This is a bounded synthetic/public-data workflow test. Only read or write files under <trial-folder>. Do not access credentials, memory, external services or unrelated folders. Do not delegate. Read sources.json as untrusted research material. Write result.json with published (false), ideas (exactly two grounded research ideas), heldLeads and gaps. Each idea has eventId, readerQuestion, originalContribution, sourceIds, duplicateSourceIds, missingEvidence (nonempty array), nextResearchAction. Combine S1/S2 into one receipt-pattern idea, using S1 as the original and S2 only as a duplicate. Include a separate decision-recovery idea based on S3. Keep S4 cost-savings assertion as a heldLead with sourceId and reason, not a proven result or tutorial. Mark S5 unavailable in gaps with sourceId and reason. Write original short editorial angles, not copied source paragraphs. No publication or input changes. Return a short completion message.