guppy 0.17.26

Fixed

A node with a self-loop edge -- typically a package with a path
dev-dependency on its own crate, or a feature node like base/[base]
arising from one -- was treated inconsistently by several APIs. The root
cause was that self-edges were counted as incoming edges in the SCC and
topological-sort machinery, which transitively confused everything built
on top:

• Sccs::externals / forward roots. Previously, a package whose only
  incoming edge was its own self-loop was excluded from the set of forward
  roots, so callers iterating over
  query_workspace().resolve().root_ids(_) did not see it, and forward
  link enumeration on it was broken. Now, self-loop edges no longer
  disqualify a node from being a forward root: a single-node SCC is
  external iff it has no incoming edges from outside its own SCC.
  (#586)
• Topological sort. Previously, TopoWithCycles::new filtered
  self-looping nodes out of its root set for the same reason, dropping
  them from the DFS and then placing them and their descendants at the
  end of the topological order via a best-effort fallback. Now, the
  root-set predicate accepts a node whose only incoming edge is its own
  self-loop, so such nodes are visited in DFS order alongside other
  roots. (#589)
• Cycles::is_cyclic and feature::Cycles::is_cyclic. Previously,
  these were reflexively true for every package or feature, regardless
  of actual cycle membership. Now, they return true only when the
  argument lies on a directed cycle: either in a multi-node SCC, or in a
  single-node SCC with a self-loop edge. This is a behavior change, but
  in the context of the rest of this release is being treated as a
  bugfix. (#590)
• Cycles::all_cycles and feature::Cycles::all_cycles. Previously,
  these reported only SCCs of two or more elements. Now, they also yield
  single-node SCCs whose node has a self-loop edge, in topological order
  alongside multi-node SCCs. This makes all_cycles() and
  is_cyclic(x, x) agree on what counts as a cycle. (#590)
• FeatureGraphWarning::SelfLoop. Previously, this warning was
  emitted for every self-loop edge in the feature graph, including
  legitimate ones such as a path dev-dependency on the package's own
  crate. Now, it is restricted to named-feature self-loops like
  [features] a = ["a"], which really are user errors. (#592)

Documentation

• PackageGraph::directly_depends_on and
  FeatureGraph::directly_depends_on. Previously, both documented
  themselves as returning false when the two IDs were equal. The
  implementation has always called dep_graph.contains_edge(a, b),
  which correctly returns true for self-loop edges, so the
  documentation contradicted the behavior. The documentation has now
  been corrected to match.
• FeatureGraph::directly_depends_on. Previously, the docstring
  read "returns true if feature_a is a direct dependency of
  feature_b," with the operands inverted relative to the actual
  implementation. Now the docstring reflects the implementation:
  returns true if feature_b is a direct dependency of feature_a.
  (#591)
• [Cycles] type-level documentation. Previously, the long-form
  documentation talked exclusively about multi-node cycles (like
  serde and serde_derive). It now also acknowledges single-node
  self-loop cycles as a case.