Offline handwriting functions as a practical biometric sensing modality for identity verification and controlled stress testing of recognition pipelines, yet recovering writer-specific stroke appearance from only a few enrollment samples remains difficult. In trajectory-based rendering used to probe such sensing systems, stroke width is often fixed or only weakly adapted, so the sensed thickness cue fails to reflect individual writing habits. The key obstacle is an operational reliability imbalance: a learned width estimator and handcrafted geometric priors disagree in trustworthiness across writers and even along a single stroke, which makes a constant fusion rule brittle for few-shot personalization. We propose a confidence-guided dynamic fusion framework whose primary method is a deterministic confidence-to-$\beta$ map (not a learned gate): point-wise pseudo-label confidence from reference images sets per-point blending weights between learned and rule-based widths, with “reliability” denoting this operational blending prior. On 73 HWDB writers under a strict writer-disjoint protocol with 240 adaptation runs, dynamic fusion reduces writer-level variance and cross-seed instability relative to fixed fusion while retaining competitive point-wise accuracy. A security-oriented sensing analysis further contrasts detector responses under generic versus writer-adapted rendering, showing that the framework can serve as a controllable probe of verification sensitivity. The method plugs into existing trajectory generators and supports reproducible evaluation of personalization fidelity and biometric sensing robustness.