/lore:curate JSONL parsing

Trap: whole-output JSON.parse() looks correct for /lore:curate, but the command emits JSONL, so multiline output fails and can silently hide reported errors. Fix: parse each non-empty line independently, ignore malformed lines, surface the first type:"error", and contain outer parsing failures. Determine score validity from primary result.json and session traces, never the auxiliary curate log.