[ChristianMurphy]

Summary

With GFM, an email address in an image title that spans lines panics in 1.0.0:

markdown::to_html_with_options("![a](b \"c@d.com\ne\")", &markdown::Options::gfm()); // Option::unwrap() on None, src/subtokenize.rs:218

CommonMark renders <p><img src="b" alt="a" title="c@d.com\ne" /></p>, and GFM doesn't autolink in titles.

The GFM autolink literal resolver scans Data outside links for email addresses and rewrites it into autolinks. It skipped Data inside a Link, but not inside an Image. So it also rewrote image title chunks, which are string content and point to the next chunk through their link field. The rewrite dropped that field, and divide_events panicked. This skips Data that has a link, the same check #207 adds to resolve_whitespace.

Output changes

Single-line image titles change too. 1.0.0 rewrote them, which could escape a character reference twice: ![a](b "c@d.com & e") gave the title c@d.com & e. It is now c@d.com & e, as micromark-extension-gfm gives it.

Performance

Callgrind, with one codegen unit, against 1.0.0 built the same way, across 7 option sets. On small documents, instructions change by −0.6% to +0.2%, within the noise of this method.

Tests

• Both inputs above assert the exact HTML.
• 50,000 seeded random titles and labels in links, images, definitions, block quotes, and list items. With test: more link-shape variants of subtokenize.rs:149 panic (GH-79) #207 alone, 1,007 of them panic, all with @ in a title. With both, none do.

As of 2026-09-29, CI on main fails for two reasons unrelated to this change. One is a new clippy lint on Rust 1.98, and the other is the swc/serde build break. #230 fixes both.

Claude, an AI assistant, helped me build this, and I reviewed it.

[Murderlon]

I authorized Codex to review this PR on my behalf.

No findings.