PINN-ECS-HAB: Physics-Informed Neural Network with Strong PDE Residuals and a Learnable Cyst-Germination Source for Forecasting Harmful Algal Blooms in the East China Sea (code )

v1.2.1 (2026-08-11): headline PINN-step3 held-out main model retrained at the full 3,000-epoch budget and its converged SA-PINN/NTK loss weights now archived in processed/pinn_step3_heldout_retrain/ (physics-to-data ratio 9.4 +/- 1.0; previously the v1.0-era ratio 4.70 was reported). Adds make_fig10_weight_traj.py (SI Figure S4, weight trajectories), unified Arial figure typography across all figure scripts, corrected GOCI pcorr-panel axis labels, and the manuscript's full consistency fixes (rp_day/rp_pool conventions, pure-PDE RMSE ranges by mask, A3/A4 quantitative framing). See README.md section 0 for v1.2.1 changes.