A <transition> element: spec 3.5, matched against events and guarded by a condition to move the configuration from one set of states to another.

event and target are both already-split lists ([String.t()]), not the raw whitespace-separated attribute strings. event holds the transition's event descriptors - each one still an unparsed token that may itself use dot-separated wildcards ("error.*") - and target holds the tokenized target ids. The re-splitting this avoids is what ../statifier/lib/statifier/event.ex:64-79 does on every match attempt in v1: tokenizing once at lowering means no consumer downstream re-splits. The second-level split of a descriptor into its dot-separated tokens is not done here - that is the compiler's job, once the descriptor is being matched against a compiled event name, not stored. An empty event list means an eventless transition; an empty target list means a targetless one - both spec-legal, and both distinguishable from "one descriptor" or "one target" without a sentinel.

cond is raw predicator source, String.t() | nil, never compiled here (docs/datamodel.md). Its ADR-0014 span lives at attribute_locations[:cond], the base offset ADR-0014's arithmetic adds a within-expression span to - not a field of this struct, matching how every other expression-bearing node stores its span.

type defaults to :external, the value lowering applies when the attribute is absent (spec 3.5's own default). That default is applied, not left for every consumer to re-derive, but it is not by itself proof the author wrote type="external": attribute_locations carries no :type key when the default applied, and does when the author wrote either value explicitly. The same rule holds for every other field with a non-nil default in this layer.

content is the transition's executable content directly, [Document.content_node()], not wrapped in a Statifier.Document.Block: there is no wrapping element in the source to give a block its own location, and the transition's own location already is that span.