Adding LLM-generated node features to graph neural networks (GNNs) is widely reported to improve accuracy on standard benchmarks. We document a contrasting observation: when LLM features are introduced through pure input concatenation (rather than joint training, distillation, or prompt-conditioning), they can systematically degrade accuracy on the same homophilous benchmarks where end-to-end LLM pipelines succeed. With an MLP backbone on the Planetoid public split and bag-of-words original features, concatenating SBERT-encoded GPT-4o-mini TAPE features reduces PubMed test accuracy by -17.0 +/- 0.3 pp and Cora by -4.3 +/- 0.6 pp (CiteSeer -0.6 +/- 0.8 pp, within seed noise). The drop attenuates as we relax each condition (GCN / GCNII / GAT backbones, random splits, smaller encoders) and reverses on medium-homophily WikiCS (+4.4 pp) and ogbn-arxiv (+11.7 pp). To predict when concatenation helps versus hurts, we report a simple measure of LLM-alone discriminability, Delta_sig. Across 9 datasets Delta_sig correlates with the concatenation cost more strongly than homophily at point estimate (r^2 = 0.38 vs. 0.06; N=9, bootstrap CIs overlap). The bootstrap-best change-point is tau = 13.8 pp, and the rule "Delta_sig <= tau predicts non-positive concat cost" classifies 7/9 datasets correctly; since 60% of bootstrap samples place tau in [5, 30] pp, we treat Delta_sig as an interpretive lens rather than a precision filter. A dimension-controlled ablation on PubMed places the LLM-feature drop between same-source PCA (-2.3 pp) and same-dim Gaussian noise (-37.3 pp), ruling out dimensionality and weight-decay artifacts. Nine PubMed configurations fit a power law |Delta_concat| proportional to (sqrt(d_l/n))^1.31 with r^2 = 0.97; the low-Delta_sig, small-n corner is exactly where the headline -17 pp PubMed deficit appears.
We examine whether graph neural network (GNN) design rules generalize across benchmark families by studying aggregator selection (sum, mean, max) on 24 node-classification datasets spanning citation, heterophilic, LINKX Facebook-100, co-purchase, and co-authorship graphs. Edge homophily is only weakly predictive of the GIN-Sum versus GIN-Mean performance gap. Label informativeness predicts this gap well on legacy benchmarks but degrades substantially when Facebook-100 graphs are included. In these dense friendship networks, near-zero label informativeness coexists with a strong preference for sum aggregation, producing gains of 7-10% and up to 13% under extended training. Stochastic block model ablations, including degree-corrected variants matching Facebook-100 degree scales, fail to reproduce this behavior, indicating that mean degree alone does not explain the effect. Among several label-independent graph statistics, the spectral gap uniquely distinguishes these graphs from other low-informativeness datasets, with the effect localized to one-hop neighborhoods and replicated across architectures. We further identify training regimes that interact with aggregator choice and show that PNA can underperform the best single-aggregator GIN on standard citation benchmarks. Our results suggest that benchmark composition, rather than numerical insufficiency, determines whether design rules appear to generalize, and that the Facebook-100 regime provides a concrete target for future adaptive aggregation methods.