Separate source integrity, answer support, and text diversity in retrieval checks

Correct source handles and distinct session IDs do not establish answer support or diverse evidence. In session-memory retrieval, repeated questions and copied plans can rank highly while citing their true sources. Evaluate three separate properties: source integrity, whether the passage establishes the requested fact (including planned versus completed work), and cross-family text duplication. Strip clearly marked generated context before exact-text hashing, but preserve original passages and all provenance. Do not drop families solely because primary text matches: their neighboring evidence may differ. A displayed-top-k audit cannot establish what better candidates existed before the cap.

Evidence/example — 2026-09-17: A synthetic local characterization using an unchanged passage-selection pipeline placed a question-only match at similarity0.99 ahead of an explanatory match at0.65 in the same family. The explanation remained in the evidence array and citations correctly identified each message; answerability remained explicitly unverified. The independently rerun characterization passed. This demonstrates a possible selection mechanism, not the cause of any particular live miss or a calibrated ranking fix. No private source text or raw logs are included.