[PHJ2000]

Rendering a Link or Image whose destination/title contains literal entity-like text (for example © or A) writes an unescaped ampersand. Parsing that Markdown again turns the text into a character, changing the AST value.

For example:

import org.commonmark.node.*;
import org.commonmark.parser.Parser;
import org.commonmark.renderer.markdown.MarkdownRenderer;

public class CommonmarkAmpersandProbe {
    public static void main(String[] args) {
        Link link = new Link("/url?value=©", "literal A");
        link.appendChild(new Text("example"));
        String markdown = MarkdownRenderer.builder().build().render(link);
        Link reparsed = (Link) Parser.builder().build().parse(markdown).getFirstChild().getFirstChild();
        if (!link.getDestination().equals(reparsed.getDestination())
                || !link.getTitle().equals(reparsed.getTitle())) {
            throw new AssertionError("Changed to: " + reparsed.getDestination() + " / " + reparsed.getTitle());
        }
        System.out.println(markdown);
        System.out.println("Literal destination and title preserved");
    }
}

Use the existing MarkdownWriter escaping for ampersands in bare/angle-bracket destinations and quoted titles. The regression tests cover links and images, named/decimal/hex entities, angle-bracket destinations, and ordinary ampersands. Ordinary query separators are also escaped in the Markdown output, while the parsed values stay the same; the renderer does not promise identical Markdown syntax.

Validation:

• Baseline: 14 of the 18 new parameter cases fail; the four ordinary-ampersand controls pass.
• JDK 21: focused renderer and spec tests pass; ./mvnw clean verify javadoc:javadoc spotless:check succeeds (4,603 test entries, 1 skipped, no failures/errors). Javadoc emits nonfatal missing sibling target/apidocs link messages; generation completes.
• Actual JDK 11: a standalone destination/title round-trip example compiles and runs. The separate Android project was not run.
• A separate AI review reproduced the 14 baseline failures and passing focused tests. It did not repeat the full reactor or JDK 11 probe.

AI disclosure: OpenAI Codex assisted with investigation, implementation, tests, and this description. The separate review was also performed by an AI session, not a human reviewer.