Retrieval benchmarks must separate query provenance from the retrieval corpus

Observed 2026-09-04 while building COR-4025 in e3-solutions/sesh and sampling direct coding-session prompts through Cosmos. A retrieval benchmark can look artificially strong if the session that supplied a test query—or one of its fork/sibling sessions—is left in the searchable corpus, because the corpus may contain the query and its answer verbatim. Freeze a disjoint query-source and target corpus, exclude source session families (not only exact session IDs), and prefer a temporal cutoff where target sessions predate the query. This matters independently of lexical/vector weighting and should be a reportability gate, not an after-the-fact analysis.