Token-cap diagnostics in this paper are evaluations of how per-passage or per-query length budgets affect retrieval and generation, reported separately from strict-parity headline numbers. The motivation is that fixed-$k$ retrieval does not imply fixed token budgets, so folding token caps into headline comparisons would conflate interface choices (how much text each system is allowed to feed downstream) with graph structure (which concepts the graph policy selects). The token-cap analysis is explicitly scoped to isolating ordering and serialization effects within a fixed budget and does not propose a new context-compression method, which keeps the paper orthogonal to the length-aware selection and prompt-compression line of RAG work (knapsack-style selection, LART, LLMLingua). Reporting token caps as a diagnostic preserves the strict-parity contract for headline claims while still surfacing how systems behave under tighter or looser context budgets.