Which Training-Data Axes Matter for Conditional Imitation Learning in CARLA? A Leave-One-Out Ablation UnderPure and Guardrailed Deployment

Conditional imitation learning (CIL) for CARLA depends on diverse expert data spanning map, weather, traffic, and control-perturbation axes, yet it remains unclear which axes actually drive closed-loop behavior, and whether ablation conclusions survive deployment guardrails. We train a matched baseline and four equal-budget leave-one-axis-out (LOO) variants of a fixed CIL architecture (v14) and evaluate each under three nested tiers: pure policy rollout, minimal traffic-rule shields, and a fully deployed stack with route blending and recovery. The factorial design comprises $18\times 5\times 3$ scenario-variant-tier cells, each repeated under n = 5 traffic-seed replicates (1350 closed-loop episodes) to estimate NPC-seed variance on every eval stack. No single withheld axis dominates pooled outcomes. LOO effects are tag- (scenario-category) and spawn- (vehicle starting location) specific: removing perturbation-labeled recovery data costs $-285$ m on geometry_stress but can gain distance on in-distribution spawns; removing multi-town data changes held-out Town05 mobility on some routes while depressing others. Axis-importance rankings reorder across tiers; Kendall $\tau$ between pure and full rankings is $0.0$, and guardrails compress or erase pure-tier gaps (e.g., drop_perturbation pooled distance $\Delta$ from $-93$ m to 0 m). Pure-tier seed replicates show that geometry-driven lane-tracking degradation under drop_perturbation is seed-stable; traffic-axis and full-tier drop_traffic loads are more seed-sensitive. We release the evaluation ledgers, parsing scripts, and analysis tooling with the paper. Training-data ablation claims should report per-scenario or tag-stratified LOO metrics under pure evaluation; guardrailed tiers are supplementary deployment checks, not substitutes for isolating what the policy learned.