Accurate grid load forecasting is safety-critical: under-predictions risk supply shortfalls, while symmetric error metrics can mask this operational asymmetry. We introduce an operator-legible evaluation framework -- Under-Prediction Rate (UPR), tail $\text{Reserve}_{99.5}^{\%}$ requirements, and explicit inflation diagnostics ($\text{Bias}_{24h}$/OPR) -- to quantify one-sided reliability risk beyond MAPE. Using this framework, we evaluate five neural architectures -- two state space models (S-Mamba, PowerMamba), two Transformers (iTransformer, PatchTST), an LSTM, and a probabilistic SSM variant (Mamba-ProbTSF) -- on a weather-aligned California Independent System Operator (CAISO) dataset spanning Nov 2023--Nov 2025 (84,498 hourly records across 5 regional transmission areas) under a rolling-origin walk-forward backtest. We develop and evaluate thermal-lag-aligned weather fusion strategies matched to each architecture's inductive bias. Our results demonstrate that standard accuracy metrics are insufficient proxies for operational safety: models with comparable MAPE can imply materially different tail reserve requirements ($\text{Reserve}_{99.5}^{\%}$). We show that explicit weather integration narrows error distributions, with the magnitude of improvement being architecturally determined -- iTransformer's cross-variate attention benefits significantly more than PatchTST's channel-independent design. Crucially, we identify a widespread susceptibility to "fake safety" in risk-averse forecasting: while probabilistic calibration reduces upper-tail errors, it achieves this by systematically inflating schedules (e.g., increasing bias by over 1,700 MW in severe cases) if left unconstrained. To solve this, we introduce Bias/OPR-constrained objectives that enable auditable trade-offs between minimizing tail risk and preventing trivial over-forecasting.