[dankoverride]

NotTaken is documented as the difference between the diagram and the run. It was built
entirely from RunState.DeclinedEdges, and declining is something the walk does — that list
only ever holds edges the walk reached and turned down. An edge leaving a pod no path arrives at
is never offered, never refused, and so never appeared at all. The largest difference between a
diagram and its run was the one difference the field could not report.

What it looks like

A model drew a 15-pod diagram from a screenshot of a reference architecture and left the
pub/sub topic with no publisher — every service-bus --> x, no x --> service-bus. It parses
clean and runs clean. One run, before and after:

	before	after
edges drawn	20	20
crossed	12	12
reported in notTaken	1	8
accounted for nowhere	7	0

A renderer greying out what notTaken names drew those 7 exactly like the edges the request
crossed — the class of lie Hop and NotTaken both exist to prevent.

What changed

NotTakenEdges now takes the graph and reports every drawn edge that is neither crossed nor
already reported as declined. The reason names the source pod rather than the edge, because
that is the thing to fix: one missing arrow into a topic strands the topic and everything only
it calls, and five of the eight here read no path from the entry point arrives at service-bus.

An unreachable edge and a refused one now read differently, which matters — only one of them is
a mistake in the picture.

A diagram with no entry point reports every edge instead of none. Nothing ran, so the difference
is the whole diagram; an empty list said the opposite.

Testing

dotnet test src/Shoebox.sln — 1201 pass, 6 new in tests/Shoebox.Api.UnitTests/Run/UnreachableEdgeTests.cs:

• every drawn arrow is crossed or reported (the invariant the field is for)
• the count and reason for a stranded source pod
• a reported edge is never also a crossed one
• an unreachable edge reads differently from a refused one
• the no-entry case
• a fully walked diagram still reports nothing — the guard on the other side, so widening this
  does not start greying out arrows the request did cross

Also ran it against a local instance and diffed the /run body for the diagram above; that is
where the table comes from.

Not in this PR

/topology/parse has the same blind spot — its notes cover cycles and say nothing about a
topic nothing publishes to, so a caller cannot see the problem until after it fires. Happy to
follow up separately; keeping this one to the run.

🤖 Generated with Claude Code