Automated vehicles (AVs) may improve urban road safety, yet aggressive calibration and mixed-traffic interactions may produce non-linear safety effects. This study examines how AV penetration, fleet composition, and behavioral calibration shape traffic safety outcomes. Using SUMO microsimulation for Charlottenburg, Berlin, with TTC-based surrogate safety analysis, LightGBM modeling, and spatial hotspot mapping, we evaluate 12 scenarios varying AV penetration (0–100%) and fleet composition (small AV12, and larger AV46). Each scenario is tested under three AV time-headway settings, τ = 0.6, 0.8, and 1.0 s, yielding 36 configurations; human-driven (HDV) vehicle headway is fixed at 1.0 s. Results reveal a non-linear safety transition. Total conflict frequency declines with increasing AV penetration, reaching reductions of approximately 45% under full automation. However, severe conflicts increase during the mixed-traffic phase under aggressive calibration, τ = 0.6 s, indicating a transition-related safety paradox. Sensitivity analysis shows that τ = 0.8 s provides the most balanced performance, improving both total and severe-conflict outcomes. Spatially, conflicts shift from network-wide dispersion under HDV conditions to localized concentration at critical junctions under AV-dominant traffic. Fleet composition further moderates safety outcomes, with AV46-dominant scenarios showing higher safety improvements and more stable kinematic patterns. LightGBM identifies relative speed difference as the dominant risk driver. Although higher AV penetration and AV46 shares reduce average severity, they may increase tail risk under short headways. AV conflicts occur at lower speeds and lower ΔV, suggesting lower crash severity despite shorter temporal margins. The findings highlight context-dependent AV operation, particularly headway regulation and fleet composition.