Federated learning is an attractive recipe for training network intrusion detectors across organisational boundaries, but most existing federated IDS proposals stop short of three things reviewers reasonably ask for: a working defence against client poisoning, a head-to-head comparison against modern federated baselines, and statistically significant accuracy gains. We address all three. We introduce FedXAI-IDS+, a two-stage federated pipeline that warms up with differentially-private FedAvg using a round-aware adaptive top-k schedule (k anneals from 0.50 to 0.10 over fifteen rounds), then hands off to a server-side LightGBM refinement protected by a Krum-style filter on the per-client summaries. Across NSL-KDD, CIC-IDS-2017, and TON_IoT — averaged over ten random seeds in an IID regime, five seeds per non-IID Dirichlet level, and three seeds per poisoning level — FedXAI-IDS+ reaches binary F1 of 0.9946, 0.9975, and 0.9987 respectively, beating FedAvg, FedProx, SCAFFOLD, and a federated XGBoost ensemble at p<0.001 by Wilcoxon signed-rank, while statistically tying centralised XGBoost and LightGBM. When two of five clients send label-flipped training data, FedXAI-IDS+ keeps F1 at 0.991, whereas SCAFFOLD drops to 0.46 and FedAvg to 0.78. Ablation isolates the LightGBM refinement as the load-bearing component and shows DP noise, top-k compression, and the Byzantine filter are accuracy-neutral under clean conditions — defences that cost nothing. Code and run artefacts are released.