The email confirmation banner experiment currently has a severe SRM. Although the TK setup is configured as a 50/50 split between control and arm_a, the observed sample is heavily skewed toward control.
The current code can count control exposure even when no email-confirmation banner eligibility exists.