Mechanistic ODE/PDE simulators provide mechanistically interpretable forward models, yet instance-level parameter inversion remains a bottleneck in cardiac electrophysiology. Action-potential (AP) waveforms are among the most readily observable electrophysiological measurements, but they provide only compact and difficult-to-decode information about the underlying ionic mechanisms. Sparse and noisy measurements, partial observability, severe non-identifiability with multi-modal posteriors, and strict simulator-call budgets make single-solution estimation unreliable and costly. This work adopts a workflow-aligned objective: amortized inference rapidly proposes a set of candidate ionic-mechanism parameters from an observed AP waveform, and forward simulation verifies these candidates to return a verifiable space of mechanistic hypotheses rather than a single point estimate. This work proposes Amortized Simulator-based Parameter Inversion with Diffusion Transformers (ASPI-DiT), a DiT-based conditional generative framework for amortized simulator-based parameter inversion in cardiac electrophysiology. The method is motivated by a Bayesian decomposition of the conditional posterior score. ASPI-DiT models parameter dependencies using self-attention over parameter tokens and injects observational guidance using cross-attention at every denoising step, which preserves multiple posterior modes under partial observation and reduces mode collapse. For scalar electrophysiological parameters, this work introduces joint-projection tokenization to better preserve parameter coupling. To evaluate ill-posed inversion under a fixed simulator-call budget, this work introduces a set-based protocol. Candidates are valid when simulated trajectories match the observation within a threshold, and evaluation measures forward-simulation consistency, hit rate, and diversity. On a cardiac electrophysiology benchmark based on the O'Hara--Rudy model, ASPI-DiT infers nine ion-channel conductance scaling factors from AP waveforms and improves the trajectory consistency, valid-solution rate, and diversity compared with amortized Simulation-based inference baselines. On real hiPSC-CM optical AP recordings, ASPI-DiT generates valid candidate sets under signal drift and reproduces observed waveforms with physiologically meaningful differences.