Deduplicate evaluated-participant lookup into a single source of truth

Summary

The "find the evaluated participant" lookup is implemented independently in two places:

• EvalonRunner.evaluatedName (EvalonRunner.scala:111-115)
• inline in ScenarioRunner.apply (ScenarioRunner.scala:88-91)

Both are byte-identical today (scenario.participants.find(_._2.participantType == ParticipantType.Evaluated).map(_._1).getOrElse("agent")), so they must be kept in sync by hand.

Why it matters

This is more than a DRY nit because the two copies feed different things:

• ScenarioRunner's copy names the actor that is spawned as the evaluated agent.
• EvalonRunner's copy is what failure classification compares against (EvalonRunner.scala:87) to decide AgentStepFailedException vs SimulationFailedException.

If one is edited (e.g. changing the .getOrElse("agent") fallback or adding a tie-break) and the other is missed, the actor spawned as "the evaluated agent" and the name used for classification diverge, silently misclassifying every agent failure as a harness failure.

Suggested fix

Extract one shared helper and have both call sites use it, so the evaluated-participant identity has a single source of truth.

Priority: low (no current bug; latent divergence risk). Found during review of #8. See also #9 (related classification correctness).

Start with EvalonRunner.scala:87 and 111-115, then compare the duplicate lookup in ScenarioRunner.scala:88-91. Extract the shared evaluated-participant lookup and update both call sites to use it. Done means both spawning and failure classification use the same source of truth, with existing tests still passing.