You should be aware that there are two known problems with K-Means and C-Means. The NP-hard (non-deterministic polynomial-time hard) problem causes issues with convergence due to optimization limitations, but is solved in the implementation of K-Means++ (yeah ESRI). The second issue is distortion of the Voroni regions that define the partition vectors. This is because class centers are based on means, making the K-Means algorithm very sensitive to outliers, non-normality and variability in data ranges. A solution to this is to use the mediod rather than the mean to specify cluster centers. The mediod statistic is designed to find centrality in n-dimensional space and, as such, is not sensitive to data range or normality. There is an implementation of K-Mediods in the R cluster package. ESRI's example K-Means script can be used as a template and modified to implement alternative models.
I am still skeptical about local autocorrelation as a solution for traditional clustering. The role of multivariate clustering is to find similarity in multivariate space and create n classes. The LISA statistic is designed to identify how values cluster in space (i.e., high-low, low-high, high-high, low-low). This becomes quite interesting in bivariate space and is technically clustering but is not considered multivariate clustering and is telling you something quite different. In your original posting it sounds like you want to cluster indicator variables into similar groups. This is exactly what multivariate clustering is designed to do.
Attached is R code for finding the optimal number of clusters (K) and creating a final cluster model using K-Mediod's. Good luck on your project.
Cheers,
Jeff
