Checks 9 and 10 (spec 3.2.1), which close the boilerplate gap that relaxed parsing left open. Lowering accepts a boilerplate-free <scxml> fragment unconditionally (Statifier.Lowering.Namespace.scxml_vocabulary?/1 treats a nil namespace as SCXML's own vocabulary), so this is the only place in the pipeline that ever refuses one.

Check 9 reads document.namespace - the root element's resolved URI (Statifier.Document's moduledoc), not the literal xmlns attribute - so a spec-conformant prefix-declared root such as <s:scxml xmlns:s="http://www.w3.org/2005/07/scxml"> passes clean even though its own xmlns field is nil. Check 10 reads document.version directly; v1's regex-based presence check is deliberately not repeated here, so a written-but-wrong version="2.0" is reported rather than silently accepted.

Both checks share the same span preference: the written attribute's own value span when the attribute was written, document.location (the <scxml> start tag) otherwise - covering both the absent case and the prefixed-xmlns case, where the declaration is xmlns:s= and has no attribute_locations[:xmlns] entry.

Check 9 has one relaxed mode (ADR-0042): when context.invoke_content_markup? is set, a document.namespace of nil also passes - Statifier.Lowering.Namespace.scxml_vocabulary?/1's own predicate, applied here instead of at lowering dispatch. A non-nil namespace that is not the SCXML URI still fails in both modes; the flag only ever widens the nil case, never the wrong-namespace case.