Battery State-of-Health Estimation Under Short Causal Observation, Unseen Cells, Condition Shifts, and Embedded Deployment

State-of-health (SOH) estimation from short measurement windows can reduce reliance on complete charge or discharge cycles, but its reliability requires evaluation across unseen batteries, operating-condition shifts, and embedded hardware. This study evaluates SOH estimation from a nominal 10 s resampled voltage, current, and temperature window using NASA battery data. The window contained a median of one raw measurement, and approximately 90.9% of grid positions were interpolated. The 10 s duration denotes input width; actual onset-to-decision time was 15 s + δ (median: 21.812 s; 95th percentile: 29.016 s). Two Random Forest (RF) models were compared using strict leave-one-battery-out validation across 34 cells and adaptation-free condition holdouts. RF20x7 achieved a battery-macro mean absolute error of 7.096 percentage points versus 7.202 for RF150. The paired difference (−0.106 percentage points; 95% CI: −0.478 to +0.275) did not establish superiority, and equivalence was not tested. In the temperature-family holdout, RF20x7 reached 13.990 percentage points, a +6.997-point increase over the matched 31-cell reference. On an ESP32-C6, mean RF inference required 62.302 µs, while feature extraction and inference required 584.141 µs on pre-resampled inputs. These results support post-resampling computational feasibility but highlight reduced accuracy under unseen conditions.