Gradient-Update Mismatch: Rethinking Conflict-Free Training of Physics-Informed Neural Networks | TickerVault