twig-sys 2.8.0

FFI bindings and native library for Twig (the Djot/Markdown/HTML/XML document engine). Used by the `twig-doc` crate.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
//! AsciiDoc — the parser. Source bytes to twig's shared `AST`, judged by
//! `conformance.zig`'s corpus comparison (`asg.encode(parse(source)) ==
//! case.asg`, structurally) against the vendored TCK corpus, exactly the
//! shapes `asg.zig`'s `decode` already proved have somewhere to live.
//!
//! ── Scope of THIS file, today ───────────────────────────────────────────────
//! Judged against a target that is finite and known: the official ASG JSON
//! Schema (vendored as `testdata/asg-schema.json`) enumerates the WHOLE block
//! and inline vocabulary, so "finished" here means "emits every shape in that
//! schema" rather than "keeps up with a spec still being written".
//!
//! Covered: the document header (title + the attribute entries directly below
//! it, including the `:!name:` unset spellings), paragraphs, section nesting
//! by `=` count including level-0 part titles, unordered lists (wrapped
//! principal text, blank lines between items), every delimited block the ASG
//! has a name for (`listing`, `literal`, `pass`, `example`, `quote`,
//! `sidebar`, `open`) plus comment blocks, line comments, thematic and page
//! breaks, and four spans in BOTH forms — `*strong*`, `_emphasis_`,
//! `` `code` `` (monospace) and `#mark#` (highlight), constrained, plus the
//! doubled unconstrained spellings (`**`, `__`, `` `` ` `` ``, `##`) that make
//! intraword formatting work.
//!
//! Not yet: ordered and callout lists, description lists, nested lists, list
//! continuation, block metadata (`[attrlist]`, `.Title`, ids, roles, options)
//! and everything it unlocks (admonitions, discrete headings, `[verse]`,
//! `[source]`), block macros (`image::`, `audio::`, `video::`, `toc::`), and
//! every inline besides those four spans and plain text — superscript,
//! subscript, attribute references, anchors, cross references, links, images,
//! footnotes, inline passthrough and line breaks all remain unimplemented, and
//! several of them (superscript, subscript, footnotes, inline images) have no
//! shape in the ASG schema yet either. Tables are not on the list at all: the
//! ASG does not model them as of draft-01.
//!
//! What's missing degrades to LITERAL SOURCE TEXT, which is the property that
//! makes a `format.zig` registry entry honest: an unimplemented `image:a.png[]`
//! renders as the characters `image:a.png[]`, visibly unhandled, rather than as
//! a mangled tree. The unconstrained spans were the one construct that broke
//! that rule — see `parseInlines` for what they used to do.
//!
//! ── Why this shape ───────────────────────────────────────────────────────
//! Bottom-up onto `AST.Builder`, the same posture `languages/rst/parser.zig`
//! took for reStructuredText. Section nesting is simpler here than rST's: an
//! AsciiDoc section's level is spelled directly in the source as the marker's
//! `=` count minus one, so there is no need for rST's order-of-first-use
//! style stack — `parseSection` just recurses, and `parseSectionsLoop` closes
//! back to a parent by comparing level numbers directly.
//!
//! ── Position matters here in a way it didn't for rST's first slice ────────
//! `doctree.zig`'s docutils pformat carries no source positions at all, so
//! rST's parser could set best-effort spans and let the corpus comparison
//! ignore them. The TCK's ASG carries a `location` on every node, and the
//! conformance harness compares through `asg.encode`, which writes it back —
//! so every span this file sets is load-bearing, not cosmetic. Where the
//! corpus doesn't exercise a shape (an unclosed listing block, a document
//! with leading blank lines), spans are still computed by the same rules,
//! just unverified.

const std = @import("std");
const Allocator = std.mem.Allocator;
const AST = @import("../../ast/ast.zig");
const Node = AST.Node;
const Builder = AST.Builder;
const Span = @import("../../span.zig");
const Document = @import("../../document.zig");

pub fn parse(allocator: Allocator, source: []const u8) Allocator.Error!Document {
    var p = try Parser.init(allocator, source);
    defer p.deinit();
    const root = try p.parseDocument();
    return p.b.finishDocument(source, root);
}

/// The TCK's `inline`-level entry point: `source` (a single line, e.g.
/// `"*s*\n"`) is scanned for inline markup directly, with no document/block
/// structure at all, and the resulting nodes are wrapped in a synthetic
/// `.doc` root — mirroring `asg.decode`'s own `.inlines` case exactly, so
/// `asg.encode(&doc, .inlines, ...)` (which reads straight through that
/// wrapper) can print it back.
pub fn parseInlineList(allocator: Allocator, source: []const u8) Allocator.Error!Document {
    var b = Builder.init(allocator);
    errdefer b.deinit();
    const text = std.mem.trimEnd(u8, source, "\n");
    const ids = try parseInlines(&b, text, 0);
    defer allocator.free(ids);
    const root = try b.addContainer(.doc, ids);
    return b.finishDocument(source, root);
}

/// One line of source, as a byte range excluding its terminating `\n`.
const LineInfo = struct { start: usize, end: usize };

/// A recognized `=`-prefixed title line — a document title (`level == 0`,
/// only matched at line 0) or a section title (`level >= 1`, matched
/// anywhere at column 0).
const HeadingMatch = struct {
    level: i32,
    text: []const u8,
    text_span: Span,
    /// Start of the marker itself (`=`'s own column), which is also the
    /// start of the construct the marker introduces (a header or a section).
    span_start: usize,
    next_line: usize,
};

/// A run of parsed blocks, plus the source extent they cover. `first_start`
/// and `last_end` are what a caller needs to span ITSELF: a section runs from
/// its own title to its last descendant's end, and a document with no header
/// starts at its first block rather than at line 1 — leading blank lines
/// belong to no node.
const FlatResult = struct { items: []Node.Id, stopped_at: usize, first_start: usize = 0, last_end: usize = 0 };
const SectionsResult = struct { items: []Node.Id, stopped_at: usize, first_start: usize = 0, last_end: usize = 0 };

/// One row of the delimited-block table — a delimiter character and what the
/// block it opens turns into. Content models split two ways, which is the only
/// structural difference between rows: `.verbatim` blocks keep their interior
/// as opaque text, `.compound` blocks parse theirs as nested blocks.
const Delimiter = struct {
    char: u8,
    /// The ASG block name (`listing`, `example`, …). Carried through to
    /// `asg.zig` verbatim; see its doc comment for how each maps to a `Kind`.
    name: []const u8,
    content: enum { verbatim, compound, dropped },

    /// `--`, the one delimiter that is not a run of four or more. Not in
    /// `DELIMITERS` because `matchDelimiter` has to test it by its whole
    /// spelling rather than by its character.
    const open: Delimiter = .{ .char = '-', .name = "open", .content = .compound };
};

/// The delimited blocks of docs/modules/blocks/pages/delimited.adoc, minus the
/// two the ASG has no node for: a table (`|===`) has no place in the schema at
/// all — the AsciiDoc ASG as of draft-01 does not model tables — and a comment
/// block produces no node by definition, so it is `.dropped` rather than
/// missing.
const DELIMITERS = [_]Delimiter{
    .{ .char = '-', .name = "listing", .content = .verbatim },
    .{ .char = '.', .name = "literal", .content = .verbatim },
    .{ .char = '+', .name = "pass", .content = .verbatim },
    .{ .char = '=', .name = "example", .content = .compound },
    .{ .char = '_', .name = "quote", .content = .compound },
    .{ .char = '*', .name = "sidebar", .content = .compound },
    .{ .char = '/', .name = "comment", .content = .dropped },
};

/// The document root's `current_level` in `parseSectionsLoop`. Not `0`:
/// AsciiDoc has real level-0 sections (a part title, `= Title` anywhere below
/// the header), and a loop that closed on `level <= 0` would never open one.
const ROOT_LEVEL: i32 = -1;

const Parser = struct {
    allocator: Allocator,
    source: []const u8,
    lines: []const LineInfo,
    b: Builder,

    fn init(allocator: Allocator, source: []const u8) Allocator.Error!Parser {
        return .{ .allocator = allocator, .source = source, .lines = try computeLines(allocator, source), .b = Builder.init(allocator) };
    }

    fn deinit(self: *Parser) void {
        self.allocator.free(self.lines);
        self.b.deinit();
    }

    // ── line helpers ─────────────────────────────────────────────────────

    fn lineText(self: *const Parser, i: usize) []const u8 {
        return self.source[self.lines[i].start..self.lines[i].end];
    }

    fn leadingSpaces(self: *const Parser, i: usize) usize {
        const t = self.lineText(i);
        var n: usize = 0;
        while (n < t.len and t[n] == ' ') n += 1;
        return n;
    }

    fn isBlankLine(self: *const Parser, i: usize) bool {
        for (self.lineText(i)) |c| {
            if (c != ' ' and c != '\t') return false;
        }
        return true;
    }

    // ── title / heading recognition ─────────────────────────────────────

    /// Any `=`-run + space + text line at column 0 — level `n - 1` for `n`
    /// leading `=` characters. Every level this matches is a real section
    /// level, INCLUDING 0: a `= Title` line below the document header is a
    /// part title, not a second document title. Only `parseDocument`'s own
    /// header scan treats a level-0 line specially, and only at the document's
    /// first non-blank line.
    fn matchHeadingLine(self: *const Parser, i: usize) ?HeadingMatch {
        if (i >= self.lines.len) return null;
        if (self.leadingSpaces(i) != 0) return null;
        const t = self.lineText(i);
        var n: usize = 0;
        while (n < t.len and t[n] == '=') : (n += 1) {}
        if (n == 0 or n >= t.len or t[n] != ' ') return null;
        const text = std.mem.trimEnd(u8, t[n + 1 ..], " \t");
        if (text.len == 0) return null;
        const text_start = self.lines[i].start + n + 1;
        return .{
            .level = @intCast(n - 1),
            .text = text,
            .text_span = Span.init(text_start, text_start + text.len),
            .span_start = self.lines[i].start,
            .next_line = i + 1,
        };
    }

    /// One `:name: value` attribute entry. `value` is trimmed of surrounding
    /// whitespace but not internally; a bare `:name:` yields an empty but
    /// NON-null value, matching the TCK's own `{"toc": ""}`, while the two
    /// negated spellings (`:!name:` and `:name!:`) yield `null` — the ASG's
    /// document `attributes` map admits null for exactly this.
    fn matchAttrEntry(self: *const Parser, i: usize) ?AST.KeyVal {
        const t = self.lineText(i);
        if (t.len < 2 or t[0] != ':') return null;
        var j: usize = 1;
        while (j < t.len and t[j] != ':') : (j += 1) {}
        if (j >= t.len or j == 1) return null;

        var key = t[1..j];
        var unset = false;
        if (key[0] == '!') {
            key = key[1..];
            unset = true;
        } else if (key[key.len - 1] == '!') {
            key = key[0 .. key.len - 1];
            unset = true;
        }
        if (key.len == 0) return null;
        return .{
            .key = key,
            .value = if (unset) null else std.mem.trim(u8, t[j + 1 ..], " \t"),
        };
    }

    // ── delimited-block / list-item recognition ─────────────────────────

    /// A delimited block's opening line, if line `i` is one. See `Delimiter`
    /// for the table this matches against.
    fn matchDelimiter(self: *const Parser, i: usize) ?struct { delim: Delimiter, text: []const u8 } {
        if (i >= self.lines.len) return null;
        if (self.leadingSpaces(i) != 0) return null;
        const t = std.mem.trimEnd(u8, self.lineText(i), " \t");
        if (t.len < 2) return null;

        // The open block is the one exception to "four or more": it is spelled
        // with exactly two hyphens, which is also why it has to be tested
        // before the run-length check below would read it as a short listing.
        if (std.mem.eql(u8, t, "--")) return .{ .delim = Delimiter.open, .text = t };
        if (t.len < 4) return null;
        for (t[1..]) |c| {
            if (c != t[0]) return null;
        }
        for (DELIMITERS) |d| {
            if (d.char == t[0]) return .{ .delim = d, .text = t };
        }
        return null;
    }

    /// Does line `i` close a block opened with exactly `opening`? A closing
    /// line must match the opening one character for character — a longer or
    /// shorter run of the same character does not close it.
    fn matchesDelim(self: *const Parser, i: usize, opening: []const u8) bool {
        if (self.leadingSpaces(i) != 0) return false;
        return std.mem.eql(u8, std.mem.trimEnd(u8, self.lineText(i), " \t"), opening);
    }

    fn isListItem(self: *const Parser, i: usize) bool {
        if (self.leadingSpaces(i) != 0) return false;
        const t = self.lineText(i);
        if (t.len < 2) return false;
        return (t[0] == '*' or t[0] == '-' or t[0] == '+') and t[1] == ' ';
    }

    /// `'''` — a thematic break. Three or more of `'`, `-`, `*` or `_` on
    /// their own line is Asciidoctor's "markdown-style" spelling too, but only
    /// `'''` is unambiguous: the other three collide with the delimited-block
    /// and list-marker syntax, which is checked first anyway.
    fn isThematicBreak(self: *const Parser, i: usize) bool {
        if (self.leadingSpaces(i) != 0) return false;
        const t = std.mem.trimEnd(u8, self.lineText(i), " \t");
        if (t.len < 3) return false;
        for (t) |c| {
            if (c != '\'') return false;
        }
        return true;
    }

    /// `<<<` — a page break.
    fn isPageBreak(self: *const Parser, i: usize) bool {
        if (self.leadingSpaces(i) != 0) return false;
        return std.mem.eql(u8, std.mem.trimEnd(u8, self.lineText(i), " \t"), "<<<");
    }

    /// `// ...` — a line comment. Produces no node at all, so it is skipped
    /// wherever a block could start. `///` and longer are a comment BLOCK's
    /// delimiter and are matched before this.
    fn isLineComment(self: *const Parser, i: usize) bool {
        if (self.leadingSpaces(i) != 0) return false;
        const t = self.lineText(i);
        return t.len >= 2 and t[0] == '/' and t[1] == '/' and !(t.len >= 4 and t[2] == '/' and t[3] == '/');
    }

    fn isBlockStart(self: *const Parser, i: usize) bool {
        return self.matchHeadingLine(i) != null or self.matchDelimiter(i) != null or
            self.isListItem(i) or self.isThematicBreak(i) or self.isPageBreak(i) or
            self.isLineComment(i);
    }

    // ── the top-level document scan ─────────────────────────────────────

    fn firstNonBlankLine(self: *const Parser, from: usize) usize {
        var i = from;
        while (i < self.lines.len and self.isBlankLine(i)) i += 1;
        return i;
    }

    fn parseDocument(self: *Parser) Allocator.Error!Node.Id {
        var header_id: ?Node.Id = null;
        var attrs_id: ?Node.Id = null;
        var header_end_offset: usize = 0;
        var start_offset: usize = 0;

        // The header, if any, is at the first NON-BLANK line — leading blank
        // lines belong to no node, and neither the document nor its header
        // starts at line 1 just because the file does.
        var body_start_line = self.firstNonBlankLine(0);
        const doc_title: ?HeadingMatch = if (self.matchHeadingLine(body_start_line)) |hm|
            (if (hm.level == 0) hm else null)
        else
            null;

        if (doc_title) |dt| {
            var header_end_line = dt.next_line;
            var attr_entries: std.ArrayList(AST.KeyVal) = .empty;
            defer attr_entries.deinit(self.allocator);
            header_end_offset = self.lines[body_start_line].end;
            while (header_end_line < self.lines.len and !self.isBlankLine(header_end_line)) {
                const entry = self.matchAttrEntry(header_end_line) orelse break;
                try attr_entries.append(self.allocator, entry);
                header_end_offset = self.lines[header_end_line].end;
                header_end_line += 1;
            }

            const title_str = try self.b.addLeaf(.{ .str = dt.text });
            self.b.setSpan(title_str, dt.text_span);
            // `level = 1`, not the ASG's own `0`: twig's shared `heading.level`
            // is 1-based across every language (djot/Markdown's `#` is 1), and
            // `languages/html/serializer.zig` prints it straight into `h{d}` —
            // storing the ASG's number here emitted `<h0>`, which is not an
            // HTML element, and put every AsciiDoc heading one off from the
            // same heading written in Markdown. `asg.zig` shifts back on encode.
            const heading_id = try self.b.addContainer(.{ .heading = .{ .level = 1 } }, &.{title_str});
            self.b.setSpan(heading_id, Span.init(dt.span_start, header_end_offset));
            header_id = heading_id;
            start_offset = dt.span_start;

            // The attributes marker is present whenever a header is, even
            // with zero entries — see `asg.zig`'s doc comment: its mere
            // presence as a child (not whether `setAttrs` ran) is what
            // signals `encode` to write the `"attributes"` key at all.
            const attrs_marker = try self.b.addNode(.{ .container = .{ .name = "document-attributes" } });
            if (attr_entries.items.len > 0) try self.b.setAttrs(attrs_marker, .{ .entries = attr_entries.items });
            attrs_id = attrs_marker;

            body_start_line = header_end_line;
        }

        const body = try self.parseSectionsLoop(body_start_line, self.lines.len, ROOT_LEVEL);
        defer self.allocator.free(body.items);

        var children: std.ArrayList(Node.Id) = .empty;
        defer children.deinit(self.allocator);
        if (attrs_id) |aid| try children.append(self.allocator, aid);
        if (header_id) |hid| try children.append(self.allocator, hid);
        try children.appendSlice(self.allocator, body.items);

        if (header_id == null and body.items.len > 0) start_offset = body.first_start;
        const end_offset = if (body.items.len > 0) body.last_end else header_end_offset;
        const root = try self.b.addContainer(.doc, children.items);
        self.b.setSpan(root, Span.init(start_offset, end_offset));
        return root;
    }

    // ── section nesting ───────────────────────────────────────────────────

    /// Parses the flat block run starting at `lo`, opening nested sections
    /// (and their whole subtrees) as their headings are found, until a
    /// heading at or above `current_level` closes this level back to its
    /// caller — the base case being `current_level == 0` at the document
    /// root, which no real section level is `<=` to.
    fn parseSectionsLoop(self: *Parser, lo: usize, hi: usize, current_level: i32) Allocator.Error!SectionsResult {
        var children: std.ArrayList(Node.Id) = .empty;
        errdefer children.deinit(self.allocator);
        var i = lo;
        var first_start: usize = 0;
        var last_end: usize = 0;
        while (true) {
            const flat = try self.parseFlatBlocks(i, hi);
            if (flat.items.len > 0) {
                if (children.items.len == 0) first_start = flat.first_start;
                last_end = flat.last_end;
            }
            try children.appendSlice(self.allocator, flat.items);
            self.allocator.free(flat.items);
            i = flat.stopped_at;
            if (i >= hi) break;

            const hm = self.matchHeadingLine(i).?; // `parseFlatBlocks` only stops early for this
            if (hm.level <= current_level) break;
            const r = try self.parseSection(hm, hi);
            if (children.items.len == 0) first_start = hm.span_start;
            try children.append(self.allocator, r.id);
            last_end = r.end_offset;
            i = r.next;
        }
        return .{
            .items = try children.toOwnedSlice(self.allocator),
            .stopped_at = i,
            .first_start = first_start,
            .last_end = last_end,
        };
    }

    fn parseSection(self: *Parser, hm: HeadingMatch, hi: usize) Allocator.Error!struct { id: Node.Id, next: usize, end_offset: usize } {
        const title_str = try self.b.addLeaf(.{ .str = hm.text });
        self.b.setSpan(title_str, hm.text_span);
        // `+ 1` for the same reason the document title above is 1 — the ASG's
        // levels are 0-based and twig's shared tree is not. `hm.level` itself
        // stays ASG-shaped, since that is what the nesting comparisons and
        // `ROOT_LEVEL` are written against.
        const heading_id = try self.b.addContainer(.{ .heading = .{ .level = @intCast(hm.level + 1) } }, &.{title_str});
        self.b.setSpan(heading_id, hm.text_span);

        const inner = try self.parseSectionsLoop(hm.next_line, hi, hm.level);
        defer self.allocator.free(inner.items);

        const all = try self.allocator.alloc(Node.Id, 1 + inner.items.len);
        defer self.allocator.free(all);
        all[0] = heading_id;
        @memcpy(all[1..], inner.items);

        const id = try self.b.addContainer(.section, all);
        const end_offset = if (inner.items.len > 0) inner.last_end else hm.text_span.end;
        self.b.setSpan(id, Span.init(hm.span_start, end_offset));
        return .{ .id = id, .next = inner.stopped_at, .end_offset = end_offset };
    }

    // ── the flat block scan ────────────────────────────────────────────

    /// Parses blocks in `[lo, hi)`, stopping at `hi` or at the first section
    /// heading (any level) — the only construct a caller (`parseSectionsLoop`)
    /// needs to inspect rather than have consumed outright.
    fn parseFlatBlocks(self: *Parser, lo: usize, hi: usize) Allocator.Error!FlatResult {
        var children: std.ArrayList(Node.Id) = .empty;
        errdefer children.deinit(self.allocator);
        var i = lo;
        var first_start: usize = 0;
        var last_end: usize = 0;
        while (i < hi) {
            if (self.isBlankLine(i)) {
                i += 1;
                continue;
            }
            if (self.matchHeadingLine(i) != null) break;

            // Where this iteration's block starts, recorded before `i` moves.
            // Only committed to `first_start` once a block is actually
            // appended: a comment produces no node, and a document whose first
            // line is a comment starts at whatever follows it.
            const block_start = self.lines[i].start;
            const had = children.items.len;

            if (self.isLineComment(i)) {
                i += 1;
                continue;
            } else if (self.isThematicBreak(i) or self.isPageBreak(i)) {
                const id = if (self.isThematicBreak(i))
                    try self.b.addNode(.thematic_break)
                else
                    try self.b.addNode(.{ .container = .{ .name = "page-break" } });
                const span = Span.init(block_start, self.lines[i].end);
                self.b.setSpan(id, span);
                try children.append(self.allocator, id);
                last_end = span.end;
                i += 1;
            } else if (self.matchDelimiter(i)) |d| {
                const r = try self.parseDelimited(i, hi, d.delim, d.text);
                if (r.id) |id| {
                    try children.append(self.allocator, id);
                    last_end = r.end_offset;
                }
                i = r.next;
            } else if (self.isListItem(i)) {
                const r = try self.parseList(i, hi);
                try children.append(self.allocator, r.id);
                last_end = r.end_offset;
                i = r.next;
            } else {
                const end = self.paragraphEnd(i, hi);
                const p = try self.parseParagraph(i, end);
                try children.append(self.allocator, p.id);
                last_end = p.end_offset;
                i = end;
            }

            if (had == 0 and children.items.len > 0) first_start = block_start;
        }
        return .{
            .items = try children.toOwnedSlice(self.allocator),
            .stopped_at = i,
            .first_start = first_start,
            .last_end = last_end,
        };
    }

    fn paragraphEnd(self: *const Parser, lo: usize, hi: usize) usize {
        var i = lo + 1;
        while (i < hi) {
            if (self.isBlankLine(i) or self.isBlockStart(i)) break;
            i += 1;
        }
        return i;
    }

    /// A paragraph's text is exactly `source[lines[lo].start..lines[end_ex-1].end]`
    /// — a hard-wrapped paragraph's embedded `\n`s are real source bytes, not
    /// synthesized, so no line-by-line reassembly is needed the way rST's
    /// indent-stripping `assembleText` requires.
    fn parseParagraph(self: *Parser, lo: usize, end_ex: usize) Allocator.Error!struct { id: Node.Id, end_offset: usize } {
        const span = Span.init(self.lines[lo].start, self.lines[end_ex - 1].end);
        const text = self.source[span.start..span.end];
        const ids = try parseInlines(&self.b, text, span.start);
        defer self.allocator.free(ids);
        const id = try self.b.addContainer(.para, ids);
        self.b.setSpan(id, span);
        return .{ .id = id, .end_offset = span.end };
    }

    /// One delimited block, from its opening delimiter line. `.verbatim`
    /// blocks keep their interior as opaque text on a `code_block`;
    /// `.compound` blocks parse theirs as nested blocks; `.dropped` (a comment
    /// block) yields no node at all, which is why `id` is optional.
    ///
    /// An UNCLOSED block runs to `hi` — the end of its enclosing block, or of
    /// the document. Asciidoctor warns and does the same; the ASG has nowhere
    /// to record the warning, so all that is left is the extent, and ending at
    /// the last content line (rather than at a delimiter that isn't there) is
    /// the only reading that keeps every location inside the source.
    fn parseDelimited(
        self: *Parser,
        delim_line: usize,
        hi: usize,
        delim: Delimiter,
        opening: []const u8,
    ) Allocator.Error!struct { id: ?Node.Id, next: usize, end_offset: usize } {
        var close_line = delim_line + 1;
        while (close_line < hi and !self.matchesDelim(close_line, opening)) close_line += 1;
        const closed = close_line < hi;
        const content_lo = delim_line + 1;
        const content_hi = close_line;
        const has_content = content_hi > content_lo;

        const next = if (closed) close_line + 1 else content_hi;

        if (delim.content == .dropped) return .{ .id = null, .next = next, .end_offset = 0 };

        const id = switch (delim.content) {
            .verbatim => blk: {
                var text: []const u8 = "";
                if (has_content) {
                    const content = Span.init(self.lines[content_lo].start, self.lines[content_hi - 1].end);
                    text = self.source[content.start..content.end];
                    const leaf = try self.b.addLeaf(.{ .code_block = .{ .lang = null, .text = text } });
                    self.b.setContentSpan(leaf, content);
                    break :blk leaf;
                }
                break :blk try self.b.addLeaf(.{ .code_block = .{ .lang = null, .text = text } });
            },
            .compound => blk: {
                const inner = try self.parseFlatBlocks(content_lo, content_hi);
                defer self.allocator.free(inner.items);
                break :blk try self.b.addContainer(kindForCompound(delim.name), inner.items);
            },
            .dropped => unreachable, // returned above
        };

        const end_offset = if (closed)
            self.lines[close_line].end
        else if (has_content)
            self.lines[content_hi - 1].end
        else
            self.lines[delim_line].end;
        self.b.setSpan(id, Span.init(self.lines[delim_line].start, end_offset));
        // The block's ASG `name`/`form`/`delimiter` are NOT recorded here.
        // `asg.zig` derives all three (`blockDelimiter`, `codeBlockName`,
        // `parentBlockName`) from the node's kind and its own source, because
        // `AST.Attrs` is the channel a document's REAL attributes travel in and
        // `languages/html/serializer.zig` renders whatever it finds there —
        // storing them put `<pre name="listing" form="delimited"
        // delimiter="----">` into rendered HTML.
        return .{ .id = id, .next = next, .end_offset = end_offset };
    }

    /// Does line `i` open an item of a list marked with `marker_char`?
    fn isItemOfList(self: *const Parser, i: usize, marker_char: u8) bool {
        if (!self.isListItem(i)) return false;
        return self.lineText(i)[0] == marker_char;
    }

    /// One list, from its first item's marker line. Two rules beyond "one line
    /// per item" (docs/modules/lists/pages/build-a-list.adoc):
    ///
    ///   * An item's PRINCIPAL TEXT continues onto any following line that is
    ///     neither blank nor the start of another block, indentation and
    ///     newlines kept verbatim — the same treatment a hard-wrapped
    ///     paragraph's interior gets, and for the same reason (the ASG's text
    ///     nodes carry locations, so their values have to be real source).
    ///   * A run of blank lines does NOT close the list as long as another
    ///     item of the SAME marker follows. Blank lines only make the list
    ///     loose, which the ASG has nowhere to record; a different marker
    ///     character starts a sibling list instead of continuing this one.
    fn parseList(self: *Parser, lo: usize, hi: usize) Allocator.Error!struct { id: Node.Id, next: usize, end_offset: usize } {
        const marker_char = self.lineText(lo)[0];
        var items: std.ArrayList(Node.Id) = .empty;
        defer items.deinit(self.allocator);
        var i = lo;
        var last_end: usize = self.lines[lo].end;
        var next_line = lo;
        while (i < hi and self.isItemOfList(i, marker_char)) {
            var end_line = i + 1;
            while (end_line < hi and !self.isBlankLine(end_line) and !self.isBlockStart(end_line)) end_line += 1;

            const text_start = self.lines[i].start + 2;
            const text_end = self.lines[end_line - 1].end;
            const text = std.mem.trimEnd(u8, self.source[text_start..text_end], " \t");
            const inline_ids = try parseInlines(&self.b, text, text_start);
            defer self.allocator.free(inline_ids);
            const item_id = try self.b.addContainer(.list_item, inline_ids);
            const item_end = text_start + text.len;
            self.b.setSpan(item_id, Span.init(self.lines[i].start, item_end));
            self.b.setSpelling(item_id, .{ .bullet = bulletFromChar(marker_char) });
            try items.append(self.allocator, item_id);
            last_end = item_end;
            next_line = end_line;

            // Look past any blank lines for another item of this same list;
            // stop (leaving `next_line` before the blanks) if there isn't one.
            i = self.firstNonBlankLine(end_line);
            if (i >= hi or !self.isItemOfList(i, marker_char)) break;
        }

        const id = try self.b.addContainer(.{ .bullet_list = .{ .tight = true } }, items.items);
        self.b.setSpan(id, Span.init(self.lines[lo].start, last_end));
        self.b.setSpelling(id, .{ .bullet = bulletFromChar(marker_char) });
        return .{ .id = id, .next = next_line, .end_offset = last_end };
    }
};

/// The twig `Kind` a compound delimited block becomes. Only `quote` has a
/// semantic kind waiting for it (`block_quote` means exactly that); the rest
/// are AsciiDoc-specific containers with no counterpart anywhere else in
/// twig's vocabulary, so they take the generic escape hatch — see `asg.zig`'s
/// doc comment for the full mapping and why. Either way the ASG's own name
/// rides in `attrs`, so `encode` reads one place rather than switching twice.
fn kindForCompound(name: []const u8) Node.Kind {
    if (std.mem.eql(u8, name, "quote")) return .block_quote;
    return .{ .container = .{ .name = name } };
}

fn bulletFromChar(ch: u8) Document.Spelling.Bullet {
    return switch (ch) {
        '*' => .star,
        '-' => .dash,
        '+' => .plus,
        else => unreachable, // only called after `isListItem` validated `ch`
    };
}

fn isWordByte(c: u8) bool {
    return std.ascii.isAlphanumeric(c) or c == '_';
}

fn emitText(b: *Builder, s: []const u8, offset: usize) Allocator.Error!Node.Id {
    const id = try b.addLeaf(.{ .str = s });
    b.setSpan(id, Span.init(offset, offset + s.len));
    return id;
}

/// The single-character constrained spans that nest — `*strong*`, `_emphasis_`,
/// `#mark#` — spelled by `parseInlines`' shared word-boundary algorithm and
/// distinguished only by delimiter byte and which `AST.InlineMark` they build.
/// `` `monospace` `` follows the identical boundary rule but is handled
/// separately (see `tryMonospace`): its ASG shape is a `code` span like these
/// three, but `AST.InlineMark` is deliberately scoped to exactly the marks
/// djot itself spells (see `languages/djot/syntax.zig`'s doc comment), so
/// monospace text is kept as a `text_leaf{.verbatim}` leaf instead of growing
/// that enum for one AsciiDoc-only member.
const MARK_SPANS = [_]struct { char: u8, mark: AST.InlineMark }{
    .{ .char = '*', .mark = .strong },
    .{ .char = '_', .mark = .emph },
    .{ .char = '#', .mark = .mark },
};
const MONOSPACE_CHAR: u8 = '`';

fn markForChar(c: u8) ?AST.InlineMark {
    for (MARK_SPANS) |m| {
        if (m.char == c) return m.mark;
    }
    return null;
}

/// Every byte that can open a span in either form. `parseInlines` dispatches on
/// this once and lets `emitSpan` sort out which node the delimiter builds, so
/// the constrained and unconstrained scans never duplicate that decision.
fn isSpanDelimiter(c: u8) bool {
    return markForChar(c) != null or c == MONOSPACE_CHAR;
}

/// May `text[i]`, a candidate delimiter byte, OPEN a constrained span? Not
/// preceded by a word character, not followed by whitespace or end of text —
/// AsciiDoc's constrained-span word-boundary rule, checked independently of
/// the matching close.
fn opensConstrained(text: []const u8, i: usize) bool {
    if (i > 0 and isWordByte(text[i - 1])) return false;
    if (i + 1 >= text.len) return false;
    const next = text[i + 1];
    return !(next == ' ' or next == '\t' or next == '\n');
}

/// The index of the delimiter byte `delim` that CLOSES a span opened at
/// `open_idx`, if any: not preceded by whitespace, not followed by a word
/// character — the closing half of the same rule.
///
/// Starts at `open_idx + 2`, so the interior is at least one byte. Asciidoctor
/// spells this interior `(\S|\S.*?\S)`, which likewise cannot match empty, and
/// the guard is what keeps `****` from encoding as two EMPTY strong spans —
/// four source bytes rendering as `<strong></strong><strong></strong>`, which
/// deletes them from the output rather than passing them through.
fn findConstrainedClose(text: []const u8, delim: u8, open_idx: usize) ?usize {
    var j = open_idx + 2;
    while (j < text.len) : (j += 1) {
        if (text[j] != delim) continue;
        const before_space = text[j - 1] == ' ' or text[j - 1] == '\t' or text[j - 1] == '\n';
        const after_word = j + 1 < text.len and isWordByte(text[j + 1]);
        if (!before_space and !after_word) return j;
    }
    return null;
}

/// The index of the FIRST byte of the doubled delimiter that closes an
/// UNCONSTRAINED span opened at `open_idx`, if any.
///
/// Unconstrained spans carry no word-boundary rule at all — that is their whole
/// point, since `sub**string**here` is how AsciiDoc spells intraword formatting
/// — so this is a plain scan for the next `delim delim` pair, non-greedy the way
/// asciidoctor's own `\*\*(.+?)\*\*` is. The search starts at `open_idx + 3`
/// because that pattern is `.+?` and not `.*?`: the interior must be at least
/// one byte, so `****` is literal text rather than an empty span.
fn findUnconstrainedClose(text: []const u8, delim: u8, open_idx: usize) ?usize {
    var j = open_idx + 3;
    while (j + 1 < text.len) : (j += 1) {
        if (text[j] == delim and text[j + 1] == delim) return j;
    }
    return null;
}

/// Build the one node a span delimiter produces, over the already-located
/// `interior`. The constrained and unconstrained scans differ ONLY in how they
/// find their delimiters — never in what they build — so both land here.
///
/// Nothing records WHICH form was used, deliberately: the ASG needs that fact
/// on the way out, but a span's own source spells it (an unconstrained span
/// opens with a doubled delimiter), so `asg.zig`'s `spanForm` reads it back off
/// the span rather than storing it. `AST.Attrs` is the channel a document's
/// REAL attributes travel in, and anything parked there renders straight into
/// `languages/html/serializer.zig`'s output as a bogus HTML attribute.
fn emitSpan(
    b: *Builder,
    delim: u8,
    interior: []const u8,
    interior_base: usize,
    outer: Span,
) Allocator.Error!Node.Id {
    const id = if (markForChar(delim)) |mark| blk: {
        const inner = try parseInlines(b, interior, interior_base);
        defer b.allocator.free(inner);
        break :blk try b.addContainer(.{ .inline_mark = mark }, inner);
    } else blk: {
        // Monospace keeps its interior verbatim instead of recursing — see
        // `MARK_SPANS`'s doc comment for why it is a leaf and not a mark.
        const leaf = try b.addLeaf(.{ .text_leaf = .{ .kind = .verbatim, .text = interior } });
        b.setContentSpan(leaf, Span.init(interior_base, interior_base + interior.len));
        break :blk leaf;
    };
    b.setSpan(id, outer);
    return id;
}

/// Scan `text` (a slice of `source` starting at byte offset `base`) for spans,
/// returning the resulting run of `str`, `inline_mark` and monospace
/// `text_leaf` nodes. See `MARK_SPANS`'s doc comment for which delimiters are
/// recognized.
///
/// ── Unconstrained is tried FIRST, and that ordering is load-bearing ─────────
/// Asciidoctor applies its unconstrained substitution patterns ahead of the
/// constrained ones, and the constrained scan is actively wrong on a doubled
/// delimiter rather than merely blind to it: it opens on the first byte of the
/// pair and closes on the NEAR half of the closing pair, leaving the far half
/// stranded as text (`**bold**` came out as `<strong>*bold</strong>*`, and
/// `__ital__` as a doubly-nested `<em>` because `isWordByte` counts `_` and
/// pushed the close one byte further right).
///
/// A doubled opener with no doubled close falls THROUGH to the constrained
/// scan rather than going literal, which is also what asciidoctor does:
/// `**bold*` is a constrained span whose interior happens to start with `*`.
fn parseInlines(b: *Builder, text: []const u8, base: usize) Allocator.Error![]Node.Id {
    var ids: std.ArrayList(Node.Id) = .empty;
    errdefer ids.deinit(b.allocator);
    var plain_start: usize = 0;
    var i: usize = 0;
    while (i < text.len) {
        if (isSpanDelimiter(text[i])) delim: {
            const doubled = i + 1 < text.len and text[i + 1] == text[i];
            if (doubled) {
                if (findUnconstrainedClose(text, text[i], i)) |close| {
                    if (i > plain_start) try ids.append(b.allocator, try emitText(b, text[plain_start..i], base + plain_start));
                    const span_id = try emitSpan(b, text[i], text[i + 2 .. close], base + i + 2, Span.init(base + i, base + close + 2));
                    try ids.append(b.allocator, span_id);
                    i = close + 2;
                    plain_start = i;
                    continue;
                }
            }
            if (!opensConstrained(text, i)) break :delim;
            const close = findConstrainedClose(text, text[i], i) orelse break :delim;

            if (i > plain_start) try ids.append(b.allocator, try emitText(b, text[plain_start..i], base + plain_start));
            const span_id = try emitSpan(b, text[i], text[i + 1 .. close], base + i + 1, Span.init(base + i, base + close + 1));
            try ids.append(b.allocator, span_id);
            i = close + 1;
            plain_start = i;
            continue;
        }
        i += 1;
    }
    if (plain_start < text.len) try ids.append(b.allocator, try emitText(b, text[plain_start..], base + plain_start));
    return ids.toOwnedSlice(b.allocator);
}

fn computeLines(allocator: Allocator, source: []const u8) Allocator.Error![]LineInfo {
    var list: std.ArrayList(LineInfo) = .empty;
    errdefer list.deinit(allocator);
    var start: usize = 0;
    var i: usize = 0;
    while (i < source.len) : (i += 1) {
        if (source[i] == '\n') {
            try list.append(allocator, .{ .start = start, .end = i });
            start = i + 1;
        }
    }
    // Only a source that does NOT end in a newline has a final partial line.
    // Appending one unconditionally would give every ordinary file a phantom
    // empty last line, which a block running to end-of-source (an unclosed
    // delimited block) would then swallow into its own extent.
    if (start < source.len) try list.append(allocator, .{ .start = start, .end = source.len });
    return list.toOwnedSlice(allocator);
}

const testing = std.testing;

test "a single paragraph" {
    var doc = try parse(testing.allocator, "A paragraph that consists of a single line.\n");
    defer doc.deinit();
    const ast = doc.ast;
    const para = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[para].kind == .para);
    const str = ast.nodes[para].first_child.?;
    try testing.expectEqualStrings("A paragraph that consists of a single line.", ast.nodes[str].kind.str);
}

test "a document title becomes the top-level heading, attribute entries attach to the document" {
    var doc = try parse(testing.allocator, "= Document Title\n:icons: font\n:toc:\n");
    defer doc.deinit();
    const ast = doc.ast;
    const attrs_marker = ast.nodes[ast.root].first_child.?;
    try testing.expectEqualStrings("document-attributes", ast.nodes[attrs_marker].kind.container.name);
    try testing.expectEqualStrings("font", ast.attrsOf(attrs_marker).get("icons").?);
    try testing.expectEqualStrings("", ast.attrsOf(attrs_marker).get("toc").?);
    const heading = ast.nodes[attrs_marker].next_sibling.?;
    // 1, not the ASG's 0: twig's shared `heading.level` is 1-based in every
    // language, so `= Title` and Markdown's `# Title` agree. See the parser's
    // own `addContainer` call for why that matters to rendered HTML.
    try testing.expectEqual(@as(u32, 1), ast.nodes[heading].kind.heading.level);
}

test "a section nests a paragraph inside it" {
    var doc = try parse(testing.allocator, "== Section Title\n\nparagraph\n");
    defer doc.deinit();
    const ast = doc.ast;
    const section = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[section].kind == .section);
    const heading = ast.nodes[section].first_child.?;
    // `==` is the ASG's level 1 and the shared tree's 2 — an `<h2>`, exactly
    // where the same section sits in Markdown's `##`.
    try testing.expectEqual(@as(u32, 2), ast.nodes[heading].kind.heading.level);
    const body = ast.nodes[heading].next_sibling.?;
    try testing.expect(ast.nodes[body].kind == .para);
}

test "a deeper heading nests inside a shallower one; a same-level heading closes it" {
    var doc = try parse(testing.allocator, "== One\n\n=== Two\n\npar\n\n== Three\n");
    defer doc.deinit();
    const ast = doc.ast;
    const sec_one = ast.nodes[ast.root].first_child.?;
    const sec_three = ast.nodes[sec_one].next_sibling.?;
    try testing.expectEqual(@as(?Node.Id, null), ast.nodes[sec_three].next_sibling);
    const heading_one = ast.nodes[sec_one].first_child.?;
    const sec_two = ast.nodes[heading_one].next_sibling.?;
    try testing.expect(ast.nodes[sec_two].kind == .section);
    try testing.expectEqual(@as(?Node.Id, null), ast.nodes[sec_two].next_sibling);
}

test "an unordered list item carries its bullet spelling" {
    var doc = try parse(testing.allocator, "* water\n");
    defer doc.deinit();
    const ast = doc.ast;
    const list = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[list].kind == .bullet_list);
    const item = ast.nodes[list].first_child.?;
    try testing.expectEqual(Document.Spelling.Bullet.star, doc.spelling(item).?.bullet);
    const str = ast.nodes[item].first_child.?;
    try testing.expectEqualStrings("water", ast.nodes[str].kind.str);
}

test "a delimited listing block keeps its interior verbatim" {
    var doc = try parse(testing.allocator, "----\ndef main\n  puts 'hello'\nend\n----\n");
    defer doc.deinit();
    const ast = doc.ast;
    const listing = ast.nodes[ast.root].first_child.?;
    try testing.expectEqualStrings("def main\n  puts 'hello'\nend", ast.nodes[listing].kind.code_block.text);
}

test "a sidebar is a generic container holding its blocks" {
    var doc = try parse(testing.allocator, "****\n* phone\n* wallet\n* keys\n****\n");
    defer doc.deinit();
    const ast = doc.ast;
    const sidebar = ast.nodes[ast.root].first_child.?;
    try testing.expectEqualStrings("sidebar", ast.nodes[sidebar].kind.container.name);
    const list = ast.nodes[sidebar].first_child.?;
    try testing.expect(ast.nodes[list].kind == .bullet_list);
}

test "degenerate inputs parse and encode without crashing" {
    // The corpus can't hold these: an empty document has no location the ASG
    // can spell (both endpoints are inclusive), so there is no expectation to
    // author — only the requirement that nothing crashes on the way through.
    const asg = @import("asg.zig");
    for ([_][]const u8{ "", "\n", "   \n", "\n\n\n", "= \n", "----\n", "****\n", "*\n" }) |source| {
        var doc = try parse(testing.allocator, source);
        defer doc.deinit();
        const json = try asg.encodeAlloc(testing.allocator, &doc, .document);
        testing.allocator.free(json);
    }
}

test "a constrained strong span" {
    var doc = try parseInlineList(testing.allocator, "*s*\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .strong);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("s", ast.nodes[str].kind.str);
}

test "a constrained emphasis span" {
    var doc = try parseInlineList(testing.allocator, "_s_\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .emph);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("s", ast.nodes[str].kind.str);
}

test "a constrained mark span" {
    var doc = try parseInlineList(testing.allocator, "#s#\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .mark);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("s", ast.nodes[str].kind.str);
}

test "a constrained monospace span is a verbatim text leaf, not a nested mark" {
    var doc = try parseInlineList(testing.allocator, "`s`\n");
    defer doc.deinit();
    const ast = doc.ast;
    const leaf = ast.nodes[ast.root].first_child.?;
    try testing.expectEqual(AST.TextLeafKind.verbatim, ast.nodes[leaf].kind.text_leaf.kind);
    try testing.expectEqualStrings("s", ast.nodes[leaf].kind.text_leaf.text);
}

test "an unconstrained span consumes BOTH delimiter bytes at each end" {
    // The regression this whole path exists for: the constrained scan used to
    // open on the first `*` of the pair and close on the near half of the
    // closing pair, yielding a `*bold` interior plus a stray trailing `*`.
    var doc = try parseInlineList(testing.allocator, "**bold**\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .strong);
    try testing.expect(ast.nodes[span].next_sibling == null);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("bold", ast.nodes[str].kind.str);
}

test "an unconstrained emphasis span does not nest inside itself" {
    // `isWordByte` counts `_`, which pushed the constrained close one byte
    // right and re-entered the scan on a `_ital_` interior — a doubled `<em>`.
    var doc = try parseInlineList(testing.allocator, "__ital__\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .emph);
    const inner = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("ital", ast.nodes[inner].kind.str);
}

test "an unconstrained span opens and closes mid-word" {
    // The point of the form: no word-boundary rule, so `sub**string**here`
    // formats where the constrained spelling refuses to.
    var doc = try parseInlineList(testing.allocator, "sub**string**here\n");
    defer doc.deinit();
    const ast = doc.ast;
    const first = ast.nodes[ast.root].first_child.?;
    try testing.expectEqualStrings("sub", ast.nodes[first].kind.str);
    const span = ast.nodes[first].next_sibling.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .strong);
    const tail = ast.nodes[span].next_sibling.?;
    try testing.expectEqualStrings("here", ast.nodes[tail].kind.str);
}

test "an unconstrained monospace span stays a verbatim leaf" {
    var doc = try parseInlineList(testing.allocator, "``mono``\n");
    defer doc.deinit();
    const ast = doc.ast;
    const leaf = ast.nodes[ast.root].first_child.?;
    try testing.expectEqual(AST.TextLeafKind.verbatim, ast.nodes[leaf].kind.text_leaf.kind);
    try testing.expectEqualStrings("mono", ast.nodes[leaf].kind.text_leaf.text);
}

test "a doubled opener with no doubled close falls through to the constrained scan" {
    // Asciidoctor's own ordering: the unconstrained pattern fails, and the
    // constrained one then matches with `*bold` as its interior.
    var doc = try parseInlineList(testing.allocator, "**bold*\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .strong);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("*bold", ast.nodes[str].kind.str);
}

test "a constrained span's interior is never empty" {
    // Without the guard, `****` encoded as two EMPTY strong spans — four source
    // bytes vanishing from the output rather than passing through.
    var doc = try parseInlineList(testing.allocator, "****\n");
    defer doc.deinit();
    const ast = doc.ast;
    const span = ast.nodes[ast.root].first_child.?;
    try testing.expect(ast.nodes[span].kind.inline_mark == .strong);
    const str = ast.nodes[span].first_child.?;
    try testing.expectEqualStrings("*", ast.nodes[str].kind.str);
}

test "delimited blocks carry no ASG bookkeeping in attrs" {
    // `attrs` renders — see `emitSpan`. Anything the ASG needs but the document
    // doesn't have must be derived in `asg.zig`, not parked here.
    var doc = try parse(testing.allocator, "----\ncode\n----\n");
    defer doc.deinit();
    for (0..doc.ast.nodes.len) |id| {
        try testing.expectEqual(@as(usize, 0), doc.ast.attrsOf(@intCast(id)).entries.len);
    }
}