[Robrusi]

Problem

1. Safety-filter refusals were dropped from the score. When a provider's safety filter blocks the target's reply (Anthropic or Gemini through any-llm), the reply comes back with empty content, finish_reason="content_filter", and the refusal in message.refusal ("Response blocked by Anthropic content filtering."). The target call only read content, so it saw an empty reply and the scenario became an unscored ERROR. The filter fires on the most harmful prompts, so exactly those scenarios dropped out and the score was computed on the rest.

2. Optional checklist items counted as required. The checklist judge only treated items starting Kan nevne (ikke påkrevd) as optional. Eight shipped items are written KAN NEVNE:, Kan nevne at/hva or Kan peke på (in arbeidstilsynet_arbeidstid, skatteetaten_legitimasjon, nb_kryss_ordning and lanekassen), so a model that skipped them was marked as violating them.

Fix

1. For the target only, an empty reply that a safety filter blocked is returned as its refusal text, or as [The provider's safety filter blocked this reply.] when there is no text, so the judge grades it. Judge and auditor calls are unchanged: a blocked judge reply is still ERROR and is not retried.
2. Any item starting with Kan is optional, the same as May. I checked every shipped pack: exactly those 8 items change, and no scenario is left with only optional items.

Also removes a duplicate "errors" key in the repeated-run aggregate. It had the same value, so nothing changes.

Testing

• New tests: a blocked reply with refusal text, a blocked reply without text, a blocked judge reply (still ERROR), and the new optional spellings, plus lookalikes that must stay required. The two target tests fail without the fix.
• Ran the CI jobs locally: lint, pytest -n auto and pytest -p no:randomly on 3.11, 3.12 and 3.13 (1990 passed), and a wheel build plus import.