Failure classification by name string misclassifies harness failures on name collision

Summary

Failure classification in EvalonRunner.scala:87 decides "agent under test failed" vs "harness failed" by comparing the failed participant's name string to evaluatedName(scenario):

if participant == EvalonRunner.evaluatedName(scenario) then
  throw AgentStepFailedException(cause, t)   // exit code 1: agent failed
else throw SimulationFailedException(participant, cause, t)  // exit code 2: harness failed, retryable

If a harness participant shares the evaluated agent's name, its failure is misclassified.

Reachability

• Participants can't collide with each other (they're Map keys), so the only exposure is an EventSourceConfig.name equal to the evaluated participant's name.
• There is no scenario validation forbidding this (same gap as #9).
• A failing event source sends ParticipantFailed(config.name, cause) (EventSourceActor.scala:115). If that name equals evaluatedName, EvalonRunner:87 throws AgentStepFailedException (exit code 1, "agent failed") instead of SimulationFailedException (exit code 2, "retryable harness failure").

Impact

Low, config-dependent, pre-existing. But the wrong exit code means automation/CI could blame the agent under test for an infrastructure outage, or skip a retry that a harness failure warrants.

Suggested fix

Classify by identity/role rather than by name string, e.g. carry a participant kind (evaluated vs simulated vs event-source) in ParticipantFailed so classification doesn't depend on names being distinct. Alternatively, validate at scenario load that event-source names and participant names are disjoint.

Found during review of #8. Related to #9 (validation gap) and the evaluated-participant-lookup dedup issue.

Start at EvalonRunner.scala:87 and trace ParticipantFailed from EventSourceActor.scala:115. Inspect how participant identity or role is represented and how scenario loading handles EventSourceConfig.name collisions. Done means a colliding event-source failure is classified as a retryable harness failure with exit code 2 rather than an agent failure with exit code 1.