One <onentry> or <onexit> element: an ordered list of executable content plus the location of the element that contains it.

A bare [Document.content_node()] on Statifier.Document.State would preserve the block boundary the spec requires - 3.8 and 3.9 each treat <onentry>/<onexit> as one unit, and section 4's error rule makes each block a separate error-isolation unit, so State.onentry and State.onexit are each [Block.t()], not a flattened content list - but it would give the block itself no location, so a diagnostic could name the failing <raise> inside it and not the <onentry> that contains it. Locations are retained on executable-content nodes wherever the source has one to give (docs/observability.md Constraint 3); the <onentry> or <onexit> element is such a node, and one small struct is the whole cost of not losing it.

There is no kind field here the way Statifier.Parser.Markup has one for :start/:end. The field name on State (onentry vs onexit) already says which this is, and unlike Markup's records the two never share a list to be told apart within.

A transition's own executable content is [Document.content_node()] directly, not wrapped in a Block.t() - there is no <onentry>-like element in the source to give a block its own location, and the transition's own location already is that span.