The "clunky API call" problem is usually a symptom of overengineering the control plane. You shouldn't be toggling a live migration with an API; you should commit a change to the variable in your pipeline's configuration repository and let your existing CI process deploy it. That's the whole system you're building. If you can't trust a merge request to change the percentage, you can't trust the pipeline it's controlling.
You're dead right about secrets drift. A central vault isn't just for the new system. It's the forcing function to modernize the old one. The integration pain with Jenkins is real, but it's a finite, scheduled pain. The alternative is the infinite, random pain of a credential leak from a forgotten .properties file in a Jenkins workspace. Prioritize the vault migration *before* the canary percentage gets serious.
Been there, migrated that
You're spot on about the control plane. Committing a percentage change to your config repo is basically a canary deployment for the canary system itself, and that's a beautiful bit of symmetry.
But I've seen teams get stuck because they make the merge request process too heavy. If changing the variable requires five approvals and a change advisory board ticket, you've just recreated the clunky API problem with extra steps. The permission to merge should be wide open for the core team running the migration.
And yes, the vault point is brutal but true. The pain of integrating Jenkins with HashiCorp Vault feels huge, but it's a known, scoped project. The pain of a secret leaking from a deprecated Jenkins job is unknown, unbounded, and happens at 2 a.m.
Try everything, keep what works.
Your 10% is cargo cult analytics. You didn't decide on 10%, you guessed.
The only number that matters is the failure detection rate your business can stomach. If a critical bug costs $50k, you can model how many canary runs you need to find it with X% confidence before you lose that money. Start there.
Duplicating secrets guarantees they will drift. The vault integration pain is your project now, not a future risk.
If it's not a retention curve, I don't care.