Cerbos doesn't support "merging" two data stores in this way. You could consider (something like) automating the injection of policies from file into your database store via some CI mechanism, and then handle the dynamic policies separately.
Generally speaking, principal policies are a special case (they override permissions for a principal with a given ID). If you wanted to act solely on principal IDs rather than roles as per more traditional RBAC+, then maintaining a set of principal policies (1 mapped per principal) and updating them via the admin API might well be the way to go. Question: would updating a rule for a role not be viable approach, rather than a rule for a principal ID?
From a scalability standpoint, the short answer is: it should be absolutely fine (the PDP is written with performance and efficiency as a primary concern).
The longer answer: the considerations would be the number of unique policies and the frequency of updates. If a human operator were updating policies on an ad-hoc basis, this is negligible. Even if there were many unique policies, Cerbos tries to maintain and continually prune an ephemeral cache of recently-used policies, so PDP instances shouldn't be heavily burdened from a memory standpoint. Even in the extreme case of many concurrent requests for a high range of unique policies; providing the PDP has access to enough memory, it should be absolutely fine.