rustledger-parser 0.16.2

//! Source-driven tests for `parse_structured` (phase 2.1-2.4).
//!
//! Each test feeds real Beancount source through the structured
//! parser and asserts the resulting tree shape against the
//! Directive-Terminator Rule (see `cst::trivia`).
//!
//! These complement (do NOT replace) the hand-constructed-tree
//! tests in `cst::trivia::tests` — those pin the policy as
//! invariants on tree shape, these pin that
//! `parse_structured(source)` actually PRODUCES trees matching
//! those invariants on real source.

// Each test references many `SyntaxKind` variants for its expected
// children sequence; a per-test glob import is the cleanest local
// shape. Clippy's enum_glob_use lint is the wrong call here.
#![allow(clippy::enum_glob_use)]

use rustledger_parser::{SyntaxKind, SyntaxNode, parse_structured};

/// Per-child kind sequence for a node. Distinguishes tokens from
/// nested nodes so a test can assert both leaf trivia and structural
/// wrapping at the same node level.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Element {
    Tok(SyntaxKind),
    Node(SyntaxKind),
}

fn elements_of(node: &SyntaxNode) -> Vec<Element> {
    node.children_with_tokens()
        .map(|el| match el {
            rowan::NodeOrToken::Token(t) => Element::Tok(t.kind()),
            rowan::NodeOrToken::Node(n) => Element::Node(n.kind()),
        })
        .collect()
}

fn tok_seq(kinds: &[SyntaxKind]) -> Vec<Element> {
    kinds.iter().copied().map(Element::Tok).collect()
}

/// Find direct-children directive nodes of any specific
/// `*_DIRECTIVE` kind under `root`.
fn directives(root: &SyntaxNode) -> Vec<SyntaxNode> {
    root.children()
        .filter(|c| {
            matches!(
                c.kind(),
                SyntaxKind::OPEN_DIRECTIVE
                    | SyntaxKind::CLOSE_DIRECTIVE
                    | SyntaxKind::BALANCE_DIRECTIVE
                    | SyntaxKind::PAD_DIRECTIVE
                    | SyntaxKind::EVENT_DIRECTIVE
                    | SyntaxKind::QUERY_DIRECTIVE
                    | SyntaxKind::NOTE_DIRECTIVE
                    | SyntaxKind::DOCUMENT_DIRECTIVE
                    | SyntaxKind::PRICE_DIRECTIVE
                    | SyntaxKind::COMMODITY_DIRECTIVE
                    | SyntaxKind::PUSHTAG_DIRECTIVE
                    | SyntaxKind::POPTAG_DIRECTIVE
                    | SyntaxKind::PUSHMETA_DIRECTIVE
                    | SyntaxKind::POPMETA_DIRECTIVE
                    | SyntaxKind::OPTION_DIRECTIVE
                    | SyntaxKind::INCLUDE_DIRECTIVE
                    | SyntaxKind::PLUGIN_DIRECTIVE
                    | SyntaxKind::CUSTOM_DIRECTIVE
                    | SyntaxKind::TRANSACTION
            )
        })
        .collect()
}

/// Round-trip property: the tree's text must equal the source for
/// every input. Asserted at the top of every test.
fn assert_round_trip(source: &str, tree: &SyntaxNode) {
    assert_eq!(
        tree.text().to_string(),
        source,
        "structured parser must round-trip byte-identically",
    );
}

// ---------- 10 dated directives ----------

#[test]
fn open_directive_with_currency() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, WHITESPACE, CURRENCY, NEWLINE
        ]),
    );
}

#[test]
fn close_directive() {
    use SyntaxKind::*;
    let source = "2024-12-31 close Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CLOSE_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, CLOSE_KW, WHITESPACE, ACCOUNT, NEWLINE]),
    );
}

#[test]
fn balance_directive() {
    use SyntaxKind::*;
    let source = "2024-06-30 balance Assets:Cash 100.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), BALANCE_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, BALANCE_KW, WHITESPACE, ACCOUNT, WHITESPACE, NUMBER, WHITESPACE,
            CURRENCY, NEWLINE,
        ]),
    );
}

#[test]
fn pad_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 pad Assets:Cash Equity:Opening-Balances\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PAD_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, PAD_KW, WHITESPACE, ACCOUNT, WHITESPACE, ACCOUNT, NEWLINE,
        ]),
    );
}

#[test]
fn event_directive() {
    use SyntaxKind::*;
    let source = "2024-01-15 event \"location\" \"Berlin\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), EVENT_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, EVENT_KW, WHITESPACE, STRING, WHITESPACE, STRING, NEWLINE,
        ]),
    );
}

#[test]
fn query_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 query \"income\" \"SELECT *\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), QUERY_DIRECTIVE);
}

#[test]
fn note_directive() {
    use SyntaxKind::*;
    let source = "2024-01-15 note Assets:Cash \"deposit\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), NOTE_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, NOTE_KW, WHITESPACE, ACCOUNT, WHITESPACE, STRING, NEWLINE,
        ]),
    );
}

#[test]
fn document_directive() {
    use SyntaxKind::*;
    let source = "2024-01-15 document Assets:Cash \"/path/to/file.pdf\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), DOCUMENT_DIRECTIVE);
}

#[test]
fn price_directive() {
    use SyntaxKind::*;
    let source = "2024-01-15 price USD 1.10 EUR\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PRICE_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, PRICE_KW, WHITESPACE, CURRENCY, WHITESPACE, NUMBER, WHITESPACE,
            CURRENCY, NEWLINE,
        ]),
    );
}

#[test]
fn commodity_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 commodity USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), COMMODITY_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE,
            WHITESPACE,
            COMMODITY_KW,
            WHITESPACE,
            CURRENCY,
            NEWLINE
        ]),
    );
}

// ---------- 4 standalone-keyword directives ----------

#[test]
fn pushtag_directive() {
    use SyntaxKind::*;
    let source = "pushtag #project-x\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PUSHTAG_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[PUSHTAG_KW, WHITESPACE, TAG, NEWLINE]),
    );
}

#[test]
fn poptag_directive() {
    use SyntaxKind::*;
    let source = "poptag #project-x\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), POPTAG_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[POPTAG_KW, WHITESPACE, TAG, NEWLINE]),
    );
}

#[test]
fn pushmeta_directive() {
    use SyntaxKind::*;
    let source = "pushmeta key: \"value\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PUSHMETA_DIRECTIVE);
}

#[test]
fn popmeta_directive() {
    use SyntaxKind::*;
    let source = "popmeta key:\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), POPMETA_DIRECTIVE);
}

// ---------- Trivia attachment tests (Directive-Terminator Rule) ----------

#[test]
fn rule_1_same_line_trailing_comment_inside_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash  ; main checking\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    // Rule 1: WS + COMMENT + terminator NEWLINE all INSIDE the directive.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, WHITESPACE, COMMENT, NEWLINE,
        ]),
    );
}

#[test]
fn rule_2_blank_line_leads_following_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash\n\
                  \n\
                  2024-01-02 open Assets:Bank\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    // Rule 1: d1 owns its own terminator NEWLINE.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, NEWLINE]),
    );
    // Rule 2: the blank-line NEWLINE leads d2.
    assert_eq!(
        elements_of(&ds[1]),
        tok_seq(&[
            NEWLINE, DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, NEWLINE
        ]),
    );
}

#[test]
fn rule_3_copyright_header_under_source_file() {
    use SyntaxKind::*;
    let source = ";; Copyright 2024\n\
                  2024-01-01 open Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    // Rule 3: header trivia is direct under SOURCE_FILE, NOT inside d1.
    assert_eq!(
        elements_of(&tree),
        vec![
            Element::Tok(COMMENT),
            Element::Tok(NEWLINE),
            Element::Node(OPEN_DIRECTIVE),
        ],
    );
}

#[test]
fn rule_4_trailing_comment_block_under_source_file() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash\n\
                  ;; closing remarks\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    // Rule 4: trailing comment block is direct under SOURCE_FILE,
    // NOT inside the file-final directive.
    assert_eq!(
        elements_of(&tree),
        vec![
            Element::Node(OPEN_DIRECTIVE),
            Element::Tok(COMMENT),
            Element::Tok(NEWLINE),
        ],
    );
}

#[test]
fn rule_5_unterminated_final_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    // Rule 5: no terminator. Directive ends at last content token.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT]),
    );
}

#[test]
fn rule_5_unterminated_with_same_line_trailing_trivia() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash  ; eol-no-nl";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    // Rules 1+5: same-line trailing trivia stays INSIDE the
    // directive even without a terminator NEWLINE.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[
            DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, WHITESPACE, COMMENT,
        ]),
    );
}

#[test]
fn mixed_directive_kinds_each_get_their_own_node() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash USD\n\
                  pushtag #x\n\
                  2024-01-02 close Assets:Cash\n\
                  poptag #x\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 4);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[1].kind(), PUSHTAG_DIRECTIVE);
    assert_eq!(ds[2].kind(), CLOSE_DIRECTIVE);
    assert_eq!(ds[3].kind(), POPTAG_DIRECTIVE);
}

// ---------- Phase 2.1b: TRANSACTION header recognition ----------

#[test]
fn transaction_with_star_flag_header_only() {
    use SyntaxKind::*;
    // `*` indicates a completed transaction. Header-only (no
    // postings yet — that's the simplest TRANSACTION shape).
    let source = "2024-01-15 * \"Coffee\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, STAR, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn transaction_with_pending_kw_flag() {
    use SyntaxKind::*;
    // `!` lexes as PENDING_KW (`Token::Pending` →
    // `SyntaxKind::PENDING_KW`), NOT as `FLAG`. It signals an
    // incomplete/warning transaction in Beancount syntax.
    let source = "2024-01-15 ! \"WIP\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    // Pin the exact token sequence so a regression that
    // mistokenizes `!` or fails to wrap the full header fires.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, PENDING_KW, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn transaction_with_txn_keyword() {
    use SyntaxKind::*;
    // Explicit `txn` keyword form.
    let source = "2024-01-15 txn \"explicit\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
}

#[test]
fn transaction_with_postings_wraps_full_multi_line_body() {
    use SyntaxKind::*;
    // Per cst::trivia's multi-line clause, TRANSACTION owns its
    // header AND every indented sub-line until non-indented
    // content (or EOF). Postings here are flat tokens inside
    // TRANSACTION; PR 2.2 will introduce POSTING / AMOUNT / etc.
    // sub-nodes.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    // SOURCE_FILE owns ONLY the TRANSACTION node — no orphaned
    // posting tokens.
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_with_metadata_and_postings() {
    use SyntaxKind::*;
    // Transactions can carry intra-transaction metadata AND
    // postings. All sub-lines inside TRANSACTION.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20note: \"morning\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_with_payee_and_narration() {
    use SyntaxKind::*;
    // Full transaction header with payee + narration + tag + link.
    let source = "2024-01-15 * \"Coffee Shop\" \"Morning coffee\" #daily ^trip1\n\
                  \x20\x20Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
}

#[test]
fn transaction_terminates_at_next_top_level_directive() {
    use SyntaxKind::*;
    // After a transaction's postings, a non-indented DATE starts
    // a NEW directive. TRANSACTION must close cleanly; the next
    // OPEN_DIRECTIVE must not be absorbed.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  2024-01-16 open Assets:Bank\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), TRANSACTION);
    assert_eq!(ds[1].kind(), OPEN_DIRECTIVE);
}

#[test]
fn transaction_terminates_at_blank_line_before_next_directive() {
    use SyntaxKind::*;
    // A blank line after a transaction's last posting ends it.
    // The blank-line NEWLINE becomes inter-directive trivia
    // leading the next directive (rule 2).
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \n\
                  2024-01-16 open Assets:Bank\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), TRANSACTION);
    assert_eq!(ds[1].kind(), OPEN_DIRECTIVE);
    // The blank-line NEWLINE leads OPEN per rule 2.
    let d2_first = elements_of(&ds[1]).first().copied();
    assert_eq!(d2_first, Some(Element::Tok(NEWLINE)));
}

#[test]
fn transaction_with_indented_comment_between_postings() {
    use SyntaxKind::*;
    // Comments interleaved with postings stay inside TRANSACTION.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20; documentation comment\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_with_implied_flag_via_bare_string() {
    use SyntaxKind::*;
    // Beancount accepts the implied-transaction shorthand:
    // `DATE WS STRING ...` with no explicit flag. The legacy
    // AST parser at parser.rs:1713 dispatches `Token::String(_)`
    // to parse_transaction_directive with an implied `*`. Common
    // in real ledgers as a convenient shorthand.
    let source = "2024-01-15 \"Coffee\"\n\
                  \x20\x20Assets:Cash 100 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    // SOURCE_FILE owns only the TRANSACTION — no orphaned posting.
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_with_hash_flag() {
    use SyntaxKind::*;
    // `#` is promoted to a transaction flag when it appears in
    // the post-DATE flag slot. The lexer's `Token::is_txn_flag`
    // includes Hash and the AST parser's `parse_flag` accepts it;
    // the CST mirrors that contract.
    let source = "2024-01-15 # \"pending hash\"\n\
                  \x20\x20Assets:Cash 100 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    // SOURCE_FILE owns only the TRANSACTION — no orphaned posting.
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_with_single_char_currency_as_flag() {
    use SyntaxKind::*;
    // NYSE/NASDAQ-style single-letter tickers (T, V, F, X, ...)
    // tokenize as CURRENCY (priority 3 over FLAG in the lexer)
    // but are accepted as transaction flags. The AST parser's
    // `parse_flag` arm `Token::Currency(s) if s.len() == 1` does
    // this; the CST mirrors it.
    let source = "2024-01-15 T \"AT&T dividend\"\n\
                  \x20\x20Assets:Brokerage 10 T\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_multi_char_currency_after_date_is_not_a_flag() {
    // Guard against the reverse: `USD` (a real currency, length 3)
    // must NOT be treated as a transaction flag. The CURRENCY arm
    // gates on length == 1.
    let source = "2024-01-15 USD \"garbled\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    // No directive recognized — falls into the passthrough branch.
    let ds = directives(&tree);
    assert!(
        ds.is_empty(),
        "multi-char CURRENCY after DATE must not be a transaction flag",
    );
}

#[test]
fn transaction_blank_line_inside_body_terminates_and_orphans_subsequent_postings() {
    use SyntaxKind::*;
    // Pins the documented blank-line termination behavior (matches
    // Python beancount). The second posting after the blank line
    // ends up flat under SOURCE_FILE, not inside the TRANSACTION.
    // PR 2.2b's POSTING-wrapping work must NOT accidentally widen
    // the body scope across blank lines; this test guards that.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 100 USD\n\
                  \n\
                  \x20\x20Liab:Card -100 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    // Exactly ONE recognized directive (the transaction). The
    // post-blank posting is flat passthrough, not a second
    // structural node.
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);

    // The TRANSACTION header tokens (DATE, STAR) are direct flat
    // children. The first posting line is wrapped in a POSTING
    // node; the second posting (post-blank) is NOT.
    let header_kinds: Vec<SyntaxKind> = elements_of(&ds[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(
        header_kinds.contains(&DATE) && header_kinds.contains(&STAR),
        "tx contains header",
    );
    // Exactly one POSTING is wrapped inside the TRANSACTION; the
    // post-blank posting is orphaned under SOURCE_FILE.
    let postings_inside_tx = ds[0].children().filter(|n| n.kind() == POSTING).count();
    assert_eq!(
        postings_inside_tx, 1,
        "only the FIRST posting is wrapped inside the tx; the second is orphaned",
    );
    // ACCOUNT count across the whole tree: 2 (one in the wrapped
    // POSTING, one orphaned flat under SOURCE_FILE).
    let total_accounts = tree
        .descendants_with_tokens()
        .filter(|e| e.kind() == ACCOUNT)
        .count();
    assert_eq!(total_accounts, 2);
}

#[test]
fn transaction_trailing_indented_comment_at_eof_stays_inside() {
    use SyntaxKind::*;
    // TRANSACTION deliberately diverges from rule 4 (which puts
    // indented trailing comments under SOURCE_FILE for the 14
    // single-line directive kinds). The transaction body
    // predicate accepts any indented non-blank line, so a
    // trailing indented comment after the last posting stays
    // inside the TRANSACTION. Compare with
    // `indented_comment_at_eof_after_no_metadata_directive_is_file_trailing`
    // earlier in this file for the OPEN_DIRECTIVE policy.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 100 USD\n\
                  \x20\x20Liab:Card -100 USD\n\
                  \x20\x20; closing note for this transaction\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
    // SOURCE_FILE owns ONLY the TRANSACTION — comment is inside.
    assert_eq!(elements_of(&tree), vec![Element::Node(TRANSACTION)]);
}

#[test]
fn transaction_unterminated_at_eof_with_postings() {
    use SyntaxKind::*;
    // No final NEWLINE on the last posting line. Per rule 5,
    // TRANSACTION wraps content up to EOF without fabricating
    // a terminator.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20Assets:Cash  -5.00 USD";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), TRANSACTION);
}

// ---------- Pass-through for still-unrecognized content ----------

// ---------- Phase 2.2a: META_ENTRY structural wrapping ----------

/// Walk all `META_ENTRY` descendants of a node, in source order.
fn meta_entries(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::META_ENTRY)
        .collect()
}

#[test]
fn meta_entry_wraps_metadata_sub_line_inside_open_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash\n  description: \"main checking\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 1);
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn meta_entry_wraps_each_of_multiple_metadata_sub_lines() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20key1: \"value1\"\n\
                  \x20\x20key2: \"value2\"\n\
                  \x20\x20key3: \"value3\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 3);
    for me in &mes {
        assert_eq!(
            elements_of(me),
            tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
        );
    }
}

#[test]
fn meta_entry_does_not_wrap_indented_comments() {
    use SyntaxKind::*;
    // An indented `;`-comment between metadata entries stays as
    // flat children of the parent directive — NOT inside a
    // META_ENTRY. META_ENTRY is reserved for metadata sub-lines
    // proper (the `WS META_KEY ... NEWLINE` shape).
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20k1: \"v1\"\n\
                  \x20\x20; doc comment\n\
                  \x20\x20k2: \"v2\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 2, "only k1 and k2 are META_ENTRYs");

    // The indented comment line lives as flat tokens (WS, COMMENT,
    // NEWLINE) inside the OPEN_DIRECTIVE between the two
    // META_ENTRY nodes.
    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    let kids: Vec<Element> = elements_of(&ds[0]);
    let comment_pos = kids
        .iter()
        .position(|e| matches!(e, Element::Tok(COMMENT)))
        .expect("indented COMMENT lives flat in the directive");
    let n_me_before_comment = kids[..comment_pos]
        .iter()
        .filter(|e| matches!(e, Element::Node(META_ENTRY)))
        .count();
    let n_me_after_comment = kids[comment_pos..]
        .iter()
        .filter(|e| matches!(e, Element::Node(META_ENTRY)))
        .count();
    assert_eq!(n_me_before_comment, 1, "k1 META_ENTRY precedes the comment");
    assert_eq!(n_me_after_comment, 1, "k2 META_ENTRY follows the comment");
}

#[test]
fn meta_entry_inside_transaction_body() {
    use SyntaxKind::*;
    // Transactions can carry intra-transaction metadata. The
    // META_ENTRY wrapping applies there too.
    let source = "2024-01-15 * \"Coffee\"\n\
                  \x20\x20note: \"morning\"\n\
                  \x20\x20Assets:Cash -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 1);
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );

    // The `note:` line is at the same indent as the posting (2
    // spaces) and appears BEFORE the posting, so it's
    // TRANSACTION-level metadata: META_ENTRY is a direct child of
    // TRANSACTION. The posting line is now wrapped in POSTING
    // (PR 2.2b).
    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    assert_eq!(txs.len(), 1);
    let n_meta_entries_in_tx = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(n_meta_entries_in_tx, 1);
    let n_postings_in_tx = txs[0].children().filter(|n| n.kind() == POSTING).count();
    assert_eq!(n_postings_in_tx, 1);
}

#[test]
fn meta_entry_at_eof_without_trailing_newline() {
    use SyntaxKind::*;
    // Per rule 5 of `cst::trivia` (unterminated final directive),
    // a metadata sub-line that ends mid-content without a final
    // NEWLINE still gets wrapped in META_ENTRY — the META_ENTRY
    // simply has no NEWLINE child. Pins the rustdoc claim.
    let source = "2024-01-01 open Assets:Cash\n  key: \"v\"";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 1);
    // The META_ENTRY contains WS + META_KEY + WS + STRING and NO
    // NEWLINE (last token reached EOF).
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING]),
    );
}

#[test]
fn meta_entry_with_value_kinds_other_than_string() {
    use SyntaxKind::*;
    // Metadata values can be a NUMBER, ACCOUNT, CURRENCY, DATE,
    // boolean, etc. META_ENTRY wraps the whole sub-line regardless
    // of the value kind — phase 3's typed AST will surface
    // `value()` accessors that decode by inspecting children.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20count: 42\n\
                  \x20\x20since: 2024-01-01\n\
                  \x20\x20mirror: Assets:Mirror\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let mes = meta_entries(&tree);
    assert_eq!(mes.len(), 3);
    // Spot-check the second (DATE-valued) entry's value-token kind.
    let date_me_kinds = elements_of(&mes[1]);
    assert!(date_me_kinds.contains(&Element::Tok(DATE)));
}

// ---------- Phase 2.2b: POSTING structural wrapping ----------

/// Walk all `POSTING` descendants of a node, in source order.
fn postings(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::POSTING)
        .collect()
}

#[test]
fn posting_wraps_account_only_line() {
    use SyntaxKind::*;
    // The simplest posting: indent + ACCOUNT (no amount). Beancount
    // calls this an "auto" posting — booking infers the amount from
    // the others. Round-trip + a single POSTING wrapper around the
    // sub-line.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    assert_eq!(
        elements_of(&ps[0]),
        tok_seq(&[WHITESPACE, ACCOUNT, NEWLINE]),
    );
}

#[test]
fn posting_wraps_account_with_amount_and_currency() {
    use SyntaxKind::*;
    // A normal posting with amount + currency. POSTING contains
    // the indent WHITESPACE, ACCOUNT, inter-token WHITESPACE, an
    // AMOUNT sub-node (PR 2.2c), then NEWLINE.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    assert_eq!(
        elements_of(&ps[0]),
        vec![
            Element::Tok(WHITESPACE),
            Element::Tok(ACCOUNT),
            Element::Tok(WHITESPACE),
            Element::Node(AMOUNT),
            Element::Tok(NEWLINE),
        ],
    );

    // AMOUNT internal shape: MINUS NUMBER WS CURRENCY.
    let amounts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amounts.len(), 1);
    assert_eq!(
        elements_of(&amounts[0]),
        tok_seq(&[MINUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn posting_wraps_each_of_multiple_postings_in_a_transaction() {
    use SyntaxKind::*;
    // Two postings → two POSTING nodes; each wraps an AMOUNT
    // sub-node around the NUMBER + CURRENCY portion (PR 2.2c).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 2);
    for p in &ps {
        assert!(
            elements_of(p)
                .iter()
                .any(|e| matches!(e, Element::Tok(ACCOUNT)))
        );
        let n_amounts = p.children().filter(|n| n.kind() == AMOUNT).count();
        assert_eq!(n_amounts, 1, "each POSTING contains exactly one AMOUNT");
    }
}

#[test]
fn posting_with_pending_flag_wraps_flag_inside_node() {
    use SyntaxKind::*;
    // `! Assets:Cash ...` — the PENDING_KW flag sits between the
    // indent and the ACCOUNT. After PR 2.2c, the amount portion is
    // wrapped in an AMOUNT sub-node; the flag and account remain
    // flat tokens inside POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20! Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    assert_eq!(
        elements_of(&ps[0]),
        vec![
            Element::Tok(WHITESPACE),
            Element::Tok(PENDING_KW),
            Element::Tok(WHITESPACE),
            Element::Tok(ACCOUNT),
            Element::Tok(WHITESPACE),
            Element::Node(AMOUNT),
            Element::Tok(NEWLINE),
        ],
    );
}

#[test]
fn posting_attached_meta_entry_lives_inside_posting() {
    use SyntaxKind::*;
    // The key PR 2.2b semantic: a META_ENTRY sub-line at STRICTLY
    // GREATER indent than the preceding POSTING attaches to that
    // POSTING (not to the TRANSACTION). Mirrors the legacy AST
    // parser's `parse_posting_metadata` (DeepIndent loop), which
    // accumulates metadata into `posting.meta`.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20note: \"posting-attached\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    assert_eq!(txs.len(), 1);
    // TRANSACTION direct children: ONE POSTING, ZERO META_ENTRY
    // (the deeper-indented META_ENTRY belongs to the POSTING, not
    // to TRANSACTION).
    let tx_posting_count = txs[0].children().filter(|n| n.kind() == POSTING).count();
    let tx_meta_count = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(tx_posting_count, 1);
    assert_eq!(tx_meta_count, 0);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    // POSTING's children include the META_ENTRY as a structural
    // child (alongside the posting's flat tokens).
    let posting_meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(posting_meta_count, 1);
}

#[test]
fn same_indent_metadata_attaches_to_preceding_posting() {
    use SyntaxKind::*;
    // A META_ENTRY at the SAME indent as the preceding POSTING is
    // posting-attached, matching Beancount. Beancount attributes
    // metadata by POSITION (any `key: value` line following a posting,
    // before the next posting, attaches to that posting) rather than
    // by relative indent, so same-column `key: value` is POSTING
    // metadata, not transaction metadata. (Verified against the
    // Python beancount loader: a same-indent `key:` line lands in the
    // preceding posting's `meta`.) This is the case that drove the
    // last `effective_date`-example compat divergence.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20note: \"on cash\"\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    assert_eq!(txs.len(), 1);
    // TRANSACTION direct children: TWO POSTINGs and NO direct
    // META_ENTRY (the metadata lives inside the first POSTING).
    let tx_posting_count = txs[0].children().filter(|n| n.kind() == POSTING).count();
    let tx_meta_count = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(tx_posting_count, 2);
    assert_eq!(tx_meta_count, 0);
    // The META_ENTRY is a child of the FIRST posting; the second has none.
    let ps = postings(&tree);
    assert_eq!(ps.len(), 2);
    assert_eq!(
        ps[0].children().filter(|n| n.kind() == META_ENTRY).count(),
        1
    );
    assert_eq!(
        ps[1].children().filter(|n| n.kind() == META_ENTRY).count(),
        0
    );
}

#[test]
fn posting_attached_multiple_meta_entries_all_inside_posting() {
    use SyntaxKind::*;
    // Multiple deeper-indented metadata lines following the same
    // POSTING all attach to it.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20key1: \"v1\"\n\
                  \x20\x20\x20\x20key2: \"v2\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let posting_meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(posting_meta_count, 2);
}

#[test]
fn posting_attached_meta_entry_terminates_at_next_posting() {
    use SyntaxKind::*;
    // After posting-attached metadata, a NEW POSTING line at the
    // standard indent closes the current POSTING and opens a new
    // one. The new POSTING starts empty (no inherited metadata).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20note: \"on cash\"\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 2);
    // First POSTING owns the META_ENTRY; second is clean.
    let first_meta = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    let second_meta = ps[1].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(first_meta, 1);
    assert_eq!(second_meta, 0);
}

#[test]
fn postings_at_increasing_indents_produce_siblings_and_meta_attributes_to_latest() {
    use SyntaxKind::*;
    // Defensive shape: Beancount normally uses uniform posting
    // indentation. But the state machine doesn't enforce
    // monotonic indent — two posting lines at different indents
    // produce sibling POSTING nodes, and a subsequent META_ENTRY
    // attributes against the MOST-RECENTLY-OPENED POSTING's
    // indent. Pins this behavior so any future "monotonic indent"
    // refactor is a visible, intentional break.
    //
    // Source:
    //   posting at 2 spaces
    //   posting at 4 spaces  (DEEPER than the first)
    //   meta at 2 spaces     (NOT strictly deeper than 4)
    //
    // Expected: two POSTING siblings; the meta closes the second
    // (its indent is shallower) and lands at TRANSACTION level.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:A\n\
                  \x20\x20\x20\x20Assets:B  10 USD\n\
                  \x20\x20note: \"transaction-level\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 2, "two POSTING siblings at different indents");

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    assert_eq!(txs.len(), 1);
    let tx_meta = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(
        tx_meta, 1,
        "meta at shallower indent than the open POSTING lands at TRANSACTION level",
    );
    // Neither POSTING owns the META_ENTRY.
    for p in &ps {
        let inner_meta = p.children().filter(|n| n.kind() == META_ENTRY).count();
        assert_eq!(inner_meta, 0);
    }
}

#[test]
fn meta_entry_before_first_posting_stays_at_transaction_level() {
    use SyntaxKind::*;
    // A META_ENTRY that appears BEFORE any POSTING (regardless of
    // indent depth, since there's no preceding POSTING to attach
    // to) is always TRANSACTION-level.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20\x20\x20note: \"before posting\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    assert_eq!(txs.len(), 1);
    let tx_meta = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    let tx_posting = txs[0].children().filter(|n| n.kind() == POSTING).count();
    assert_eq!(tx_meta, 1);
    assert_eq!(tx_posting, 1);
    // The POSTING itself has no META_ENTRY child.
    let ps = postings(&tree);
    assert_eq!(
        ps[0].children().filter(|n| n.kind() == META_ENTRY).count(),
        0
    );
}

#[test]
fn deeper_indented_comment_stays_inside_posting_with_following_meta() {
    use SyntaxKind::*;
    // Doc-comment-for-following-posting-metadata idiom: an indented
    // `;` comment at indent STRICTLY GREATER than the open POSTING
    // (and at the same depth as the subsequent posting-attached
    // META_ENTRY) belongs to the POSTING. Both the comment AND the
    // META_ENTRY land inside the POSTING node.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20; comment about note\n\
                  \x20\x20\x20\x20note: \"deeper\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);

    // The deeper-indented META_ENTRY is attached to the POSTING.
    let posting_meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(posting_meta_count, 1);

    // The deeper-indented COMMENT token is also inside POSTING.
    let posting_comment_count = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(
        posting_comment_count, 1,
        "deeper-indented `;` comment stays inside POSTING with following meta",
    );

    // TRANSACTION's direct children have ZERO orphaned META_ENTRY
    // or COMMENT.
    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    let tx_meta = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    let tx_comment = txs[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(tx_meta, 0);
    assert_eq!(tx_comment, 0);
}

#[test]
fn deeper_indented_comment_stays_inside_posting_even_without_following_meta() {
    use SyntaxKind::*;
    // Rule is purely indent-based: a deeper-indented comment
    // belongs to the open POSTING regardless of whether a
    // META_ENTRY follows. Pins the rule's edge so the predicate
    // can't drift to "only attach when followed by meta".
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20; trailing posting doc\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 2);
    let first_comment_count = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    let second_comment_count = ps[1]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(first_comment_count, 1);
    assert_eq!(second_comment_count, 0);
}

#[test]
fn posting_with_indented_comment_between_postings_terminates_posting() {
    use SyntaxKind::*;
    // An indented `;`-comment between two posting lines is
    // TRANSACTION-level inter-posting trivia: it closes the
    // current POSTING. The comment ends up as flat tokens between
    // the two POSTING nodes (matches the existing
    // `transaction_with_indented_comment_between_postings`
    // structural intent).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD\n\
                  \x20\x20; doc comment\n\
                  \x20\x20Expenses:Food  5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 2);
    // The COMMENT token lives between the two POSTING nodes as a
    // flat child of TRANSACTION, NOT inside either POSTING.
    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    let tx_kids = elements_of(&txs[0]);
    let first_posting_idx = tx_kids
        .iter()
        .position(|e| matches!(e, Element::Node(POSTING)))
        .unwrap();
    let comment_idx = tx_kids
        .iter()
        .position(|e| matches!(e, Element::Tok(COMMENT)))
        .expect("indented comment is a flat TRANSACTION child");
    assert!(
        comment_idx > first_posting_idx,
        "comment follows first POSTING"
    );
}

#[test]
fn posting_at_eof_without_trailing_newline_still_wrapped() {
    use SyntaxKind::*;
    // Per rule 5 of `cst::trivia` (unterminated final directive),
    // a POSTING that reaches EOF mid-content without a final
    // NEWLINE still gets wrapped — the POSTING simply has no
    // NEWLINE child.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5.00 USD";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    // POSTING children: WS, ACCOUNT, WS, AMOUNT (no trailing
    // NEWLINE because of rule 5).
    assert_eq!(
        elements_of(&ps[0]),
        vec![
            Element::Tok(WHITESPACE),
            Element::Tok(ACCOUNT),
            Element::Tok(WHITESPACE),
            Element::Node(AMOUNT),
        ],
    );
    // AMOUNT internal shape, also unterminated.
    let amounts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(
        elements_of(&amounts[0]),
        tok_seq(&[MINUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn star_flagged_posting_wraps_flag_inside_node() {
    use SyntaxKind::*;
    // `* Account ...` (STAR-flagged posting) is also a valid
    // beancount posting shape. The STAR sits between the indent
    // and the ACCOUNT inside POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20* Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&ps[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.starts_with(&[WHITESPACE, STAR, WHITESPACE, ACCOUNT]));
}

#[test]
fn flagged_posting_with_question_mark_wraps_flag_inside_node() {
    use SyntaxKind::*;
    // `? Account ...` — the `?` flag emits a FLAG token (the
    // single-letter alphabetic flags P/S/T/C/U/R/M are tokenized
    // as CURRENCY by lexer priority 3 — covered by
    // `single_char_currency_flagged_posting_wraps_currency_as_flag`).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20? Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&ps[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.starts_with(&[WHITESPACE, FLAG, WHITESPACE, ACCOUNT]));
}

#[test]
fn hash_flagged_posting_wraps_hash_inside_node() {
    use SyntaxKind::*;
    // `# Account ...` is a valid Beancount posting flag (legacy
    // `parse_flag` accepts `Token::Hash`; `identify_directive`
    // accepts HASH as a transaction trigger). Pin that
    // `starts_posting_sub_line` recognizes it so the line is
    // wrapped in POSTING rather than falling through as flat
    // tokens.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20# Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&ps[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.starts_with(&[WHITESPACE, HASH, WHITESPACE, ACCOUNT]));
}

#[test]
fn hash_flagged_posting_attached_meta_entry_lives_inside_posting() {
    use SyntaxKind::*;
    // Combines HASH flag with posting-attached META_ENTRY (the
    // shape the bare hash_flagged_posting_wraps_hash_inside_node
    // test alone couldn't catch a regression on). If a future
    // change drops HASH from `starts_posting_sub_line`, this test
    // fails because the line would no longer open a POSTING and
    // the deeper-indented META_ENTRY would orphan to TRANSACTION.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20# Assets:Cash  -5.00 USD\n\
                  \x20\x20\x20\x20note: \"hash-flagged\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let posting_meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(posting_meta_count, 1);

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    let tx_meta = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(tx_meta, 0);
}

#[test]
fn deeper_indented_trailing_comment_at_eof_stays_inside_posting() {
    use SyntaxKind::*;
    // Doc-comment-attribution rule extended to the EOF case: a
    // deeper-indented `;` comment that is the LAST sub-line of the
    // file (no final NEWLINE) still attaches to the open POSTING.
    // Per rule 5 of `cst::trivia` (recursive application: an
    // unterminated POSTING ends at its last content token without a
    // NEWLINE child of its own).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  \x20\x20\x20\x20; deep trailing";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let posting_comment_count = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(
        posting_comment_count, 1,
        "EOF-trailing deep `;` comment is a child of POSTING",
    );

    // No COMMENT orphaned to TRANSACTION level.
    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    let tx_comment = txs[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(tx_comment, 0);
}

#[test]
fn deeper_indented_emacs_directive_attaches_to_open_posting() {
    use SyntaxKind::*;
    // `is_comment_token` includes EMACS_DIRECTIVE (`#+`). The
    // indented-comment branch in `emit_transaction_body` routes it
    // through the same indent-attribution rule as COMMENT: deeper-
    // indented than the open POSTING = stays INSIDE POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  \x20\x20\x20\x20#+STARTUP: overview\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let emacs_inside_posting = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    assert_eq!(
        emacs_inside_posting, 1,
        "EMACS_DIRECTIVE recognized as comment-class trivia, attaches by indent",
    );
}

#[test]
fn deeper_indented_shebang_attaches_to_open_posting() {
    use SyntaxKind::*;
    // Companion to the EMACS_DIRECTIVE test: pin that SHEBANG
    // (`#!`) is also recognized as comment-class trivia via
    // `is_comment_token` and follows the same indent-attribution
    // rule. Catches a regression that drops SHEBANG from the
    // helper while leaving EMACS_DIRECTIVE in place.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  \x20\x20\x20\x20#!/usr/bin/env something\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let shebang_inside_posting = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == SHEBANG)
        .count();
    assert_eq!(
        shebang_inside_posting, 1,
        "SHEBANG recognized as comment-class trivia, attaches by indent",
    );
}

#[test]
fn deeper_indented_percent_comment_attaches_to_open_posting() {
    use SyntaxKind::*;
    // PERCENT_COMMENT (`%`) is included in `is_comment_token` but
    // every other comment-attribution test uses `;`. Pin the `%`
    // path so a regression that demotes PERCENT_COMMENT (e.g., via
    // a typo or split refactor) fails here.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  \x20\x20\x20\x20% percent-style doc\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let pct_inside_posting = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == PERCENT_COMMENT)
        .count();
    assert_eq!(
        pct_inside_posting, 1,
        "PERCENT_COMMENT recognized as comment-class trivia, attaches by indent",
    );
}

#[test]
fn directive_body_absorbs_indented_emacs_directive_when_block_has_meta() {
    use SyntaxKind::*;
    // The `is_comment_token` widening also affects
    // `upcoming_indented_block_has_meta` and
    // `is_indented_directive_continuation` for NON-transaction
    // directives. Pin that an indented `#+STARTUP` line inside an
    // OPEN_DIRECTIVE that ALSO contains a meta line is absorbed as
    // a continuation (rather than orphaning to SOURCE_FILE).
    // Mirrors the existing `indented_comment_before_first_metadata`
    // / `indented_comment_between_metadata_lines` tests, which use
    // `;` only; this pins the SHEBANG/EMACS_DIRECTIVE branch.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20#+STARTUP: overview\n\
                  \x20\x20key: \"v\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);

    // The EMACS_DIRECTIVE token lives inside the OPEN_DIRECTIVE,
    // not orphaned anywhere else in the tree. Use `descendants`
    // symmetrically on both sides so a future refactor that wraps
    // SOURCE_FILE trivia in any nested node doesn't make the
    // orphan check vacuously pass.
    let emacs_total = tree
        .descendants_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    let emacs_in_directive = ds[0]
        .descendants_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    assert_eq!(emacs_total, 1, "exactly one EMACS_DIRECTIVE in the tree");
    assert_eq!(
        emacs_in_directive, 1,
        "EMACS_DIRECTIVE absorbed by OPEN_DIRECTIVE"
    );

    // The block_has_meta look-ahead is what kept the EMACS line
    // inside the directive: the subsequent `key: "v"` becomes a
    // META_ENTRY child of OPEN_DIRECTIVE. Assert that META_ENTRY
    // actually appears so a regression that breaks the META_KEY
    // arm (closing the directive AFTER the EMACS line but BEFORE
    // the meta) fails here, not silently.
    let meta_entries_in_directive = ds[0]
        .descendants()
        .filter(|n| n.kind() == META_ENTRY)
        .count();
    assert_eq!(meta_entries_in_directive, 1, "META_ENTRY also absorbed");
}

#[test]
fn directive_body_does_not_absorb_indented_emacs_directive_when_no_meta() {
    use SyntaxKind::*;
    // Complementary case: when an OPEN_DIRECTIVE has NO meta block,
    // an indented EMACS_DIRECTIVE / SHEBANG / `;`-comment that
    // follows the header is NOT a continuation (per the
    // block_has_meta gate). Pins that the widening did not
    // accidentally make these tokens unconditional continuations.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20#+STARTUP: trailing only\n\
                  2024-01-02 open Assets:Bank\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(
        ds.len(),
        2,
        "two OPEN_DIRECTIVES, separated by EMACS_DIRECTIVE rule-2 trivia"
    );
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[1].kind(), OPEN_DIRECTIVE);

    // The EMACS_DIRECTIVE is rule-2 inter-directive trivia: it
    // attaches as LEADING trivia of the SECOND directive (per
    // `cst::trivia`), NOT as a continuation of the first.
    // Symmetric `descendants` walks on both sides and a total-
    // count sanity check guard against future structural changes
    // that wrap trivia in a nested node.
    let emacs_total = tree
        .descendants_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    let emacs_in_first = ds[0]
        .descendants_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    let emacs_in_second = ds[1]
        .descendants_with_tokens()
        .filter(|e| e.kind() == EMACS_DIRECTIVE)
        .count();
    assert_eq!(emacs_total, 1, "exactly one EMACS_DIRECTIVE in the tree");
    assert_eq!(
        emacs_in_first, 0,
        "EMACS_DIRECTIVE is NOT absorbed by header-only directive"
    );
    assert_eq!(
        emacs_in_second, 1,
        "EMACS_DIRECTIVE leads the next directive as rule-2 trivia"
    );
}

#[test]
fn catch_all_indented_unknown_content_closes_posting_and_emits_flat() {
    use SyntaxKind::*;
    // Catch-all `else` branch of emit_transaction_body: an indented
    // sub-line that is neither posting, meta, nor comment closes
    // any open POSTING and emits flat at TRANSACTION level.
    // Examples: a stray bare STRING on its own indented line. Pin
    // the behavior so PR 2.2c (AMOUNT continuations etc.) doesn't
    // silently shift attribution without an explicit test update.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  \x20\x20\"stray string on own line\"\n\
                  \x20\x20Expenses:Food 1 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(
        ps.len(),
        2,
        "stray indented STRING closes POSTING; next POSTING opens fresh"
    );

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    // The stray STRING token is a flat child of TRANSACTION (not
    // inside either POSTING). TRANSACTION's direct STRING tokens
    // include the header narration "x" PLUS the stray, for a total
    // of 2.
    let tx_strings: usize = txs[0]
        .children_with_tokens()
        .filter(|e| e.kind() == STRING)
        .count();
    assert_eq!(tx_strings, 2);
    for p in &ps {
        let inside_string = p
            .children_with_tokens()
            .filter(|e| e.kind() == STRING)
            .count();
        assert_eq!(inside_string, 0);
    }
}

#[test]
fn same_indent_comment_between_posting_and_deeper_meta_orphans_meta() {
    use SyntaxKind::*;
    // The same-indent `;` comment between a posting and a deeper-
    // indented META_KEY closes the POSTING (per the
    // indent-attribution rule: comment indent is not strictly
    // greater than posting indent). The subsequent deeper-indented
    // META_KEY then has no open POSTING and lands at TRANSACTION
    // level. Matches the legacy AST parser's
    // `parse_posting_metadata` loop, which terminates posting-
    // attached metadata at any indented sub-line that is not a
    // DeepIndent META_KEY (a same-indent COMMENT being one such
    // terminator). Python beancount parity is NOT verified here —
    // a future compat audit may find Python attaches the deeper
    // META to the still-open posting, in which case this test is
    // the touch-point. Pinned so a future refactor can't silently
    // flip the attribution without a test update.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash -5 USD\n\
                  \x20\x20; explicit break at posting indent\n\
                  \x20\x20\x20\x20key: \"orphaned\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let posting_meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(
        posting_meta_count, 0,
        "same-indent comment ends posting-attached meta block; deeper meta orphans to TRANSACTION",
    );

    let txs: Vec<SyntaxNode> = tree
        .children()
        .filter(|c| c.kind() == TRANSACTION)
        .collect();
    let tx_meta = txs[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(tx_meta, 1);
}

#[test]
fn single_char_currency_flagged_posting_wraps_currency_as_flag() {
    use SyntaxKind::*;
    // `P Account ...` — `P` tokenizes as CURRENCY (lexer priority
    // 3) but functions as a posting flag, mirroring the transaction
    // header's same Currency-vs-Flag tie-break. POSTING still wraps
    // the line.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20P Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&ps[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.starts_with(&[WHITESPACE, CURRENCY, WHITESPACE, ACCOUNT]));
}

// ---------- Phase 2.2c: AMOUNT / COST_SPEC / PRICE_ANNOTATION ----------

/// Walk all `AMOUNT` descendants of a node, in source order.
fn amounts(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::AMOUNT)
        .collect()
}

/// Walk all `COST_SPEC` descendants of a node, in source order.
fn cost_specs(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::COST_SPEC)
        .collect()
}

/// Walk all `PRICE_ANNOTATION` descendants of a node, in source order.
fn price_annotations(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::PRICE_ANNOTATION)
        .collect()
}

#[test]
fn amount_wraps_positive_number_and_currency() {
    use SyntaxKind::*;
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  100.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_negative_number_and_currency_with_sign_token() {
    use SyntaxKind::*;
    // Lexer emits MINUS + NUMBER for negative amounts (enables
    // arithmetic expressions). The sign token lives INSIDE the
    // AMOUNT wrapper.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -100.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[MINUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_explicit_plus_sign_and_currency() {
    use SyntaxKind::*;
    // `+100 USD` — the lexer emits PLUS + NUMBER; both live
    // inside AMOUNT.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  +100.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[PLUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_number_only_no_currency() {
    use SyntaxKind::*;
    // Incomplete amount: NUMBER without a trailing currency.
    // AMOUNT wraps the NUMBER alone.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  100.00\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(elements_of(&amts[0]), tok_seq(&[NUMBER]));
}

#[test]
fn amount_wraps_currency_only_no_number() {
    use SyntaxKind::*;
    // Incomplete amount: bare CURRENCY (currency-only amount).
    // AMOUNT wraps the CURRENCY alone.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(elements_of(&amts[0]), tok_seq(&[CURRENCY]));
}

#[test]
fn auto_posting_with_no_amount_has_no_amount_node() {
    // Account-only "auto" posting — the booker fills in the
    // amount from the others. No AMOUNT wrapper because there's
    // nothing to wrap.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 0, "auto posting has no AMOUNT child");
    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
}

#[test]
fn cost_spec_wraps_simple_per_unit_cost() {
    use SyntaxKind::*;
    // Per-unit cost spec: `{NUMBER WS CURRENCY}`. COST_SPEC's
    // direct children include the brace tokens and the inner
    // amount tokens flat (per-2.2c design — contents are
    // unstructured until phase 3).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {500.00 USD}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    assert_eq!(
        elements_of(&cs[0]),
        tok_seq(&[L_BRACE, NUMBER, WHITESPACE, CURRENCY, R_BRACE]),
    );
}

#[test]
fn cost_spec_wraps_total_double_brace() {
    use SyntaxKind::*;
    // `{{ ... }}` total-cost form.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {{5000.00 USD}}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    assert_eq!(
        elements_of(&cs[0]),
        tok_seq(&[L_DOUBLE_BRACE, NUMBER, WHITESPACE, CURRENCY, R_DOUBLE_BRACE]),
    );
}

#[test]
fn cost_spec_wraps_per_unit_plus_total_brace_hash() {
    use SyntaxKind::*;
    // `{# ... }` per-unit + total form (HASH separator inside).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {# 5000.00 USD}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    // L_BRACE_HASH opener, content tokens, R_BRACE close.
    let kinds: Vec<SyntaxKind> = elements_of(&cs[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&L_BRACE_HASH));
    assert!(kinds.contains(&R_BRACE));
}

#[test]
fn cost_spec_unclosed_at_eof_still_wraps_per_rule_5() {
    use SyntaxKind::*;
    // Per rule 5, an unclosed brace at EOF still gets wrapped;
    // the COST_SPEC simply has no matching close-brace child.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {500.00 USD";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&cs[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&L_BRACE));
    assert!(!kinds.contains(&R_BRACE), "no close brace consumed");
    assert!(kinds.contains(&CURRENCY));
}

#[test]
fn price_annotation_wraps_per_unit_with_nested_amount() {
    use SyntaxKind::*;
    // `@ NUMBER WS CURRENCY` — per-unit price. PRICE_ANNOTATION
    // contains AT + WHITESPACE + AMOUNT.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL @ 500.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    assert_eq!(
        elements_of(&prices[0]),
        vec![
            Element::Tok(AT),
            Element::Tok(WHITESPACE),
            Element::Node(AMOUNT),
        ],
    );
    // The nested AMOUNT has its own NUMBER/CURRENCY.
    let inner_amount = prices[0].children().find(|n| n.kind() == AMOUNT).unwrap();
    assert_eq!(
        elements_of(&inner_amount),
        tok_seq(&[NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn price_annotation_wraps_total_at_at() {
    use SyntaxKind::*;
    // `@@ NUMBER WS CURRENCY` — total price. The opener distinguishes
    // per-unit (AT) from total (AT_AT).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL @@ 5000.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    let first_child_kind = elements_of(&prices[0]).first().copied();
    assert_eq!(first_child_kind, Some(Element::Tok(AT_AT)));
    let inner_amount = prices[0].children().find(|n| n.kind() == AMOUNT).unwrap();
    assert_eq!(
        elements_of(&inner_amount),
        tok_seq(&[NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn posting_with_amount_cost_spec_and_price_annotation_all_three() {
    use SyntaxKind::*;
    // Full posting form: units AMOUNT, COST_SPEC, PRICE_ANNOTATION
    // in canonical order. Each gets its own wrapper inside POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {500.00 USD} @ 510.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let posting_kids = elements_of(&ps[0]);
    // Locate the three sub-nodes.
    let amount_count = posting_kids
        .iter()
        .filter(|e| matches!(e, Element::Node(AMOUNT)))
        .count();
    let cost_count = posting_kids
        .iter()
        .filter(|e| matches!(e, Element::Node(COST_SPEC)))
        .count();
    let price_count = posting_kids
        .iter()
        .filter(|e| matches!(e, Element::Node(PRICE_ANNOTATION)))
        .count();
    assert_eq!(amount_count, 1, "one units AMOUNT");
    assert_eq!(cost_count, 1, "one COST_SPEC");
    assert_eq!(price_count, 1, "one PRICE_ANNOTATION");

    // Source-order: AMOUNT before COST_SPEC before PRICE_ANNOTATION.
    let amount_idx = posting_kids
        .iter()
        .position(|e| matches!(e, Element::Node(AMOUNT)))
        .unwrap();
    let cost_idx = posting_kids
        .iter()
        .position(|e| matches!(e, Element::Node(COST_SPEC)))
        .unwrap();
    let price_idx = posting_kids
        .iter()
        .position(|e| matches!(e, Element::Node(PRICE_ANNOTATION)))
        .unwrap();
    assert!(amount_idx < cost_idx);
    assert!(cost_idx < price_idx);
}

#[test]
fn posting_amount_and_trailing_comment_keeps_comment_outside_amount() {
    use SyntaxKind::*;
    // A trailing same-line comment after the amount stays as a
    // flat POSTING child, NOT inside AMOUNT (the amount wrapper
    // closes at the last currency / number).
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  100 USD ; trailing\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    // The COMMENT token lives at POSTING level, not inside AMOUNT.
    let amount_comments = amts[0]
        .descendants_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(amount_comments, 0, "comment is not absorbed into AMOUNT");
    let ps = postings(&tree);
    let posting_comments = ps[0]
        .children_with_tokens()
        .filter(|e| e.kind() == COMMENT)
        .count();
    assert_eq!(posting_comments, 1, "comment is a POSTING flat child");
}

#[test]
fn posting_attached_meta_entry_after_amount_still_attaches() {
    use SyntaxKind::*;
    // Combines AMOUNT wrapping with the indent-aware meta
    // attribution from PR 2.2b: a deeper-indented META_KEY after
    // a posting with an AMOUNT still attaches to the POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {500.00 USD}\n\
                  \x20\x20\x20\x20note: \"posting-attached\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amount_count = ps[0].children().filter(|n| n.kind() == AMOUNT).count();
    let cost_count = ps[0].children().filter(|n| n.kind() == COST_SPEC).count();
    let meta_count = ps[0].children().filter(|n| n.kind() == META_ENTRY).count();
    assert_eq!(amount_count, 1);
    assert_eq!(cost_count, 1);
    assert_eq!(meta_count, 1);
}

#[test]
fn hash_flagged_posting_with_amount_wraps_both() {
    use SyntaxKind::*;
    // HASH flag + AMOUNT together — pins that the flag-flagged
    // arm of starts_posting_sub_line still allows AMOUNT wrapping
    // inside the same POSTING.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20# Assets:Cash  -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amount_count = ps[0].children().filter(|n| n.kind() == AMOUNT).count();
    assert_eq!(amount_count, 1);
    // The HASH and ACCOUNT are flat siblings, then AMOUNT.
    let first_four: Vec<Element> = elements_of(&ps[0]).into_iter().take(4).collect();
    assert_eq!(
        first_four,
        vec![
            Element::Tok(WHITESPACE),
            Element::Tok(HASH),
            Element::Tok(WHITESPACE),
            Element::Tok(ACCOUNT),
        ],
    );
}

#[test]
fn amount_with_only_negative_number_no_currency() {
    use SyntaxKind::*;
    // `MINUS NUMBER` without a CURRENCY (incomplete amount).
    // AMOUNT wraps MINUS NUMBER only.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -100\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(elements_of(&amts[0]), tok_seq(&[MINUS, NUMBER]));
}

#[test]
fn price_annotation_without_amount_still_wraps_opener_only() {
    use SyntaxKind::*;
    // Degenerate `@` with no following amount (malformed).
    // PRICE_ANNOTATION wraps the AT opener and stops; round-trip
    // is preserved.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL @\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    assert_eq!(elements_of(&prices[0]), tok_seq(&[AT]));
}

#[test]
fn cost_spec_with_inner_label_and_date_stays_flat_internally() {
    use SyntaxKind::*;
    // `{NUMBER CURRENCY, "label", DATE}` — multi-component cost.
    // COST_SPEC's internal structure is flat for phase 2.2c;
    // phase 3 typed-AST will surface accessors.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {500.00 USD, \"lot1\", 2024-01-15}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    let kinds: Vec<SyntaxKind> = elements_of(&cs[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&NUMBER));
    assert!(kinds.contains(&CURRENCY));
    assert!(kinds.contains(&STRING));
    assert!(kinds.contains(&DATE));
    assert!(kinds.contains(&L_BRACE));
    assert!(kinds.contains(&R_BRACE));
}

#[test]
fn amount_wraps_number_and_currency_with_no_whitespace_between() {
    use SyntaxKind::*;
    // The lexer's NUMBER and CURRENCY regexes are exclusive
    // (NUMBER stops at a non-digit, CURRENCY starts on an
    // uppercase letter), so `1USD` lexes as adjacent NUMBER +
    // CURRENCY with NO WHITESPACE between. Real corpus shape
    // (e.g., beancount-import fixtures). AMOUNT must still wrap
    // the CURRENCY despite the missing separator — regression for
    // the round-1 review finding where emit_amount required a WS
    // token between NUMBER and CURRENCY.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 1USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let amts = amounts(&tree);
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[NUMBER, CURRENCY]),
        "AMOUNT wraps both NUMBER and adjacent CURRENCY (no WS between)",
    );
}

#[test]
fn price_annotation_with_negative_amount_wraps_sign_inside_nested_amount() {
    use SyntaxKind::*;
    // `@ -5 USD` — negative price. The nested AMOUNT contains
    // MINUS NUMBER WS CURRENCY. Pins the sign-detection path
    // inside emit_price_annotation.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL @ -5.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    let inner_amount = prices[0].children().find(|n| n.kind() == AMOUNT).unwrap();
    assert_eq!(
        elements_of(&inner_amount),
        tok_seq(&[MINUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn cost_spec_empty_braces_wraps_open_close_pair_only() {
    use SyntaxKind::*;
    // `{}` — empty cost spec (Beancount accepts this as a
    // "no-cost" marker). COST_SPEC contains exactly the L_BRACE +
    // R_BRACE pair.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    assert_eq!(elements_of(&cs[0]), tok_seq(&[L_BRACE, R_BRACE]));
}

#[test]
fn cost_spec_with_merge_star_keeps_star_inside_node() {
    use SyntaxKind::*;
    // `{*}` — merge marker. Per the legacy parser, the STAR
    // inside braces signals lot merging. emit_cost_spec keeps
    // STAR as a flat content token between the braces.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL {*}\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let cs = cost_specs(&tree);
    assert_eq!(cs.len(), 1);
    assert_eq!(elements_of(&cs[0]), tok_seq(&[L_BRACE, STAR, R_BRACE]));
}

#[test]
fn price_annotation_at_eof_without_newline_still_wraps_opener_only() {
    use SyntaxKind::*;
    // Per rule 5, an unterminated PRICE_ANNOTATION at EOF (no
    // newline, no amount) still wraps — the PRICE_ANNOTATION
    // contains just the AT opener.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 10 HOOL @";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    assert_eq!(elements_of(&prices[0]), tok_seq(&[AT]));
}

#[test]
fn total_price_annotation_at_at_eof_without_newline_still_wraps_opener_only() {
    use SyntaxKind::*;
    // Companion to the single-`@` EOF test: pin that `@@` (total
    // price opener) at EOF also wraps. Both code paths emit
    // through the same opener branch in emit_price_annotation,
    // but the AT_AT-specific path was unpinned.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash 10 HOOL @@";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let prices = price_annotations(&tree);
    assert_eq!(prices.len(), 1);
    assert_eq!(elements_of(&prices[0]), tok_seq(&[AT_AT]));
}

#[test]
fn balance_and_price_directive_header_amounts_stay_flat_not_wrapped() {
    use SyntaxKind::*;
    // Phase 2.2c scopes AMOUNT wrapping to POSTING only. BALANCE
    // and PRICE directive headers emit their inline NUMBER +
    // CURRENCY tokens flat via `emit_through_terminator`, NOT
    // wrapped in an AMOUNT node. Phase 3 typed-AST
    // `Balance::amount()` / `Price::amount()` will need a token-
    // walking strategy distinct from `Posting::amount()`. Pinned
    // here so a future refactor that unifies the code path (e.g.,
    // calling emit_amount from the directive header) is a visible,
    // intentional break rather than a silent design shift.
    //
    // Each sub-case also asserts the directive's specific kind, so
    // a regression that drops keyword recognition (and falls
    // through to flat passthrough under SOURCE_FILE) doesn't
    // trivially satisfy the `amts.len() == 0` assertion.
    let source = "2024-06-30 balance Assets:Cash 100.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);
    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), BALANCE_DIRECTIVE);
    assert_eq!(
        amounts(&tree).len(),
        0,
        "BALANCE header NUMBER + CURRENCY are flat children of BALANCE_DIRECTIVE",
    );

    let price_source = "2024-01-01 price USD 1.10 EUR\n";
    let price_tree = parse_structured(price_source);
    assert_round_trip(price_source, &price_tree);
    let price_ds = directives(&price_tree);
    assert_eq!(price_ds.len(), 1);
    assert_eq!(price_ds[0].kind(), PRICE_DIRECTIVE);
    assert_eq!(
        amounts(&price_tree).len(),
        0,
        "PRICE header NUMBER + CURRENCY are flat children of PRICE_DIRECTIVE",
    );
}

#[test]
fn amount_wraps_arithmetic_no_spaces() {
    use SyntaxKind::*;
    // Phase 2.4 closes the 2.2c.1 divergence: `10+5 USD` is a
    // SINGLE AMOUNT containing the full expression run, matching
    // Python beancount's `parse_expr` (verified: `bean-check`
    // accepts the form). All tokens are flat children of AMOUNT;
    // phase 3 typed-AST will surface evaluated value via inspecting
    // the children.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10+5 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[NUMBER, PLUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_arithmetic_with_spaces_around_op() {
    use SyntaxKind::*;
    // `100 + 5 USD` (spaces around operator) — AMOUNT consumes the
    // whole `NUMBER WS PLUS WS NUMBER` run plus the trailing
    // `WS CURRENCY`.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  100 + 5 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[
            NUMBER, WHITESPACE, PLUS, WHITESPACE, NUMBER, WHITESPACE, CURRENCY
        ]),
    );
}

#[test]
fn amount_wraps_signed_arithmetic_negative_outer() {
    use SyntaxKind::*;
    // `-10+5 USD` — leading MINUS, then NUMBER PLUS NUMBER, then
    // CURRENCY. All inside a single AMOUNT.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -10+5 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[MINUS, NUMBER, PLUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_parenthesized_subexpression() {
    use SyntaxKind::*;
    // `-(10+5) USD` — MINUS L_PAREN NUMBER PLUS NUMBER R_PAREN
    // WS CURRENCY, all inside one AMOUNT. Phase 2.4 handles paren
    // groups via balanced-depth scanning.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -(10+5) USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[
            MINUS, L_PAREN, NUMBER, PLUS, NUMBER, R_PAREN, WHITESPACE, CURRENCY
        ]),
    );
}

#[test]
fn amount_wraps_multiplication_and_division() {
    use SyntaxKind::*;
    // STAR / SLASH operators work the same as PLUS / MINUS.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10*2/4 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[NUMBER, STAR, NUMBER, SLASH, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_wraps_nested_parens_via_depth_tracking() {
    use SyntaxKind::*;
    // `((1+2))` — nested parens. emit_amount_operand tracks depth
    // so the inner R_PAREN doesn't prematurely close the outer.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  ((1+2)) USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[
            L_PAREN, L_PAREN, NUMBER, PLUS, NUMBER, R_PAREN, R_PAREN, WHITESPACE, CURRENCY
        ]),
    );
}

#[test]
fn amount_wraps_real_corpus_arithmetic_shape() {
    use SyntaxKind::*;
    // The canonical real-corpus shape `NUMBER WS STAR WS NUMBER
    // WS CURRENCY` (`700.00 * 0.1 BRL`) — appears in 6+ files in
    // `tests/compatibility/files/apyb-financeiro/`. Pins the
    // exact multi-space-around-STAR pattern that drives the
    // manifest churn from this PR; the other arithmetic tests
    // cover synthetic shapes.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Bancos:BB  700.00 * 0.1 BRL\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    assert_eq!(
        elements_of(&amts[0]),
        tok_seq(&[
            NUMBER, WHITESPACE, STAR, WHITESPACE, NUMBER, WHITESPACE, CURRENCY
        ]),
    );
}

#[test]
fn price_annotation_inner_amount_wraps_arithmetic() {
    use SyntaxKind::*;
    // `@ 5+1 USD` — the per-unit price is itself an arithmetic
    // expression. emit_price_annotation delegates to emit_amount,
    // so the inner AMOUNT should wrap the whole expression as one
    // node. Pinned because no corpus file currently exercises
    // arithmetic in PRICE_ANNOTATION; without a test, a future
    // refactor that special-cases the inner-amount path would
    // silently regress without manifest signal.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  10 HOOL @ 5+1 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    let prices: Vec<SyntaxNode> = ps[0]
        .children()
        .filter(|n| n.kind() == PRICE_ANNOTATION)
        .collect();
    assert_eq!(prices.len(), 1);
    let inner_amount = prices[0].children().find(|n| n.kind() == AMOUNT).unwrap();
    assert_eq!(
        elements_of(&inner_amount),
        tok_seq(&[NUMBER, PLUS, NUMBER, WHITESPACE, CURRENCY]),
    );
}

#[test]
fn amount_with_unclosed_paren_at_newline_stops_per_rule_5() {
    use SyntaxKind::*;
    // Rule 5 (unterminated final content) extends to paren
    // expressions inside AMOUNT: when an unclosed `(` hits NEWLINE,
    // emit_amount_operand stops emitting and the AMOUNT closes
    // with depth>0 (no R_PAREN child). The NEWLINE goes back to
    // the surrounding scope (POSTING line terminator). Pins the
    // exact tree shape so a future refactor that changes the stop
    // condition is a visible, intentional break.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  (10+5 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 1);
    let amts: Vec<SyntaxNode> = ps[0].children().filter(|n| n.kind() == AMOUNT).collect();
    assert_eq!(amts.len(), 1);
    // AMOUNT contains the open paren and its consumed-before-
    // NEWLINE content; NO R_PAREN child.
    let kinds: Vec<SyntaxKind> = elements_of(&amts[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&L_PAREN));
    assert!(!kinds.contains(&R_PAREN));
    // POSTING's NEWLINE terminator is OUTSIDE the AMOUNT (sibling).
    let posting_kids = elements_of(&ps[0]);
    let newline_after_amount = posting_kids
        .iter()
        .position(|e| matches!(e, Element::Tok(NEWLINE)));
    let amount_idx = posting_kids
        .iter()
        .position(|e| matches!(e, Element::Node(AMOUNT)))
        .unwrap();
    assert!(
        newline_after_amount.is_some_and(|n| n > amount_idx),
        "NEWLINE terminator follows the AMOUNT, not consumed inside it",
    );
}

#[test]
fn mixed_shape_sibling_postings_each_wrap_their_own_amount_or_not() {
    use SyntaxKind::*;
    // Three postings in one transaction with different shapes:
    // auto (no amount), basic (NUMBER + CURRENCY), full (with
    // COST_SPEC and PRICE_ANNOTATION). Pins that emit_posting_line
    // re-initializes its state for each posting line and doesn't
    // leak an open AMOUNT / COST_SPEC scope across siblings.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash\n\
                  \x20\x20Assets:Bank  -5.00 USD\n\
                  \x20\x20Income:Misc  10 HOOL {500.00 USD} @ 510.00 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ps = postings(&tree);
    assert_eq!(ps.len(), 3);

    // First: auto posting, no AMOUNT.
    let p0_amounts = ps[0].children().filter(|n| n.kind() == AMOUNT).count();
    let p0_costs = ps[0].children().filter(|n| n.kind() == COST_SPEC).count();
    let p0_prices = ps[0]
        .children()
        .filter(|n| n.kind() == PRICE_ANNOTATION)
        .count();
    assert_eq!(p0_amounts, 0);
    assert_eq!(p0_costs, 0);
    assert_eq!(p0_prices, 0);

    // Second: basic posting, one AMOUNT, no COST_SPEC / PRICE.
    let p1_amounts = ps[1].children().filter(|n| n.kind() == AMOUNT).count();
    let p1_costs = ps[1].children().filter(|n| n.kind() == COST_SPEC).count();
    let p1_prices = ps[1]
        .children()
        .filter(|n| n.kind() == PRICE_ANNOTATION)
        .count();
    assert_eq!(p1_amounts, 1);
    assert_eq!(p1_costs, 0);
    assert_eq!(p1_prices, 0);

    // Third: full posting, one of each.
    let p2_amounts = ps[2].children().filter(|n| n.kind() == AMOUNT).count();
    let p2_costs = ps[2].children().filter(|n| n.kind() == COST_SPEC).count();
    let p2_prices = ps[2]
        .children()
        .filter(|n| n.kind() == PRICE_ANNOTATION)
        .count();
    assert_eq!(p2_amounts, 1);
    assert_eq!(p2_costs, 1);
    assert_eq!(p2_prices, 1);
}

#[test]
fn commodity_with_metadata_wraps_full_multi_line_directive() {
    use SyntaxKind::*;
    // Per cst::trivia, a directive that carries indented metadata
    // sub-lines spans MULTIPLE LINES — the directive's last content
    // token is the last content token of its LAST sub-line, not
    // the header. The COMMODITY_DIRECTIVE node must therefore span
    // the header AND the metadata line.
    let source = "2024-01-01 commodity HOOL\n  name: \"Hooli Common shares.\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), COMMODITY_DIRECTIVE);
    // PR 2.2a: the metadata sub-line is now wrapped in META_ENTRY.
    // The directive owns the header tokens followed by the
    // META_ENTRY node (which contains the indented metadata's
    // tokens internally).
    assert_eq!(
        elements_of(&ds[0]),
        vec![
            Element::Tok(DATE),
            Element::Tok(WHITESPACE),
            Element::Tok(COMMODITY_KW),
            Element::Tok(WHITESPACE),
            Element::Tok(CURRENCY),
            Element::Tok(NEWLINE),
            Element::Node(META_ENTRY),
        ],
    );

    // Drill into the META_ENTRY: it owns the indent + key +
    // value tokens + terminator NEWLINE.
    let me = ds[0]
        .children()
        .find(|n| n.kind() == META_ENTRY)
        .expect("directive contains a META_ENTRY child");
    assert_eq!(
        elements_of(&me),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );

    // SOURCE_FILE owns ONLY the directive — no orphaned metadata.
    assert_eq!(elements_of(&tree), vec![Element::Node(COMMODITY_DIRECTIVE)]);
}

#[test]
fn open_with_multiple_metadata_lines_wraps_all_inside_directive() {
    use SyntaxKind::*;
    let source = "2024-01-01 open Assets:Cash USD\n\
                  \x20\x20description: \"main checking\"\n\
                  \x20\x20priority: \"high\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    // OPEN_DIRECTIVE node should contain header + BOTH metadata
    // lines (no orphaned content under SOURCE_FILE).
    assert_eq!(elements_of(&tree), vec![Element::Node(OPEN_DIRECTIVE)]);
}

#[test]
fn directive_with_metadata_then_next_directive() {
    use SyntaxKind::*;
    // After a metadata-carrying directive, the next directive
    // starts cleanly — the metadata-loop must stop when the indent
    // pattern ends.
    let source = "2024-01-01 open Assets:Cash USD\n\
                  \x20\x20description: \"main\"\n\
                  2024-01-02 close Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[1].kind(), CLOSE_DIRECTIVE);
}

#[test]
fn indented_comment_after_no_metadata_directive_leads_next_directive() {
    use SyntaxKind::*;
    // An indented comment AFTER a directive that has no metadata
    // is inter-directive trivia per rule 2 — it leads the NEXT
    // directive, NOT trailing into the previous one. The widening
    // of is_indented_directive_continuation must be gated on a
    // prior META_KEY in the body; otherwise this comment is
    // wrongly absorbed into the preceding directive.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20; documentation for the next directive\n\
                  2024-01-02 close Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[1].kind(), CLOSE_DIRECTIVE);

    // d1 OWNS its header NEWLINE only — no trailing trivia.
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, NEWLINE]),
        "rule 2: indented comment after header-only directive must NOT be absorbed; \
         it's inter-directive trivia leading the next directive",
    );

    // d2 leads with the indented comment + its NEWLINE.
    let d2_first = elements_of(&ds[1])
        .iter()
        .take_while(|e| !matches!(e, Element::Tok(DATE)))
        .copied()
        .collect::<Vec<_>>();
    assert_eq!(
        d2_first,
        tok_seq(&[WHITESPACE, COMMENT, NEWLINE]),
        "rule 2: leading trivia of d2 must include the inter-directive comment",
    );
}

#[test]
fn indented_comment_at_eof_after_no_metadata_directive_is_file_trailing() {
    use SyntaxKind::*;
    // An indented comment at EOF following a header-only directive
    // is file-trailing trivia per rule 4 — it attaches to
    // SOURCE_FILE, NOT inside the directive. v3's overbroad
    // widening incorrectly absorbed this; the META_KEY gate
    // restores rule 4 conformance.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20; trailing indented comment\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[DATE, WHITESPACE, OPEN_KW, WHITESPACE, ACCOUNT, NEWLINE]),
        "directive owns ONLY its header + terminator NEWLINE",
    );

    // SOURCE_FILE owns the trailing WS + COMMENT + NEWLINE.
    assert_eq!(
        elements_of(&tree),
        vec![
            Element::Node(OPEN_DIRECTIVE),
            Element::Tok(WHITESPACE),
            Element::Tok(COMMENT),
            Element::Tok(NEWLINE),
        ],
        "rule 4: indented trailing comment is file-trailing under SOURCE_FILE",
    );
}

#[test]
fn indented_comment_before_first_metadata_stays_inside_directive() {
    use SyntaxKind::*;
    // The "documentation-comment-for-the-following-field" idiom
    // — an indented `;` line BEFORE the first META_KEY. v4's per-
    // line `body_has_meta` couldn't see the META_KEY that came
    // after the comment, so v4 silently closed the directive at
    // the comment and orphaned the metadata. v5's prospective
    // upcoming_indented_block_has_meta scan catches it.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20; documentation for the next field\n\
                  \x20\x20description: \"main checking\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    // The OPEN_DIRECTIVE owns the entire input — header, the
    // documentation comment, AND the metadata line. SOURCE_FILE
    // has no orphaned children.
    assert_eq!(elements_of(&tree), vec![Element::Node(OPEN_DIRECTIVE)]);
    // Specifically: NO bare META_KEY appears as a direct child of
    // SOURCE_FILE (would mean the v4 orphaning regression).
    let sf_token_kinds: Vec<SyntaxKind> = elements_of(&tree)
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(
        !sf_token_kinds.contains(&META_KEY),
        "META_KEY orphaned to SOURCE_FILE: {sf_token_kinds:?}",
    );
}

#[test]
fn indented_comment_between_metadata_lines_stays_inside_directive() {
    use SyntaxKind::*;
    // Beancount idiom: documentation comments between metadata
    // entries. They MUST stay inside the directive — otherwise the
    // metadata that follows is orphaned to SOURCE_FILE, losing
    // structural ownership and producing a tree where bare
    // META_KEY tokens sit directly under SOURCE_FILE.
    let source = "2024-01-01 open Assets:Cash\n\
                  \x20\x20k1: \"v1\"\n\
                  \x20\x20; doc comment for k2\n\
                  \x20\x20k2: \"v2\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    // The entire multi-line body — header + k1 + indented comment
    // + k2 — must be inside the OPEN_DIRECTIVE. SOURCE_FILE owns
    // ONLY the directive node.
    assert_eq!(elements_of(&tree), vec![Element::Node(OPEN_DIRECTIVE)]);
    // Specifically: no META_KEY appears as a direct child of
    // SOURCE_FILE (would mean orphaning).
    let sf_children: Vec<SyntaxKind> = elements_of(&tree)
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(
        !sf_children.contains(&META_KEY),
        "META_KEY orphaned to SOURCE_FILE: {sf_children:?}",
    );
}

#[test]
fn blank_line_between_metadata_lines_terminates_directive() {
    use SyntaxKind::*;
    // A blank line breaks the indented-metadata run; the second
    // metadata line is NOT part of the same directive. Conservative
    // interpretation: stop at the first non-indented-meta line.
    let source = "2024-01-01 open Assets:Cash USD\n\
                  \x20\x20description: \"main\"\n\
                  \n\
                  2024-01-02 close Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[1].kind(), CLOSE_DIRECTIVE);
    // The blank-line NEWLINE leads d2 per rule 2.
    let d2_first = elements_of(&ds[1]).first().copied();
    assert_eq!(d2_first, Some(Element::Tok(NEWLINE)));
}

#[test]
fn malformed_date_then_keyword_on_next_line_is_not_a_directive() {
    // Beancount directive headers are single-line: `DATE keyword
    // ...` on ONE line. If a DATE is followed by a NEWLINE (then
    // a keyword on the next line), the identifier MUST NOT
    // recognize it as a directive — otherwise emit_through_terminator
    // would stop at the first NEWLINE and produce a node
    // containing only `[DATE, NEWLINE]`, orphaning the keyword.
    let source = "2024-01-01\nopen Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);
    // Neither line is a recognized directive — both pass through
    // flat.
    let ds = directives(&tree);
    assert!(
        ds.is_empty(),
        "DATE alone on a line is malformed; identifier must not pretend it starts an OPEN_DIRECTIVE just because the next non-trivia token (skipping the NEWLINE) happens to be OPEN_KW",
    );
}

#[test]
fn recognized_and_passthrough_can_coexist() {
    use SyntaxKind::*;
    // All four directive shapes recognized: OPTION (PR 2.3),
    // OPEN, TRANSACTION (PR 2.1b), CLOSE. Pure-passthrough lines
    // (error-recovery shapes) are now exceptional.
    let source = "option \"title\" \"My Ledger\"\n\
                  2024-01-01 open Assets:Cash\n\
                  2024-01-15 * \"Coffee\"\n\
                  2024-01-16 close Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 4);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    assert_eq!(ds[1].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[2].kind(), TRANSACTION);
    assert_eq!(ds[3].kind(), CLOSE_DIRECTIVE);
}

// ---------- Phase 2.3: edge directives (OPTION / INCLUDE / PLUGIN / CUSTOM) ----------

#[test]
fn option_directive() {
    use SyntaxKind::*;
    let source = "option \"title\" \"My Ledger\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[OPTION_KW, WHITESPACE, STRING, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn include_directive() {
    use SyntaxKind::*;
    let source = "include \"shared/2024.beancount\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), INCLUDE_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[INCLUDE_KW, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn plugin_directive_without_config() {
    use SyntaxKind::*;
    // Plugin without config string: just `plugin "module"`.
    let source = "plugin \"beancount.plugins.implicit_prices\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PLUGIN_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[PLUGIN_KW, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn plugin_directive_with_config_string() {
    use SyntaxKind::*;
    // Plugin with optional config string: `plugin "module" "config"`.
    let source = "plugin \"my.plugin\" \"{\\\"key\\\": 42}\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PLUGIN_DIRECTIVE);
    assert_eq!(
        elements_of(&ds[0]),
        tok_seq(&[PLUGIN_KW, WHITESPACE, STRING, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn custom_directive_with_string_values() {
    use SyntaxKind::*;
    // CUSTOM with a type name string + arbitrary string values.
    let source = "2024-01-01 custom \"budget\" \"Food\" \"500 USD\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CUSTOM_DIRECTIVE);
    // Header tokens: DATE WS CUSTOM_KW WS STRING WS STRING WS STRING NEWLINE.
    let kinds: Vec<SyntaxKind> = elements_of(&ds[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&DATE));
    assert!(kinds.contains(&CUSTOM_KW));
    assert_eq!(kinds.iter().filter(|&&k| k == STRING).count(), 3);
}

#[test]
fn custom_directive_with_mixed_value_types() {
    use SyntaxKind::*;
    // CUSTOM accepts a heterogeneous trailing value list: STRING,
    // ACCOUNT, NUMBER + CURRENCY (amount), DATE, BOOL_TRUE /
    // BOOL_FALSE. All stay flat inside CUSTOM_DIRECTIVE (no AMOUNT
    // wrapper at the directive-header level per phase 2.2c scope).
    let source = "2024-01-01 custom \"limits\" \"cap\" Assets:Cash 500 USD 2024-12-31 TRUE\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CUSTOM_DIRECTIVE);
    // No AMOUNT wrapper inside CUSTOM_DIRECTIVE (directive headers
    // emit flat — same as BALANCE / PRICE per
    // `balance_and_price_directive_header_amounts_stay_flat_not_wrapped`).
    let amount_count = ds[0].descendants().filter(|n| n.kind() == AMOUNT).count();
    assert_eq!(amount_count, 0);
}

#[test]
fn option_directive_with_metadata_wraps_multi_line() {
    use SyntaxKind::*;
    // Per the Directive-Terminator Rule (and PR 2.1a's body
    // shape), an OPTION_DIRECTIVE with trailing indented META_KEY
    // sub-lines spans multiple lines. The META_ENTRY wrapping
    // from PR 2.2a applies.
    let source = "option \"title\" \"My Ledger\"\n\
                  \x20\x20doc: \"primary file\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    let metas = ds[0]
        .descendants()
        .filter(|n| n.kind() == META_ENTRY)
        .count();
    assert_eq!(metas, 1);
}

#[test]
fn plugin_directive_terminates_at_next_top_level() {
    use SyntaxKind::*;
    // Adjacent top-level directives don't merge.
    let source = "plugin \"a\"\n\
                  include \"b.bean\"\n\
                  option \"c\" \"d\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 3);
    assert_eq!(ds[0].kind(), PLUGIN_DIRECTIVE);
    assert_eq!(ds[1].kind(), INCLUDE_DIRECTIVE);
    assert_eq!(ds[2].kind(), OPTION_DIRECTIVE);
}

#[test]
fn custom_directive_unterminated_at_eof_still_wraps_per_rule_5() {
    use SyntaxKind::*;
    // Rule 5: an unterminated CUSTOM directive at EOF (no final
    // newline) still gets wrapped — the directive simply has no
    // NEWLINE child.
    let source = "2024-01-01 custom \"type\" \"value\"";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CUSTOM_DIRECTIVE);
    let has_trailing_newline = elements_of(&ds[0])
        .iter()
        .any(|e| matches!(e, Element::Tok(NEWLINE)));
    assert!(!has_trailing_newline);
}

#[test]
fn include_directive_with_metadata_wraps_multi_line() {
    use SyntaxKind::*;
    // The body / metadata code path is shared across all edge
    // directives; pinning INCLUDE-with-meta complements the
    // OPTION-with-meta test and guards against a future refactor
    // that special-cases dated vs keyword directives.
    let source = "include \"shared/2024.beancount\"\n\
                  \x20\x20note: \"shared accounts\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), INCLUDE_DIRECTIVE);
    let mes: Vec<SyntaxNode> = ds[0]
        .descendants()
        .filter(|n| n.kind() == META_ENTRY)
        .collect();
    assert_eq!(mes.len(), 1);
    // Pin full META_ENTRY shape so a regression that produces a
    // structurally-wrong META_ENTRY but the same count still fails.
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn plugin_directive_with_metadata_wraps_multi_line() {
    use SyntaxKind::*;
    let source = "plugin \"my.plugin\" \"cfg\"\n\
                  \x20\x20tolerance: \"0.01\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PLUGIN_DIRECTIVE);
    let mes: Vec<SyntaxNode> = ds[0]
        .descendants()
        .filter(|n| n.kind() == META_ENTRY)
        .collect();
    assert_eq!(mes.len(), 1);
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );
}

#[test]
fn custom_directive_with_metadata_wraps_multi_line() {
    use SyntaxKind::*;
    let source = "2024-01-01 custom \"budget\" \"food\"\n\
                  \x20\x20source: \"manual\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CUSTOM_DIRECTIVE);
    let mes: Vec<SyntaxNode> = ds[0]
        .descendants()
        .filter(|n| n.kind() == META_ENTRY)
        .collect();
    assert_eq!(mes.len(), 1);
    assert_eq!(
        elements_of(&mes[0]),
        tok_seq(&[WHITESPACE, META_KEY, WHITESPACE, STRING, NEWLINE]),
    );
}

/// For trailing-inline-comment tests: assert that a directive
/// contains exactly one COMMENT child AND that COMMENT appears
/// BEFORE the directive's NEWLINE terminator. Catches a
/// regression that reorders trailing tokens (e.g., closes the
/// directive before consuming the same-line comment, putting the
/// COMMENT logically after the terminator).
fn assert_directive_has_trailing_comment_before_newline(directive: &SyntaxNode) {
    use SyntaxKind::*;
    let kids = elements_of(directive);
    let comment_idx = kids
        .iter()
        .position(|e| matches!(e, Element::Tok(COMMENT)))
        .expect("directive must contain a COMMENT child");
    let newline_idx = kids
        .iter()
        .position(|e| matches!(e, Element::Tok(NEWLINE)))
        .expect("directive must contain its NEWLINE terminator");
    assert!(
        comment_idx < newline_idx,
        "trailing same-line COMMENT must precede the NEWLINE inside the directive",
    );
    let comment_count = kids
        .iter()
        .filter(|e| matches!(e, Element::Tok(COMMENT)))
        .count();
    assert_eq!(comment_count, 1);
}

#[test]
fn option_directive_with_trailing_inline_comment_attaches_inside() {
    use SyntaxKind::*;
    // Rule 1 of `cst::trivia`: a same-line trailing `;` comment
    // attaches INSIDE the directive. The body code path
    // (`emit_through_terminator`) handles this uniformly for all
    // directive shapes; pinning the four new kinds protects
    // against a regression that splits the path.
    let source = "option \"title\" \"My Ledger\" ; an explanation\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    assert_directive_has_trailing_comment_before_newline(&ds[0]);
}

#[test]
fn include_directive_with_trailing_inline_comment_attaches_inside() {
    use SyntaxKind::*;
    let source = "include \"shared.beancount\" ; main shared file\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), INCLUDE_DIRECTIVE);
    assert_directive_has_trailing_comment_before_newline(&ds[0]);
}

#[test]
fn plugin_directive_with_trailing_inline_comment_attaches_inside() {
    use SyntaxKind::*;
    let source = "plugin \"my.plugin\" ; description\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), PLUGIN_DIRECTIVE);
    assert_directive_has_trailing_comment_before_newline(&ds[0]);
}

#[test]
fn custom_directive_with_trailing_inline_comment_attaches_inside() {
    use SyntaxKind::*;
    let source = "2024-01-01 custom \"budget\" \"food\" ; monthly cap\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 1);
    assert_eq!(ds[0].kind(), CUSTOM_DIRECTIVE);
    assert_directive_has_trailing_comment_before_newline(&ds[0]);
}

#[test]
fn all_four_edge_directives_mixed_with_dated_directives() {
    use SyntaxKind::*;
    // Smoke test: all 4 new edge directives plus an OPEN and a
    // TRANSACTION in one source. Pins that CUSTOM (the only dated
    // edge directive — dispatched via the DATE-peek arm of
    // identify_directive) coexists cleanly with both keyword-head
    // edge directives and the legacy dated/standalone ones.
    let source = "option \"title\" \"X\"\n\
                  include \"shared.bean\"\n\
                  plugin \"my.plugin\"\n\
                  2024-01-01 open Assets:Cash\n\
                  2024-01-15 * \"tx\"\n\
                  \x20\x20Assets:Cash 1 USD\n\
                  2024-01-20 custom \"note\" \"end of test\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 6);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    assert_eq!(ds[1].kind(), INCLUDE_DIRECTIVE);
    assert_eq!(ds[2].kind(), PLUGIN_DIRECTIVE);
    assert_eq!(ds[3].kind(), OPEN_DIRECTIVE);
    assert_eq!(ds[4].kind(), TRANSACTION);
    assert_eq!(ds[5].kind(), CUSTOM_DIRECTIVE);
}

// ---------- Phase 2.4: ERROR_NODE wrapping ----------

/// Walk all `ERROR_NODE` descendants in source order.
fn error_nodes(node: &SyntaxNode) -> Vec<SyntaxNode> {
    node.descendants()
        .filter(|n| n.kind() == SyntaxKind::ERROR_NODE)
        .collect()
}

#[test]
fn unknown_keyword_line_wraps_in_error_node() {
    use SyntaxKind::*;
    // `bogus ...` is not a Beancount directive shape. PR 2.4 wraps
    // the whole line in an ERROR_NODE so downstream consumers can
    // identify malformed regions instead of having to scan flat
    // SOURCE_FILE children for stray content.
    let source = "bogus \"x\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 1);
    // All tokens of the line live INSIDE the ERROR_NODE.
    let kinds: Vec<SyntaxKind> = elements_of(&errs[0])
        .iter()
        .filter_map(|e| match e {
            Element::Tok(k) => Some(*k),
            Element::Node(_) => None,
        })
        .collect();
    assert!(kinds.contains(&STRING));
    assert!(kinds.contains(&NEWLINE));
}

#[test]
fn unrecognized_dated_keyword_wraps_in_error_node() {
    // `2024-01-01 unknown ...` — DATE is recognized but the
    // following keyword isn't one of the dated-directive shapes
    // (open/close/balance/... or custom). identify_directive
    // returns None; the line wraps as ERROR_NODE.
    let source = "2024-01-01 zzz \"data\"\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 1);
    let ds = directives(&tree);
    assert_eq!(ds.len(), 0, "no real directive recognized");
}

#[test]
fn error_node_coexists_with_recognized_directives() {
    use SyntaxKind::*;
    // ERROR_NODEs sit alongside recognized directives at
    // SOURCE_FILE level. Trivia attachment follows the same rule
    // as recognized directives.
    let source = "option \"title\" \"X\"\n\
                  bogus line\n\
                  2024-01-01 open Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 1);
    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), OPTION_DIRECTIVE);
    assert_eq!(ds[1].kind(), OPEN_DIRECTIVE);
}

#[test]
fn error_node_unterminated_at_eof_still_wraps_per_rule_5() {
    use SyntaxKind::*;
    // Rule 5: unterminated final line at EOF still wraps; the
    // ERROR_NODE simply has no NEWLINE child.
    let source = "bogus content without newline";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 1);
    let has_newline = elements_of(&errs[0])
        .iter()
        .any(|e| matches!(e, Element::Tok(NEWLINE)));
    assert!(!has_newline);
}

#[test]
fn multiple_consecutive_error_lines_each_get_their_own_error_node() {
    // Two adjacent unrecognized lines produce TWO ERROR_NODE
    // siblings (one per line), mirroring how adjacent recognized
    // directives stay as separate sibling nodes.
    let source = "bogus one\n\
                  zzz two\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 2);
}

#[test]
fn error_node_leading_trivia_attaches_inside_per_rule_2() {
    use SyntaxKind::*;
    // Per rule 2, leading trivia (the blank-line NEWLINE between
    // two top-level items) attaches as LEADING content INSIDE the
    // FOLLOWING node — same rule for ERROR_NODE as for recognized
    // directives.
    let source = "2024-01-01 open Assets:Cash\n\
                  \n\
                  bogus content\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let errs = error_nodes(&tree);
    assert_eq!(errs.len(), 1);
    // ERROR_NODE's first child is the blank-line NEWLINE that
    // serves as its leading trivia.
    let first = elements_of(&errs[0]).first().copied();
    assert_eq!(first, Some(Element::Tok(NEWLINE)));
}

#[test]
fn error_node_adjacent_to_multi_line_transaction_doesnt_bleed() {
    use SyntaxKind::*;
    // Multi-line TRANSACTION coexistence: the transaction body
    // (header + posting lines) must terminate cleanly when the
    // next top-level line is unrecognized, so the unrecognized
    // line wraps as its own ERROR_NODE without being absorbed
    // into the transaction. emit_transaction_body's stop
    // condition is non-indented top-level content; a regression
    // there could silently merge an ERROR_NODE line into the
    // preceding transaction.
    let source = "2024-01-15 * \"x\"\n\
                  \x20\x20Assets:Cash  -5 USD\n\
                  \x20\x20Expenses:Food  5 USD\n\
                  bogus content here\n\
                  2024-01-16 * \"y\"\n\
                  \x20\x20Assets:Cash  -3 USD\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    let ds = directives(&tree);
    assert_eq!(ds.len(), 2);
    assert_eq!(ds[0].kind(), TRANSACTION);
    assert_eq!(ds[1].kind(), TRANSACTION);

    let errs = error_nodes(&tree);
    assert_eq!(
        errs.len(),
        1,
        "exactly one ERROR_NODE between the two transactions"
    );

    // The ERROR_NODE is a direct sibling of TRANSACTION under
    // SOURCE_FILE, NOT nested inside either transaction.
    let tx_inner_errors = ds[0]
        .descendants()
        .filter(|n| n.kind() == ERROR_NODE)
        .count()
        + ds[1]
            .descendants()
            .filter(|n| n.kind() == ERROR_NODE)
            .count();
    assert_eq!(tx_inner_errors, 0);

    // Both transactions retain their full posting count.
    let p0_count = ds[0].descendants().filter(|n| n.kind() == POSTING).count();
    let p1_count = ds[1].descendants().filter(|n| n.kind() == POSTING).count();
    assert_eq!(p0_count, 2);
    assert_eq!(p1_count, 1);
}

// ---------- Edge cases ----------

#[test]
fn empty_source() {
    let tree = parse_structured("");
    assert_round_trip("", &tree);
    assert_eq!(tree.kind(), SyntaxKind::SOURCE_FILE);
    assert!(directives(&tree).is_empty());
}

#[test]
fn only_trivia_no_directives() {
    use SyntaxKind::*;
    let source = ";; only a comment\n\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    assert!(directives(&tree).is_empty());
    // All under SOURCE_FILE.
    assert_eq!(
        elements_of(&tree),
        vec![
            Element::Tok(COMMENT),
            Element::Tok(NEWLINE),
            Element::Tok(NEWLINE)
        ],
    );
}

#[test]
fn bom_under_source_file_directive_follows() {
    use SyntaxKind::*;
    let source = "\u{FEFF}2024-01-01 open Assets:Cash\n";
    let tree = parse_structured(source);
    assert_round_trip(source, &tree);

    // BOM is file-leading; first directive comes after.
    assert_eq!(
        elements_of(&tree),
        vec![Element::Tok(BOM), Element::Node(OPEN_DIRECTIVE)],
    );
}