Add a 6th Trace Theater scenario — prompt injection variant (benign)

[caiovicentino]

We have a DAN-style refusal scenario in Trace Theater. Add a second variant in the safety category: a benign jailbreak pattern like 'role-play grandma reading chemistry bedtime stories' (public, documented).

What to do

1. Fork lib/trace-scenarios.ts
2. Append a new TraceScenario object (use the existing DAN scenario as template)
3. Pick 8-10 new unique feature IDs (grep the file to avoid collision)
4. Write plausible activations + at least 1 counterfactual
5. Open PR

Why

Gives users a comparison point across two distinct injection styles. Validates the feature-level refusal signal generalizes.

Mentor

@caiovicentino — review within 48h