HTML image alt text is reparsed as directive option syntax #1213

What version of myst-parser are you using?

5.1.0, reproduced from current master at 723cffcf84213f0cb58695b27eec9ad72052b53a.

What version dependencies are you using?

Python 3.11.16 on macOS. The regression was exercised with Sphinx 9.0.4 / docutils 0.22.4 and Sphinx 8.2.3 / docutils 0.21.2; markdown-it-py 4.2.0 and mdit-py-plugins 0.6.1.

Describe the bug

With html_image enabled, decoded HTML image attributes are serialized into directive option text and parsed again. Valid literal alt text can therefore be truncated or dropped and produce warnings:

• A # sign becomes A, with a directive-comments warning.
• "quoted" text loses its alt attribute, with an invalid-options-format warning.
• First followed by a newline and second becomes First, with a forbidden-content warning.

This loses accessibility text and can fail warning-strict builds.

To reproduce

from io import StringIO

from docutils import nodes
from docutils.core import publish_doctree

from myst_parser.parsers.docutils_ import Parser

for alt in ("ordinary text", "A # sign", ""quoted" text", "First\nsecond"):
    warnings = StringIO()
    document = publish_doctree(
        f'<img src="example.png" alt="{alt}">\n',
        parser=Parser(),
        settings_overrides={
            "myst_enable_extensions": ["html_image"],
            "warning_stream": warnings,
        },
    )
    image = next(document.findall(nodes.image))
    print(repr(alt), "->", repr(image.get("alt")), repr(warnings.getvalue()))

Expected behavior

Preserve the decoded alt value exactly, without interpreting it as directive option syntax. The same problem occurs for inline HTML images.

This differs from #1210 and #1211, which concern Markdown image text_special tokens in renderInlineAsText. This reproduction starts with an HTML <img> tag and fails during HTML-to-directive conversion.

This report and the accompanying patch were prepared with AI assistance.