Real BEAM-native benchmark harness comparing distributed-engine topologies for process discovery (directly-follows graph, DFG), under a virtual-cost model (Ex4pmEngine.Bench.VirtualCost: compute/store/send operation counts) rather than wall-clock time.

A toy event log is a list of {case_id, activity, timestamp} tuples. Discovery computes the DFG: for each case, sort its events by timestamp, then count {activity_a, activity_b} directly-follows edge frequencies across all cases. The DFG is %{ {a, b} => frequency }.

Two real topologies are implemented:

• :centralized — a single process holds the whole log, sorts each case, and builds the DFG directly. No inter-process communication.
• :edge — the log is partitioned across N real BEAM processes (one Task per simulated node), each computes a local partial DFG over its own shard of cases, then sends its partial DFG back to the coordinator, which merges them by summing edge frequencies. This exercises real message passing (Task.async/Task.await, which is a real send/receive round-trip), not just a sequential map/reduce in one process.

Both topologies are asserted (by the test suite) to produce the same final merged DFG for the same input log — the correctness invariant this harness exists to check before its virtual-cost numbers are trusted.