Strict-parity evaluation is a methodological contract for comparing graph-aware retrieval systems against non-graph baselines. In a strict-parity comparison, the encoder, candidate pool, cutoff $k$, matching rule, and split policy are fixed across systems in the headline numbers, so any reported difference is attributable to the graph component rather than to interface mismatches. Token caps (per-passage or per-query length budgets) are reported separately as a diagnostic rather than folded into the headline comparison, because they conflate interface choices with graph structure. The contract is designed to surface cases where prior graph-RAG gains came from a richer retrieval interface rather than from graph reasoning itself.