Conformance¶

Parsing and conformance answer different questions. The WHATWG tree-construction algorithm is defined to never fail: it recovers from any byte stream into a tree, recording each recovery as a ParseError. Conformance is the other half – the authoring requirements a document must meet even though it parsed cleanly. <img src=x> builds a perfectly good img element; it is still non-conforming, because the standard requires an alt attribute. check() is that second surface: the rules the parser deliberately does not enforce.

A severity model, not a bool¶

The reference validator for this space is the Nu Html Checker (validator.nu), and it does not answer yes or no. It classifies each finding – an error for a violated requirement, a warning for an authoring recommendation, an info note for advice – and the document is invalid exactly when at least one finding is an error. check() adopts that shape. A ConformanceReport is a valid verdict plus every ConformanceMessage, and each message carries a stable code ("img-missing-alt") so a tool can match on it, a Severity, a human-readable message, and a source line and column. The verdict reads only the errors: turning a warning into a hard failure is the caller’s policy, made by inspecting warnings, not the checker’s.

What it checks¶

The rules cover the authoring mistakes that survive a parse, grouped the way the standard and WAI-ARIA 1.2 define them:

Text alternatives. An img, an area with an href, or an input of type image without an alt attribute (error).
Obsolete markup. A non-conforming element (center, font, marquee, …) or a presentational attribute (align, bgcolor, link on body, frame on table, …) the standard’s obsolete-features table retired (error).
Identity. A duplicate id (error).
ARIA. A role whose value is not a defined non-abstract ARIA role (error), or one that merely restates the element’s implicit role, such as <nav role=navigation> (warning).
Structure. An empty heading, or a section or article with no heading and no accessible name (warning).
Document. A whole document with no non-empty title (error) or no lang on html (warning). These apply only to a Document; a subtree runs the per-element rules alone.

Why the C core¶

The check is a single pre-order pass over the tree the parser already built, and it runs entirely in the extension. It is iterative, not recursive, so no depth of nesting can exhaust the C stack – the checker needs no depth cap. Every name test (an obsolete element, an ARIA role token, a presentational attribute) resolves against a static table by a loop rather than a chain of comparisons, so a tag spelling the parser never interns is still caught and the cost stays a handful of integer compares per node. The Python layer only shapes the C findings into the report records and derives the verdict.