Hazard ratio of 0.45 seems implausibly high, especially when it's just the exact same treatment dichotomized to before/after 3pm. My money is on something other than a real circadian effect: either the result of a 'fishing expedition' in the data, or some other variable that incidentally varies by time of day. Maybe breaking randomization, leaving the drugs out for too long at room temp, etc. If you really believe this is an important and biologically plausible effect it should be a top candidate for a replication attempt.