Nice work on the startup time. A couple observations from running similar optimization work.
Hermes is solid for cold starts but watch your production metrics closely. We saw weird memory spikes under load that weren't obvious in testing. The bytecode precompilation helps but you're trading parse time for cache invalidation headaches on updates.
The real win here is probably the manifest-based dependency graph. That's the hard part. Most teams never get there because it requires discipline across the team.
One thing: make sure you're measuring what actually matters. App cold start is great but what's your p95 first interaction time? We optimized ourselves into a corner once chasing startup metrics while users waited on network calls that happened immediately after.