
                .any(|c| c.contains("comment-for-bad")),
            "comment-for-bad leaked across failed posting onto Assets:B: {:?}",
            b.value.comments
        );
        assert!(
            !b.value
                .comments
                .iter()
                .any(|c| c.contains("another-comment")),
            "another-comment leaked: {:?}",
            b.value.comments
        );
    }

    #[test]
    fn arithmetic_overflow_in_amount_emits_diagnostic() {
        // F5: when `is_arithmetic` is true but the evaluator
        // gives up (overflow, div-by-zero), the converter used
        // to silently produce CurrencyOnly. Now an explicit
        // SyntaxError fires so the user sees the actual root
        // cause instead of just a downstream "doesn't balance".
        // Decimal max is 28 digits - `9999999999999999999999999999 *
        // 9999999999999999999999999999` overflows.
        let huge = "9999999999999999999999999999 * 9999999999999999999999999999";
        let src = format!("2024-01-15 * \"big\"\n  Expenses:X   {huge} USD\n  Assets:Y\n");
        let result = parse_via_cst(&src);
        let arith_errs = result
            .errors
            .iter()
            .filter(|e| match &e.kind {
                crate::ParseErrorKind::SyntaxError(s) => s.contains("arithmetic"),
                _ => false,
            })
            .count();
        assert_eq!(
            arith_errs, 1,
            "expected one arithmetic-error diagnostic, got: {:?}",
            result.errors
        );
    }

    // ---- 14 emission-gap regressions (#1281 round-3 review) ----

    #[test]
    fn date_with_single_digit_month_parses() {
        let result = parse_via_cst("2024-1-15 open Assets:Checking\n");
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        assert_eq!(open.date, naive_date(2024, 1, 15).unwrap());
    }

    #[test]
    fn date_with_single_digit_day_parses() {
        let result = parse_via_cst("2024-01-5 open Assets:Cash USD\n");
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        assert_eq!(open.date, naive_date(2024, 1, 5).unwrap());
    }

    #[test]
    fn date_with_single_digit_month_and_day_parses() {
        let result = parse_via_cst("2024-1-1 open Assets:Cash USD\n");
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Open(open) = &result.directives[0].value else {
            panic!("expected Open");
        };
        assert_eq!(open.date, naive_date(2024, 1, 1).unwrap());
    }

    #[test]
    fn date_with_month_out_of_range_emits_invalid_date_value() {
        let result = parse_via_cst("2024-13-01 open Assets:Cash USD\n");
        let invalid_date: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::InvalidDateValue(s) => Some(s.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(invalid_date.len(), 1, "errors: {:?}", result.errors);
        let msg = &invalid_date[0];
        assert!(
            msg.contains("month") && msg.contains("out of range"),
            "msg: {msg}"
        );
    }

    #[test]
    fn date_with_invalid_leap_year_emits_invalid_date_value() {
        let result = parse_via_cst("2023-02-29 open Assets:Cash USD\n");
        let invalid_date: Vec<_> = result
            .errors
            .iter()
            .filter_map(|e| match &e.kind {
                crate::ParseErrorKind::InvalidDateValue(s) => Some(s.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(invalid_date.len(), 1, "errors: {:?}", result.errors);
        let msg = &invalid_date[0];
        assert!(
            msg.contains("day") && msg.contains("out of range") && msg.contains("2023-02"),
            "msg: {msg}"
        );
    }

    #[test]
    fn date_with_completely_invalid_value_still_emits_error() {
        // `2024-13-45` has BOTH month and day out of range; any
        // error variant satisfies the original integration test's
        // `!result.errors.is_empty()` assertion.
        let result = parse_via_cst("2024-13-45 open Assets:Bank\n");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn open_directive_without_account_emits_error() {
        // `2024-01-01 open` with no account is rejected by legacy
        // via the top-level error-recovery path. CST emits the
        // catch-all `SyntaxError` from `parse_via_cst`'s
        // is_directive_producing/errors_before tracker.
        let result = parse_via_cst("2024-01-01 open\n");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn open_directive_with_lowercase_account_emits_error() {
        // `lowercase:invalid` doesn't match the ACCOUNT regex
        // (uppercase first letter required), so the open directive
        // has no ACCOUNT child. Same catch-all path as the no-
        // account case.
        let result = parse_via_cst("2024-01-01 open lowercase:invalid\n");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn incomplete_open_at_eof_emits_error() {
        // Regression for the PR #740 "incomplete-at-EOF" finding:
        // `2024-01-01 open` at EOF with no trailing newline must
        // not be silently dropped.
        let result = parse_via_cst("2024-01-01 open");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn balance_directive_without_amount_emits_error() {
        let result = parse_via_cst("2024-01-15 balance Assets:Checking\n");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn pad_directive_without_source_account_emits_error() {
        let result = parse_via_cst("2024-01-15 pad Assets:Checking\n");
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn cost_spec_n_hash_t_parses_as_compound() {
        use rust_decimal_macros::dec;
        // This test previously pinned `{N # T}` -> Total{T} — the #1700
        // misparse (beancount's compound_amount weighs N*per + total,
        // so dropping the per-unit silently misweighed every compound
        // spec). It now pins the corrected as-written form.
        let src = "2024-01-01 open Assets:Stock\n\
                   2024-01-01 open Assets:Cash USD\n\
                   2024-01-15 *\n  \
                   Assets:Stock  10 STK {50 # 1500 USD}\n  \
                   Assets:Cash  -1500.00 USD\n";
        let result = parse_via_cst(src);
        assert!(result.errors.is_empty(), "errors: {:?}", result.errors);
        let Directive::Transaction(txn) = &result.directives[2].value else {
            panic!("expected Transaction at index 2");
        };
        let cost = txn.postings[0]
            .value
            .cost
            .as_ref()
            .expect("cost spec present");
        assert_eq!(
            cost.number,
            Some(CostNumber::Compound {
                per_unit: dec!(50),
                total: dec!(1500)
            }),
            "the `{{N # T CCY}}` form must carry both components as written"
        );
    }

    #[test]
    fn unclosed_cost_brace_emits_error() {
        let src = "2024-01-01 open Assets:Stock\n\
                   2024-01-01 open Assets:Cash USD\n\
                   2024-01-15 *\n  \
                   Assets:Stock 10 AAPL {150 USD\n  \
                   Assets:Cash -1500 USD\n";
        let result = parse_via_cst(src);
        let has_unclosed: bool = result
            .errors
            .iter()
            .any(|e| e.message().contains("unclosed cost"));
        assert!(
            has_unclosed,
            "expected 'unclosed cost' error, got: {:?}",
            result.errors
        );
    }

    #[test]
    fn unclosed_cost_brace_at_eof_emits_error() {
        let src = "2024-01-01 open Assets:Stock\n\
                   2024-01-01 open Assets:Cash USD\n\
                   2024-01-15 *\n  \
                   Assets:Stock 10 AAPL {150 USD";
        let result = parse_via_cst(src);
        let has_unclosed: bool = result
            .errors
            .iter()
            .any(|e| e.message().contains("unclosed cost"));
        assert!(
            has_unclosed,
            "expected 'unclosed cost' error at EOF, got: {:?}",
            result.errors
        );
    }

    #[test]
    fn leading_decimal_in_posting_amount_emits_error() {
        // `.50 USD` (no integer part before the decimal) must be
        // rejected by both parsers; valid `0.50 USD` still works
        // (covered by other tests).
        let src = "2024-01-15 * \"Test\"\n  \
                   Expenses:Food  .50 USD\n  \
                   Assets:Checking\n";
        let result = parse_via_cst(src);
        assert!(!result.errors.is_empty(), "errors: {:?}", result.errors);
    }

    #[test]
    fn transaction_with_metadata_on_directive_and_posting() {
        let src = "2024-01-15 * \"x\"\n  \
                   tag1: \"hello\"\n  \
                   Assets:Cash  -5 USD\n    \
                       receipt: \"abc123\"\n";
        let result = parse_via_cst(src);
        let Directive::Transaction(t) = &result.directives[0].value else {
            panic!("expected Transaction");
        };
        assert_eq!(
            t.meta.get("tag1"),
            Some(&MetaValue::String("hello".to_string()))
        );
        let p_meta = &t.postings[0].value.meta;
        assert_eq!(
            p_meta.get("receipt"),
            Some(&MetaValue::String("abc123".to_string()))
        );
    }

    /// Pins the `ERROR_NODE` exclusion contract on
    /// `account_occurrences`. The rustdoc on `ParseResult::
    /// account_occurrences` distinguishes two failure modes:
    ///
    /// - **Typed-conversion failure** (e.g. `InvalidBookingMethod`
    ///   on an `open` whose booking string is garbage): the CST is
    ///   intact, the `ACCOUNT` node is NOT inside `ERROR_NODE`, so
    ///   the token IS tracked. The LSP rename can still hit it
    ///   during mid-edit.
    /// - **CST-recovery wrap**: a directive so garbled that the
    ///   CST wraps the region in `ERROR_NODE`. The `ACCOUNT` token
    ///   is inside `ERROR_NODE`, NOT tracked.
    ///
    /// The two policies are deliberate. This test pins both.
    #[test]
    fn account_occurrences_policy_for_failing_directives() {
        // Case A: typed-conversion failure. `open Assets:Bank
        // "GARBAGE"` parses syntactically but fails the booking-
        // method whitelist. The ACCOUNT token IS tracked.
        let src = "2024-01-01 open Assets:Bank \"GARBAGE\"\n";
        let r = parse_via_cst(src);
        assert!(
            r.account_occurrences
                .iter()
                .any(|o| o.value == "Assets:Bank"),
            "typed-conversion failure should keep the ACCOUNT token in \
             account_occurrences (got {:?}); rename mid-edit relies on this",
            r.account_occurrences,
        );

        // Case B: CST-recovery wrap. `opn Assets:Bank USD` (typo
        // `opn`) is unrecognized at the directive position and the
        // recovery walker wraps it in ERROR_NODE. The ACCOUNT
        // token is excluded.
        let src = "2024-01-01 opn Assets:Bank USD\n";
        let r = parse_via_cst(src);
        assert!(
            !r.account_occurrences
                .iter()
                .any(|o| o.value == "Assets:Bank"),
            "ERROR_NODE-wrapped ACCOUNT should be EXCLUDED from \
             account_occurrences (got {:?}); rename should not hit garbled \
             mid-edit syntax",
            r.account_occurrences,
        );
    }

    // ---- cost-spec token latches and the `{*}` merge machine ----
    //
    // Added after the 2026-08-01 mutation run: every mutant in these two
    // machines survived. The merge rule was tested only through the `ast.rs`
    // copy (now deleted, it delegates here), and nothing at all exercised the
    // first-token latches, so a cost spec carrying a duplicate date, label or
    // currency was untested in either tree walker.

    /// Parse one posting's cost spec, or panic with the source for context.
    fn cost_of(src: &str) -> CostSpec {
        let result = parse_via_cst(src);
        let Some(Directive::Transaction(txn)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected a transaction from {src:?}");
        };
        txn.postings
            .first()
            .and_then(|p| p.cost.as_deref().cloned())
            .unwrap_or_else(|| panic!("expected a cost spec from {src:?}"))
    }

    fn posting_with_cost(spec: &str) -> String {
        format!("2020-01-01 * \"t\"\n  Assets:A 1 HOOL {spec}\n  Assets:B\n")
    }

    /// A repeated DATE, STRING or CURRENCY keeps the FIRST occurrence. Malformed
    /// input is the only way to get here, and "first wins" is what keeps the
    /// green and red walkers agreeing on it.
    #[test]
    fn cost_spec_latches_the_first_date_label_and_currency() {
        let cost = cost_of(&posting_with_cost(
            "{2 USD, 2020-06-01, 2021-02-02, \"first\", \"second\"}",
        ));
        assert_eq!(cost.date, naive_date(2020, 6, 1), "the FIRST date wins");
        assert_eq!(cost.label.as_deref(), Some("first"), "the FIRST label wins");
        assert_eq!(
            cost.currency
                .as_ref()
                .map(rustledger_core::Currency::as_str),
            Some("USD"),
            "the FIRST currency wins"
        );

        // Currency specifically, with nothing else competing.
        let cost = cost_of(&posting_with_cost("{2 USD, EUR}"));
        assert_eq!(
            cost.currency
                .as_ref()
                .map(rustledger_core::Currency::as_str),
            Some("USD")
        );
    }

    /// The latch is on the first token of a KIND, not the first that parses.
    ///
    /// `9999-99-99` lexes as a DATE and fails to parse. The latch must still
    /// close, leaving the date empty — falling through to a later, valid DATE
    /// would make the two walkers disagree on malformed input, which is the
    /// divergence class this design exists to prevent.
    #[test]
    fn cost_spec_latch_closes_on_an_unparsable_first_token() {
        let cost = cost_of(&posting_with_cost("{2 USD, 9999-99-99, 2021-02-02}"));
        assert_eq!(
            cost.date, None,
            "an unparsable first DATE must not let a later one through"
        );
    }

    /// The merge flag is decided by the first non-whitespace, non-opener token
    /// after an opener. Each row pins one arm of that machine.
    #[test]
    fn cost_spec_merge_flag_is_decided_by_the_first_token_after_an_opener() {
        for (spec, expected, why) in [
            ("{*}", true, "bare star directly after the opener"),
            (
                "{ * }",
                true,
                "whitespace never decides, so the star still does",
            ),
            ("{{*}}", true, "`{{` is an opener too"),
            (
                "{2 USD, *}",
                false,
                "the number decided it first; a later star cannot re-arm",
            ),
            (
                "{500 * 2 USD}",
                false,
                "a star past the first token is multiplication",
            ),
        ] {
            assert_eq!(
                cost_of(&posting_with_cost(spec)).merge,
                expected,
                "{spec}: {why}"
            );
        }
    }

    /// The red-tree accessor and the token-level canonical must agree, since
    /// they are now one machine fed by two walkers. Asserts on the exact pair
    /// rather than on either alone, so deleting the delegation is caught.
    #[test]
    fn ast_is_merge_agrees_with_the_converted_cost_spec() {
        for spec in [
            "{*}",
            "{ * }",
            "{{*}}",
            "{2 USD, *}",
            "{500 * 2 USD}",
            "{2 USD}",
        ] {
            let src = posting_with_cost(spec);
            let converted = cost_of(&src).merge;

            let parsed = crate::parse(&src);
            let root = ast::SourceFile::cast(parsed.syntax_node()).expect("source file");
            let from_ast = root
                .syntax()
                .descendants()
                .find_map(ast::CostSpec::cast)
                .map_or_else(|| panic!("no CostSpec node in {src:?}"), |cs| cs.is_merge());

            assert_eq!(
                from_ast, converted,
                "{spec}: ast::CostSpec::is_merge disagrees with the converted CostSpec"
            );
        }
    }

    /// `MergeFlag` guards `past_opener` because a token stream may begin before
    /// the opener. Both tree walkers happen to start AT the opener, so this is
    /// unreachable through them and only a direct feed can pin it -- but the
    /// guard is what stops a leading `*` (or any leading token) from deciding
    /// the flag, and it costs nothing to keep it honest.
    #[test]
    fn merge_flag_ignores_tokens_before_the_opener() {
        use crate::SyntaxKind as K;

        // A star BEFORE any opener is not a merge marker: nothing has opened
        // yet, so it cannot be the first token after an opener.
        let mut flag = MergeFlag::default();
        for kind in [K::STAR, K::L_BRACE, K::R_BRACE] {
            flag.feed(kind);
        }
        assert!(
            !flag.is_merge(),
            "a star before the opener must not decide the flag"
        );

        // And a non-star before the opener must not close the machine early,
        // or the real `{*}` that follows would be missed.
        let mut flag = MergeFlag::default();
        for kind in [K::NUMBER, K::L_BRACE, K::STAR, K::R_BRACE] {
            flag.feed(kind);
        }
        assert!(
            flag.is_merge(),
            "a token before the opener must not consume the decision"
        );
    }

    /// Every diagnostic span in this module is built as `offset + bom_offset`,
    /// and a wrong offset puts the editor's squiggle on the wrong text.
    ///
    /// The error paths were already exercised, but only for the PRESENCE of an
    /// error, so the 2026-08-01 mutation run could flip `+` to `-` or `*` in
    /// five different span computations without a single failure. Each row
    /// below drives one of them and asserts the exact range, once with no BOM
    /// and once with one, so the addition itself is pinned rather than merely
    /// the arithmetic happening to agree at zero.
    #[test]
    fn diagnostic_spans_point_at_the_offending_text_with_and_without_a_bom() {
        // (label, source, the substring the span must cover)
        let cases = [
            (
                "price with two numbers",
                "2024-01-15 price HOOL 1 2 USD\n",
                "1 2",
            ),
            (
                "balance with two numbers",
                "2024-01-15 balance Assets:Cash 1 2 USD\n",
                "1 2",
            ),
            (
                "posting with a second amount",
                "2024-01-15 *\n  Assets:A 5 USD + 3 USD\n  Assets:B\n",
                // Underlined from the END of the first amount on purpose, so
                // the reader sees `5 USD + 3 USD` and not just the tail.
                " + 3 USD",
            ),
            (
                // A `+`/`-` binds to the amount as its sign, so the orphan that
                // actually reaches this path is a stray comma - someone writing
                // `1,234` with the separator outside the number.
                "orphaned comma before a posting amount",
                "2024-01-15 *\n  Assets:A , 1,234.00 USD\n  Assets:B\n",
                ",",
            ),
        ];

        for (label, src, needle) in cases {
            for bom in [false, true] {
                let full = if bom {
                    format!("\u{FEFF}{src}")
                } else {
                    src.to_string()
                };
                let result = parse_via_cst(&full);
                let bom_len = if bom { "\u{FEFF}".len() } else { 0 };

                let expected_start = src
                    .find(needle)
                    .unwrap_or_else(|| panic!("{label}: {needle:?} not in the fixture"))
                    + bom_len;
                let expected_end = expected_start + needle.len();

                let hit = result
                    .errors
                    .iter()
                    .find(|e| e.span.start == expected_start && e.span.end == expected_end);
                assert!(
                    hit.is_some(),
                    "{label} (bom={bom}): expected an error spanning {expected_start}..{expected_end} \
                     (the {needle:?}), got {:?}",
                    result
                        .errors
                        .iter()
                        .map(|e| (e.span.start, e.span.end))
                        .collect::<Vec<_>>()
                );
            }
        }
    }

    /// A leading `-` on a `price`/`balance` number is a separate MINUS token,
    /// so the converter has to re-apply the sign the AST accessor drops. Both
    /// the negation and the scanner that finds it were untested.
    #[test]
    fn negative_numbers_in_price_and_balance_keep_their_sign() {
        // SPACED, so the sign is its own MINUS token and the AST accessor
        // hands back an unsigned number. `-1.50` written closed up lexes as a
        // single signed NUMBER and never reaches the scanner at all.
        let result = parse_via_cst("2024-01-15 price HOOL - 1.50 USD\n");
        let Some(Directive::Price(p)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected a Price, got {:?}", result.directives);
        };
        assert_eq!(p.amount.number, rust_decimal_macros::dec!(-1.50));

        let result = parse_via_cst("2024-01-15 balance Assets:Cash - 1.50 USD\n");
        let Some(Directive::Balance(b)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected a Balance, got {:?}", result.directives);
        };
        assert_eq!(b.amount.number, rust_decimal_macros::dec!(-1.50));

        // And the positive case must stay positive: a scanner that reports
        // "minus" for everything would pass the assertions above alone.
        let result = parse_via_cst("2024-01-15 price HOOL 1.50 USD\n");
        let Some(Directive::Price(p)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected a Price");
        };
        assert_eq!(p.amount.number, rust_decimal_macros::dec!(1.50));
    }

    /// `price` puts the BASE currency BEFORE the number, so the scan that
    /// rejects a two-number value may only stop at a currency once a number has
    /// been seen. Getting that guard wrong makes every `price` directive look
    /// malformed, or stops rejecting the thing it exists to reject.
    #[test]
    fn price_base_currency_before_the_number_is_not_a_malformed_value() {
        let result = parse_via_cst("2024-01-15 price HOOL 1.50 USD\n");
        assert!(
            result.errors.is_empty(),
            "a well-formed price must not be reported as malformed: {:?}",
            result.errors
        );
        assert_eq!(result.directives.len(), 1);

        // Two numbers still must be rejected.
        let result = parse_via_cst("2024-01-15 price HOOL 1 2 USD\n");
        assert!(
            has_syntax_error(&result, "malformed amount"),
            "two numbers must still be refused: {:?}",
            result.errors
        );
    }

    /// Only `+`, `-` and `,` are orphanable. A posting FLAG sits between the
    /// account and the amount too, and treating it as an orphan would reject
    /// perfectly ordinary input.
    #[test]
    fn a_posting_flag_is_not_an_orphaned_amount_prefix() {
        let result = parse_via_cst("2024-01-15 *\n  ! Assets:A 5 USD\n  Assets:B\n");
        assert!(
            !has_syntax_error(&result, "unexpected token before posting amount"),
            "a posting flag is not an orphan: {:?}",
            result.errors
        );
    }

    /// Both diagnostics inside posting-amount conversion carry spans built with
    /// the BOM offset, and neither was pinned. An arithmetic expression that
    /// cannot be evaluated and a number past the Decimal ceiling are the two
    /// ways in.
    #[test]
    fn posting_amount_diagnostics_point_at_the_offending_amount() {
        // 30 digits: past `rust_decimal`'s ~28-digit ceiling.
        let huge = "1".repeat(30);
        let cases = [
            (
                // The span covers the whole AMOUNT node, currency included:
                // the expression is what is wrong, but the amount is what the
                // reader has to replace.
                "unevaluatable arithmetic",
                "2024-01-15 *\n  Assets:A (1/0) USD\n  Assets:B\n".to_string(),
                "(1/0) USD".to_string(),
            ),
            (
                "number past the Decimal ceiling",
                format!("2024-01-15 *\n  Assets:A {huge} USD\n  Assets:B\n"),
                huge,
            ),
        ];

        for (label, src, needle) in cases {
            for bom in [false, true] {
                let full = if bom {
                    format!("\u{FEFF}{src}")
                } else {
                    src.clone()
                };
                let bom_len = if bom { "\u{FEFF}".len() } else { 0 };
                let result = parse_via_cst(&full);

                let start = src.find(&needle).expect("needle present") + bom_len;
                let end = start + needle.len();
                assert!(
                    result
                        .errors
                        .iter()
                        .any(|e| e.span.start == start && e.span.end == end),
                    "{label} (bom={bom}): expected a span {start}..{end}, got {:?}",
                    result
                        .errors
                        .iter()
                        .map(|e| (e.span.start, e.span.end))
                        .collect::<Vec<_>>()
                );
            }
        }
    }

    /// The trailing currency closes a directive value, and it must only do so
    /// once a number has been seen (a `price` names its base currency first).
    /// Without the break, a stray number after the currency would be counted
    /// and a well-formed directive rejected.
    #[test]
    fn a_trailing_currency_closes_the_value_scan() {
        let result = parse_via_cst("2024-01-15 price HOOL 1.50 USD 2\n");
        assert!(
            !has_syntax_error(&result, "malformed amount"),
            "the scan must stop at the closing currency, so the stray `2` is not \
             a second number of the VALUE: {:?}",
            result.errors
        );
    }

    /// Only tokens AFTER the account can be orphans, and only `+`, `-` and `,`
    /// qualify. Both halves of that were untested, so each row here would be
    /// reported as an orphan by a slightly wrong predicate.
    #[test]
    fn orphan_detection_ignores_pre_account_and_non_sign_tokens() {
        let orphan_reported = |src: &str| {
            has_syntax_error(
                &parse_via_cst(src),
                "unexpected token before posting amount",
            )
        };

        assert!(
            !orphan_reported("2024-01-15 *\n  , Assets:A 1 USD\n  Assets:B\n"),
            "a comma BEFORE the account is not an orphaned amount prefix"
        );
        assert!(
            !orphan_reported("2024-01-15 *\n  Assets:A \"note\" 1 USD\n  Assets:B\n"),
            "a non-sign token between account and amount is not an orphan"
        );
        // The genuine orphan still is one, so the assertions above cannot pass
        // by the detector simply never firing.
        assert!(
            orphan_reported("2024-01-15 *\n  Assets:A , 1 USD\n  Assets:B\n"),
            "a comma after the account IS an orphan"
        );
    }

    /// A posting's trailing comment is collected up to the newline. Stopping on
    /// the wrong condition silently drops every one of them.
    #[test]
    fn posting_trailing_comment_is_captured() {
        let result = parse_via_cst("2024-01-15 *\n  Assets:A 1 USD ; why\n  Assets:B\n");
        let Some(Directive::Transaction(txn)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected a transaction");
        };
        let first = &txn.postings[0];
        assert!(
            first.trailing_comments.iter().any(|c| c.contains("why")),
            "expected the trailing comment on the posting, got {:?}",
            first.trailing_comments
        );
    }

    /// The sign scanner behind `price`/`balance` fallback conversion.
    ///
    /// Reaching it takes work: `directive_arithmetic_value` runs first and
    /// handles ordinary unary minus, so `- 1.50` never gets here. The fallback
    /// only runs when the arithmetic parse declines, and `- 1.50 - USD` is one
    /// such shape -- error recovery tolerates the trailing operator, the
    /// arithmetic parse gives up, and the AST accessor then hands back an
    /// UNSIGNED number that this scanner has to re-sign.
    ///
    /// Probed for rather than assumed: the whole test suite and all 995 corpus
    /// files leave this branch cold, so it looked like dead code until an
    /// adversarial sweep found the inputs. Worth stating, because deleting it
    /// on that first impression would have been wrong.
    #[test]
    fn price_and_balance_fallback_re_signs_a_leading_minus() {
        let number_of = |src: &str| -> Decimal {
            let r = parse_via_cst(src);
            match r.directives.first().map(|d| &d.value) {
                Some(Directive::Price(p)) => p.amount.number,
                Some(Directive::Balance(b)) => b.amount.number,
                other => panic!("expected price/balance from {src:?}, got {other:?}"),
            }
        };

        // MINUS before the number: re-signed.
        assert_eq!(
            number_of("2024-01-15 price HOOL - 1.50 - USD\n"),
            rust_decimal_macros::dec!(-1.50)
        );
        assert_eq!(
            number_of("2024-01-15 balance Assets:C - 1.50 - USD\n"),
            rust_decimal_macros::dec!(-1.50)
        );

        // NUMBER first: the scan stops there, so a LATER minus must not flip
        // the sign. Without this the "stop at the number" arm is free to vanish.
        assert_eq!(
            number_of("2024-01-15 price HOOL 1.50 - USD\n"),
            rust_decimal_macros::dec!(1.50)
        );
        assert_eq!(
            number_of("2024-01-15 balance Assets:C 1.50 - USD\n"),
            rust_decimal_macros::dec!(1.50)
        );
    }

    /// The red conversion path is a mirror of the green one, kept for the
    /// `green_eq_red` differential fuzz target. Production parses via green, so
    /// a test written against `parse_via_cst` exercises the mirror only where
    /// the two SHARE a helper -- which is why the 2026-08-01 mutation run
    /// showed red-only code uncovered even though its green twin was tested.
    ///
    /// These drive `parse_red_only` directly, and assert the two paths agree,
    /// so the mirror cannot rot silently.
    #[test]
    fn red_path_matches_green_on_posting_comments_and_orphan_detection() {
        let orphan_msg = "unexpected token before posting amount";

        // Trailing comment on a posting line, collected by the red converter's
        // own scan up to the terminating NEWLINE.
        let src = "2024-01-15 *\n  Assets:A 1 USD ; why\n  Assets:B\n";
        for (label, result) in [("green", parse_via_cst(src)), ("red", parse_red_only(src))] {
            let Some(Directive::Transaction(txn)) = result.directives.first().map(|d| &d.value)
            else {
                panic!("{label}: expected a transaction");
            };
            assert!(
                txn.postings[0]
                    .trailing_comments
                    .iter()
                    .any(|c| c.contains("why")),
                "{label}: trailing comment lost, got {:?}",
                txn.postings[0].trailing_comments
            );
        }

        // Orphan detection, through the red converter: a comma after the
        // account is one, a comma before it and a non-sign token are not.
        let orphan_reported = |src: &str| has_syntax_error(&parse_red_only(src), orphan_msg);
        assert!(
            orphan_reported("2024-01-15 *\n  Assets:A , 1 USD\n  Assets:B\n"),
            "red: a comma after the account IS an orphan"
        );
        assert!(
            !orphan_reported("2024-01-15 *\n  , Assets:A 1 USD\n  Assets:B\n"),
            "red: a comma BEFORE the account is not"
        );
        assert!(
            !orphan_reported("2024-01-15 *\n  Assets:A \"note\" 1 USD\n  Assets:B\n"),
            "red: a non-sign token between account and amount is not"
        );
    }

    // ---- metadata and custom values: the other token-level canonical ----
    //
    // `meta_value_from_tokens` is the twin of `cost_spec_from_tokens` and had
    // the same shape of gap: first-of-kind latches and a sign machine that no
    // test touched. `value_tokens_to_meta` is the sibling used by custom
    // directives and the red path.

    fn meta_of(entries: &str) -> rustledger_core::Metadata {
        let src = format!("2024-01-15 open Assets:A\n{entries}");
        let result = parse_via_cst(&src);
        let Some(Directive::Open(open)) = result.directives.first().map(|d| &d.value) else {
            panic!("expected an Open from {src:?}, errors {:?}", result.errors);
        };
        open.meta.clone()
    }

    fn custom_values(line: &str) -> Vec<MetaValue> {
        let result = parse_via_cst(line);
        let Some(Directive::Custom(c)) = result.directives.first().map(|d| &d.value) else {
            panic!(
                "expected a Custom from {line:?}, errors {:?}",
                result.errors
            );
        };
        c.values.clone()
    }

    /// Every value kind a metadata entry can carry. Deleting any one arm made
    /// that kind silently fall through to the next candidate in the priority
    /// order, which is invisible unless the kind is asserted directly.
    #[test]
    fn metadata_values_cover_every_kind() {
        let meta = meta_of(
            "  str: \"hello\"\n  num: 42\n  amt: 42 USD\n  dt: 2024-06-01\n  \
             acct: Assets:B\n  cur: USD\n  yes: TRUE\n  no: FALSE\n  \
             tg: #mytag\n  lk: ^mylink\n",
        );
        let got = |k: &str| meta.get(k).cloned().unwrap_or(MetaValue::None);

        assert_eq!(got("str"), MetaValue::String("hello".into()));
        assert_eq!(got("num"), MetaValue::Int(42));
        assert_eq!(
            got("amt"),
            MetaValue::Amount(Amount::new(rust_decimal_macros::dec!(42), "USD"))
        );
        assert_eq!(got("dt"), MetaValue::Date(naive_date(2024, 6, 1).unwrap()));
        assert_eq!(got("acct"), MetaValue::Account(Account::new("Assets:B")));
        assert_eq!(got("cur"), MetaValue::Currency(Currency::new("USD")));
        assert_eq!(got("yes"), MetaValue::Bool(true));
        assert_eq!(got("no"), MetaValue::Bool(false));
        assert_eq!(got("tg"), MetaValue::Tag(Tag::new("mytag")));
        assert_eq!(got("lk"), MetaValue::Link(Link::new("mylink")));
    }

    /// First-of-kind latching, the same rule `cost_spec_from_tokens` uses. A
    /// repeated token of any kind keeps the FIRST, and nothing exercised that
    /// for metadata, so every latch guard could be flipped freely.
    #[test]
    fn metadata_latches_the_first_token_of_each_kind() {
        // Two-character keys because beancount requires them, and so do we
        // since #1955. These are fixture names only; the test is about
        // first-of-kind LATCHING and nothing here depends on key length.
        let meta = meta_of(
            "  ss: \"one\" \"two\"\n  nn: 1 2\n  cc: USD EUR\n  dd: 2024-06-01 2025-07-02\n  \
             aa: Assets:First Assets:Second\n  bb: TRUE FALSE\n  tt: #first #second\n",
        );
        let got = |k: &str| meta.get(k).cloned().unwrap_or(MetaValue::None);

        assert_eq!(got("ss"), MetaValue::String("one".into()));
        assert_eq!(got("nn"), MetaValue::Int(1));
        assert_eq!(got("dd"), MetaValue::Date(naive_date(2024, 6, 1).unwrap()));
        assert_eq!(got("aa"), MetaValue::Account(Account::new("Assets:First")));
        assert_eq!(got("bb"), MetaValue::Bool(true), "TRUE came first");
        assert_eq!(got("tt"), MetaValue::Tag(Tag::new("first")));
        // `c` pairs a number-less currency run: the FIRST currency wins.
        assert_eq!(got("cc"), MetaValue::Currency(Currency::new("USD")));
    }

    /// The sign machine: a MINUS after the key negates the number.
    ///
    /// The third case CHANGED with #1944. It used to assert that `42 - 1` is
    /// `Int(42)` — "a minus past the number is not a sign; the first NUMBER
    /// closes it". That described the truncation faithfully but was never the
    /// right answer: beancount evaluates it and reports **41**, verified
    /// directly against the oracle before this expectation was touched. The
    /// old assertion was pinning a bug as intended behavior, which is why it
    /// took a differential comparison rather than a reading to notice.
    ///
    /// The sign machine itself is unchanged and still pinned by the first two
    /// cases: a leading MINUS is now consumed by the expression evaluator
    /// instead of a separate flag, and reaches the same values.
    #[test]
    fn metadata_minus_applies_only_before_the_number() {
        let meta = meta_of("  neg: -42\n  negamt: -42 USD\n  after: 42 - 1\n");
        let got = |k: &str| meta.get(k).cloned().unwrap_or(MetaValue::None);

        assert_eq!(got("neg"), MetaValue::Int(-42));
        assert_eq!(
            got("negamt"),
            MetaValue::Amount(Amount::new(rust_decimal_macros::dec!(-42), "USD")),
            "the sign applies to the amount too"
        );
        assert_eq!(
            got("after"),
            MetaValue::Int(41),
            "an expression in a metadata value is evaluated, matching beancount"
        );
    }

    /// `value_tokens_to_meta` walks a token run and returns the NEXT index, so
    /// a wrong advance either drops values or repeats them. Custom directives
    /// are the surface that reads several values in a row, which makes the
    /// advance observable.
    #[test]
    fn custom_directive_values_advance_one_value_at_a_time() {
        assert_eq!(
            custom_values("2024-01-15 custom \"b\" FALSE TRUE FALSE\n"),
            vec![
                MetaValue::Bool(false),
                MetaValue::Bool(true),
                MetaValue::Bool(false)
            ],
            "each bool consumes exactly one token"
        );

        assert_eq!(
            custom_values("2024-01-15 custom \"b\" 42 USD TRUE\n"),
            vec![
                MetaValue::Amount(Amount::new(rust_decimal_macros::dec!(42), "USD")),
                MetaValue::Bool(true)
            ],
            "NUMBER + CURRENCY consumes TWO tokens and the next value still lands"
        );

        // A lone CURRENCY is its own token, not an amount fragment -- it
        // consumes exactly one, so `TRUE` still lands. It is no longer
        // EMITTED, because a bare currency is rejected as a custom value by
        // both tools and a rejected value must not reach `values` (#2190).
        // The advance is what this test is about, and the surviving `TRUE`
        // is what proves it.
        assert_eq!(
            custom_values("2024-01-15 custom \"b\" USD TRUE\n"),
            vec![MetaValue::Bool(true)],
            "a lone CURRENCY consumes one token, so the next value still lands"
        );

        // The tag and link are rejected here and so are not emitted, but they
        // still consume one token each -- which is exactly what the date and
        // account landing afterwards proves.
        assert_eq!(
            custom_values("2024-01-15 custom \"b\" -42 #tag ^link 2024-06-01 Assets:B\n"),
            vec![
                MetaValue::Int(-42),
                MetaValue::Date(naive_date(2024, 6, 1).unwrap()),
                MetaValue::Account(Account::new("Assets:B")),
            ],
            "MINUS + NUMBER consumes two tokens; the rest follow in order"
        );
    }

    /// The bool and tag/link latches, in the order that actually exercises the
    /// SECOND arm of each pair. `TRUE FALSE` only proves the first arm latches;
    /// reversing it is what pins the guard on the other one.
    #[test]
    fn metadata_latches_bool_and_taglink_in_either_order() {
        // Two-character keys, per #1955; `tl` / `lt` already were.
        let meta = meta_of("  bb: FALSE TRUE\n  tl: #tag ^link\n  lt: ^link #tag\n");
        let got = |k: &str| meta.get(k).cloned().unwrap_or(MetaValue::None);

        assert_eq!(
            got("bb"),
            MetaValue::Bool(false),
            "FALSE came first, so the later TRUE must not overwrite it"
        );
        assert_eq!(
            got("tl"),
            MetaValue::Tag(Tag::new("tag")),
            "tag and link share one slot; the tag came first"
        );
        assert_eq!(
            got("lt"),
            MetaValue::Link(Link::new("link")),
            "and the link wins when it comes first"
        );
    }

    /// `extract_custom_values` advances by the index the discriminator returns.
    /// A helper that failed to advance would spin the loop forever, so the
    /// caller clamps. This pins that a long run of values terminates and is
    /// read in order -- a hang here would let malformed input stall the parser.
    #[test]
    fn custom_values_terminate_on_a_long_run() {
        let values = custom_values(
            "2024-01-15 custom \"b\" 1 USD 2 EUR TRUE FALSE #a ^b 2024-06-01 Assets:X \"s\"\n",
        );
        // Nine values are READ; the tag and the link are rejected and dropped,
        // leaving seven. `1 USD` and `2 EUR` each PAIR into an Amount, so
        // neither currency is bare and neither is dropped. The point of the
        // test is that the run terminates and stays in order, which the
        // surviving sequence still shows.
        assert_eq!(
            values.len(),
            7,
            "every value consumed exactly once, got {values:?}"
        );
        assert_eq!(
            values.first(),
            Some(&MetaValue::Amount(Amount::new(
                rust_decimal_macros::dec!(1),
                "USD"
            )))
        );
        assert_eq!(values.last(), Some(&MetaValue::String("s".into())));
    }

    /// The scan skips the directive header (date, keyword, and the type-name
    /// string) before reading values, steps past tokens that are not values,
    /// and terminates when there are none. Each of those is a separate step in
    /// the loop and none had a test.
    #[test]
    fn custom_directive_scan_skips_the_header_and_non_values() {
        assert_eq!(
            custom_values("2024-01-15 custom \"b\"\n"),
            vec![],
            "the type name is the header, not a value, and no values is valid"
        );

        assert_eq!(
            custom_values("2024-01-15 custom \"b\" \"x\" 42\n"),
            vec![MetaValue::String("x".into()), MetaValue::Int(42)],
            "the FIRST string is the type name; a later one IS a value"
        );

        // A `*` is not a value, so the scan must step over it rather than
        // stall. Both positions matter: before any value, and between two.
        assert_eq!(
            custom_values("2024-01-15 custom \"b\" * 42\n"),
            vec![MetaValue::Int(42)],
            "a non-value token before the first value is stepped over"
        );
        assert_eq!(
            custom_values("2024-01-15 custom \"b\" 42 * 7\n"),
            vec![MetaValue::Int(42), MetaValue::Int(7)],
            "and between two values"
        );
    }
}

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

    /// A `custom` directive must not report a value it has just rejected.
    ///
    /// `Custom.values` used to carry the tag, link or bare currency that the
    /// same parse had raised an error for, so a consumer reading `values`
    /// without also reading `errors` saw data the parser had declared invalid
    /// (#2190).
    ///
    /// All three types are rejected by beancount too, measured across its
    /// eight custom value types -- so this is not a stricter grammar, only an
    /// honest one.
    #[test]
    fn a_rejected_custom_value_is_not_emitted() {
        for (src, what) in [
            ("2024-01-05 custom \"b\" #atag\n", "tag"),
            ("2024-01-05 custom \"b\" ^alink\n", "link"),
            ("2024-01-05 custom \"b\" USD\n", "bare currency"),
        ] {
            let res = crate::parse(src);
            assert_eq!(res.errors.len(), 1, "{what} must be rejected: {src:?}");
            let values = res
                .directives
                .iter()
                .find_map(|d| match &d.value {
                    Directive::Custom(c) => Some(c.values.clone()),
                    _ => None,
                })
                .expect("the directive still parses, minus the bad value");
            assert!(
                values.is_empty(),
                "the rejected {what} must not survive into values; got {values:?}",
            );
        }
    }

    /// Only the rejected value is dropped. A directive mixing good and bad
    /// values keeps the good ones, so this is not "discard the whole line".
    #[test]
    fn valid_custom_values_survive_alongside_a_rejected_one() {
        let res = crate::parse("2024-01-05 custom \"b\" \"ok\" #atag 42\n");
        assert_eq!(res.errors.len(), 1, "the tag is rejected");
        let values = res
            .directives
            .iter()
            .find_map(|d| match &d.value {
                Directive::Custom(c) => Some(c.values.clone()),
                _ => None,
            })
            .expect("custom directive");
        assert_eq!(
            values,
            vec![MetaValue::String("ok".to_string()), MetaValue::Int(42)],
            "the string and the integer survive; only the tag is dropped",
        );
    }

    /// A `#tag` is still a legal METADATA value -- beancount accepts one there
    /// and stores it, so the shared `value_tokens_to_meta` must keep producing
    /// `MetaValue::Tag`. Only the custom-directive path filters it.
    ///
    /// Pinned because the fix above filters a value the two paths share, and
    /// filtering it one level lower would silently break metadata.
    #[test]
    fn a_tag_is_still_a_valid_metadata_value() {
        let res = crate::parse("2024-01-20 note Assets:A \"n\"\n  mytag: #kept\n");
        assert!(
            res.errors.is_empty(),
            "tag metadata parses: {:?}",
            res.errors
        );
        let meta = res
            .directives
            .iter()
            .find_map(|d| match &d.value {
                Directive::Note(n) => Some(n.meta.clone()),
                _ => None,
            })
            .expect("note directive");
        assert_eq!(
            meta.get("mytag"),
            Some(&MetaValue::Tag(Tag::new("kept"))),
            "the tag survives as a Tag in metadata; got {meta:?}",
        );
    }
}