The Mond River delta is an arid-zone river–delta–sea system on the northern Persian Gulf coast where episodic freshwater inflow, tidal exchange, hypersaline upstream runoff and high evaporation generate strongly time-dependent salinity states. This paper develops a physics-guided scenario-screening workflow for salinity intrusion and conservative pollutant transport under limited local bathymetric and continuous monitoring data. The study integrates corrected daily hydrologic records for 2004–2024, published Mond-estuary salinity observations, a conservative finite-volume hydrodynamic benchmark specification, a mass-conserving recurrent-learning model, scenario analysis and uncertainty propagation. The salinity component is bounded by published salinity observations from CTD/current-meter stations and reported Mond regimes, including freshwater-lens, positive-estuary, inverse and salt-plug states. Consequently, outputs are reported as published-observation-bounded envelopes rather than fully operational field-calibrated predictions. The PGML configuration uses PT-LSTM-style conservation of water volume and conservative pollutant mass, with a chronological 70/15/15 training, validation and independent-testing split, Adam optimization, grid-searched penalty weights and early stopping. Scenario envelopes indicate that dry-flow and combined-stress conditions can shift the inner delta toward high-salinity regimes consistent with published inverse/salt-plug behavior, whereas wet-event freshwater lenses may produce much lower salinity over short periods. Sensitivity analysis identifies river discharge, sea level, offshore salinity and river-boundary salinity as dominant uncertainty sources. The study provides a transparent pre-operational framework, a minimum salinity-monitoring network and reproducible supplementary datasets/scripts for future calibration of the Mond River–Delta–Sea system.