You have a very strange (i.e., negative) estimate for rho_A and it seems that you are using PLSc.
With such a strange reliability estimate (similar to Haywood cases in CB-SEM where you have negative variances) you will get inadmissible solutions for some of your estimates such as path coefficients and loadings and subsequent estimates such as composite reliability and AVE. In contrast, Cronbach’s Alpha is not based on any PLS results, but only on correlations between the indicators. Hence, it is not affected by the strange PLSc estimates.
Solution: Estimate your model with normal PLS or define the problematic construct as composite (formative) in PLSc. Indicators pointing inwards.
The reasons is that PLSc makes very strict assumptions about the data structure and requires common factor model data. If this is not fulfilled, it gives strange estimates. Usually, this is a sign that your construct is simply not a common factor (reflective) and should be estimated differently.