Previously we presented a new approach to solving dynamic constraint satisfaction problems (DCSPs) based on detection of major bottlenecks in a problem using a weighted-degree method called "random probing". The present work extends this approach and the analysis of the performance of this algorithm. We first show that despite a reduction in search effort, variability in search effort with random probing after problem perturbation is still pronounced, reflected in low correlations between performance measures on the original and perturbed problems. Using an analysis that separates effects based on promise and fail-firstness, we show that such variability is mostly due to variation in promise. Moreover, the stability of fail-firstness is greater when random probiing is used than with non-adaptive heuristics. We then present an enhancement of our original probing procedure, called "random probing with solution guidance", which improves average performance (as well as solution stability). Finally, we present an analysis of the nearest solution in the perturbed problem to the solution found for the original (base) problem. These results show why solution repair methods do poorly when problems are in a critical complexity region, since there may be no solutions similar to the original one in the perturbed problem. They also show that on average probing with solution guidance finds solutions with near-maximal stability under these conditions.