Evaluation launch, judging, and failure handling

Curate-log parsing must never abort an evaluation. Parse /lore:curate JSONL one non-empty line at a time, surface the first type:error record, ignore malformed lines, and catch outer failures. Determine score validity from primary artifacts such as result.json and session traces, not this auxiliary log. When JUDGE_API_KEY is configured, use the real external Anthropic provider with Sonnet as the default judge; never let a weak or non-Anthropic answering model grade its own output.