v0.6.0

Fixed

• GraphGenerator.CompleteBipartite now rejects a negative right-side count
  with ArgumentOutOfRangeException as documented, instead of silently
  producing a smaller edgeless graph when the two counts summed to a
  non-negative total.
• GraphDocument.EdgeData documentation no longer claims document order
  always matches the graph's edge enumeration order: multigraphs enumerate
  edges grouped by source vertex, so foreign documents with interleaved
  edge sources must be correlated by document position. The exporter
  documentation now states the actual (still deterministic) edge order.

Changed

• The assembly is strong-name signed from this release (the key ships in
  the repository — strong naming is identity, not security). .NET 8+
  ignores strong names for loading, but strongly-named consumers can now
  reference the package.

Added

• Parallel analysis overloads taking System.Threading.Tasks.ParallelOptions
  (degree of parallelism + cancellation, composing with the 0.5.0 token
  design) on the per-source algorithms: BetweennessCentrality (hop-count
  and weighted), ClosenessCentrality, Diameter, Radius, Center,
  Periphery, and AveragePathLength. Sequential paths are unchanged and
  remain the reference implementations; per-vertex results (closeness,
  eccentricity-based metrics) are bit-identical to sequential, accumulated
  ones (betweenness) differ only by floating-point merge order. Benchmarks
  compare both paths.

• Attribute round-trips for GraphML and node-link JSON: one
  GraphAttribute<T> declaration (typed factories: String/Bool/Int/Long/
  Float/Double) drives both exporters via VertexAttributes/
  EdgeAttributes on the existing option records; GraphMl.ParseDocument
  and GraphJson.ParseDocument return a GraphDocument carrying the graph
  plus typed per-vertex and per-edge attribute data. Existing Parse*
  entry points are unchanged; the JSON path stays reflection-free.

• Random-graph generators with realistic structure: BarabasiAlbert
  (preferential attachment; connected, simple, exactly m·(n−m) edges) and
  WattsStrogatz (ring lattice with in-place rewiring, so the edge count
  is always n·k/2). Seeded and deterministic like the existing generators.

• Clustering coefficients: LocalClusteringCoefficient,
  ClusteringCoefficients, AverageClusteringCoefficient, and
  GlobalClusteringCoefficient (transitivity). Direction is ignored,
  self-loops never count, and multigraph neighbors count once.

• Spectral centrality: EigenvectorCentrality (shifted power iteration,
  so bipartite graphs cannot oscillate) and KatzCentrality
  (attenuation alpha + base beta; the DAG-safe alternative), both
  L2-normalized with PageRank-style optional parameters and cancellation.

• DAG path algorithms via single-pass topological relaxation:
  DagShortestPathsFrom, DagLongestPathsFrom (both reuse
  SingleSourceShortestPaths), and CriticalPath (the maximum-weight path
  anywhere in the DAG). Negative weights are supported; cyclic input throws
  GraphCycleException.

• Graph set operations in Graph1x.Algorithms: Subgraph (induced by a
  vertex selection; unknown vertices ignored), Union (operands must agree
  on direction; result family and comparer come from the first), and
  Complement (simple graphs only; never emits self-loops). Results are new
  graphs matching the source's direction/parallel-edge policy, with
  directed-typed overloads preserving IDirectedGraph dispatch.