Why a storm that makes no rain at the ground still starts fires: raindrop evaporation (virga), charge build-up, a strike, and the fire that follows.
Re-runs the evaporation model for every humidity from 10 to 95 % with your other settings and reports the flip from wet to dry storm.
No ground strikes yet.
The science
Virga. A drop of radius r in sub-saturated air shrinks by diffusion and latent-heat limited evaporation (Rogers & Yau): r dr/dt = (RH-1)/(F_k+F_d), with F_d = ρ_w R_v T/(D e_s) and F_k = (L/R_vT − 1) ρ_w L/(kT). So r² = r₀² − K t with K = 2(1−RH)/(F_k+F_d). Falling at terminal speed v(D) (a fit, Stokes for tiny drops) it lasts z = ∫ v dt; for constant v this is v r₀²/K (used as a check). Ventilation, drop breakup, a vertical temperature profile and evaporative cooling of the air are ignored, so real drops fall somewhat further than shown. Drop sizes follow a Marshall-Palmer distribution.
Charging. Ice-graupel collisions in the updraft separate charge. Here the charging current grows as I = I₀ (w/10 m/s)³ (I₀ = 0.55 A is an illustrative scale). Two charge layers in an active pocket of radius R form a field E = Q/(ε₀πR²); a flash starts when E reaches the in-cloud initiation field ≈ 2.5×10⁵ V/m (typical observed value, far below the 3×10⁶ V/m of sea-level air) and removes a fraction of Q. Energy W = ½ ΔQ E d with d = 3 km (order 10⁹ J, a typical value).
Ignition. Only ground flashes with a long continuing current heat fuel enough: P_cc = 1 − exp(−τ/40 ms), τ drawn from a log-normal. Fuel receptivity is a logistic in dead-fuel moisture, 50 % near 12 % moisture, lower in wind. These coefficients are illustrative, not calibrated to a fire agency model; they show the trend, not a forecast.
Fire spread. The front uses a Richards ellipse with the Anderson (1983) length-to-breadth ratio L/B = 0.936e^{0.2566U} + 0.461e^{−0.1548U} − 0.397 (U in mph), rate falling exponentially with fuel moisture, plus a Rothermel-type slope multiplier. Minimum arrival times come from a Dijkstra search on a 16-neighbour grid. Spotting, crowning and fire-weather feedbacks are not modelled.
Terrain height uses a smooth analytic surface (vertical scale true; 1 scene unit = 10 m). Not a safety tool: use official fire-danger services.