Baifusi, a long-inhabited mountain-river town in China, coexisted with floods, yet damaging inundation has become concentrated in the recent engineering era. This study asks why flood risk has risen markedly over the last decade. We combine hydrological monitoring, field floodmarks, historical satellite images, drainage calculations, reservoir-operation rule learning, operation logs, surveyed terrain reconstruction, one-dimensional–two-dimensional coupled hydrodynamic scenario modelling, threshold analysis and interpretable surrogate-model evidence. The evidence shows that rainfall magnitude alone is insufficient: the 2016 event had a rare 24 h rainfall return period but a moderate flood-peak return period, while the 2023 event combined high Najitan release, Qiedao tributary inflow, Jinlongtan downstream backwater and local drainage blockage. Matched scenarios show a persistent 1.56–2.08 m tributary increment, whereas the downstream-backwater increment decreases from 1.67 m in the 5-year scenario to 0.03 m in the 100-year scenario. Relative to the 357.01 m hospital-entrance threshold, effective protection capacity decreases from about 42 years under the mainstream-only boundary to about 17 years with tributary inflow and 10 years under the full compound boundary. Archival images indicate a shift from sparse riverfront occupation in 2009 to a dense riverfront building belt by 2023. Pump-station calculations show that internal drainage capacity is 2.62–3.00 m³ s⁻¹, so ponding and backflow become critical when river stage is high. Flood-risk management in engineered mountain towns should move beyond a single return-period defence standard towards a process-based system that jointly manages upstream releases, tributary timing, downstream stage, drainage outlets and exposure control.