Two independently built event‑camera systems were compared on the SPARK 2026 benchmark for 6‑DoF spacecraft pose estimation. The direct‑regression network predicts translation and rotation in one pass, using a global voxel for translation, a cropped view for rotation, and a physics‑motivated scale cue. The modular pipeline first detects 2‑D keypoints with HRNet, adapts the model with CycleGAN and DANN, pulls the pose with EPnP+RANSAC, and smooths it over time with an EKF. In the hidden‑test protocol the direct system achieved a relative translation error of 0.1335 m, an orientation error of 0.9101 rad, a composite score of 1.0436, and full 100 % coverage. The modular system achieved a lower translation error of 0.0285 m but a higher orientation error of 1.1767 rad, a composite score of 1.2052, and also 100 % coverage. These numbers illustrate a component‑wise reversal: direct regression reduces orientation error by 22.7 % while modular processing reduces translation error by 78.7 %. Because the benchmark adds a unitless translation error to a radian rotation, the ranking is dominated by orientation, explaining why the direct system tops the leaderboard despite its larger position error. The results motivate a hybrid design that would couple direct orientation estimation with geometry‑constrained translation recovery.