I think it's safe, but we need real cryptographers to figure this out (does Matthew Green approving miniLock count?). With proper Curve25519 implementations key generation timing shouldn't be secret-depended: it's the same scalar multiplication (crypto_scalarmult_base).