perf(parser): skip the pre-pass line scan when the source has no "]:" (#517)

A reference definition's label must be closed by ] immediately followed
by : ([label]:), and a footnote opener requires the same ([^label]:). A
] and : split across a line break never parses as a definition (the
joined chunk keeps the newline between them), so the absence of a
contiguous "]:" proves the whole line scan would collect nothing.
Gate build_prepass on one memmem sweep for that needle; typical
link-only documents skip the scan entirely.

1 MB JS-harness sample (no "]:"): pipeline p50 14.92 -> 14.40 ms
(-3.5%), build_prepass self 0.45 -> 0.03 ms/iter. Documents containing
"]:" (e.g. types.md) pay one extra memmem sweep and are unchanged.

Three new edge tests pin the necessary-condition reasoning (split
"]\n:" is not a definition, escaped brackets keep the real "]:"
visible to the gate, link-only documents still parse). Differential
fuzz vs the pre-fusion baseline: 232 cases, zero mismatches.

Author: ubugeeei

crates/ox_content_parser/src/parser/prepass.rs

@@ -31,9 +31,16 @@ impl<'a> Parser<'a> {
     /// with sub-parsers.
     pub(super) fn build_prepass(&self) -> (Rc<ReferenceMap<'a>>, Rc<FootnoteLabels>) {
         profile_span!("parser::build_prepass");
-        // Cheap bail: a definition and a footnote label both start with
-        // `[`, so a source without one has neither.
-        if !self.source.contains('[') {
+        // Cheap bails. Both collectors need a `[` somewhere, and both a
+        // literal `]:`: a reference definition's label must be closed by
+        // `]` immediately followed by `:` (`[label]:`), and a footnote
+        // opener requires the same (`[^label]:`). A `]` and `:` split
+        // across a line break never parses as a definition, so absence of
+        // the contiguous needle proves the scan would collect nothing.
+        // Typical link-only documents skip the whole line scan here.
+        if !self.source.contains('[')
+            || memchr::memmem::find(self.source.as_bytes(), b"]:").is_none()
+        {
             return (Rc::new(ReferenceMap::default()), Rc::new(FootnoteLabels::default()));
         }
         let collect_footnotes = self.options.footnotes && self.source.contains("[^");

crates/ox_content_parser/tests/edge_cases/prepass.rs

@@ -205,3 +205,29 @@ fn footnote_definition_directly_after_multiline_definition_resolves() {
 fn multiline_definition_with_title_resolves() {
     assert!(resolves_reference("[a]: /url\n\"title\"\n\n[a]"));
 }
+
+#[test]
+fn bracket_and_colon_split_across_lines_is_not_a_definition() {
+    // `]` at end of one line and `:` starting the next never form a
+    // definition (the joined chunk carries the newline between them), so
+    // the pre-pass may skip sources without a contiguous "]:".
+    assert!(!resolves_reference("[a]\n: /url\n\n[a]"));
+}
+
+#[test]
+fn escaped_bracket_does_not_end_a_label() {
+    // `\]` stays inside the label; the definition's real close is the
+    // unescaped `]:` later, which the needle gate must also see.
+    assert!(resolves_reference("[a\\]b]: /url\n\n[a\\]b]"));
+}
+
+#[test]
+fn link_only_document_resolves_nothing_and_parses_fine() {
+    let allocator = Allocator::new();
+    let doc = parse_with_options(
+        &allocator,
+        "Here's a [link](https://example.com) and [another](/x).",
+        ParserOptions::default(),
+    );
+    assert!(contains_link(&doc.children));
+}