Conflicting gene_count_corr results with scv.pp.normalize_per_cell and csr_vcorrcoef source function

[thompjac24]

When working with a subset of the adata object I get gene_count_corr results that do not make sense with the input data. I believe that this applies to a large number of genes but will use our reporter gene transcript as an example.
When running scv.pp.normalize_per_cell with default settings on the full dataset (20321 cells) I get gene_count_corr= .3, a value that agrees with trend of the data.

When I subset the dataset to the cell types where we expect to find the reporter (4751 cells), I now get gene_count_corr= -.14, despite a nearly identical graph.

I went to the source code for csr_vcorrcoef (here) and replicated it with my data subset and performing the calculation in this way I get gene_count_corr= .10.

With these 4751 cells (and an analogous set of cells from a different condition) I keep getting a low number of 'velocity genes' and I think that this errant coefficient calculation is contributing to the lack of results in my data. For example, the following two genes are not considered 'velocity genes' despite the promising phase portraits.

I made sure that I am running scvelo 0.2.4 and have exhausted all ideas to troubleshoot on my own. Please let me know if you can help.