Would stratified sampling fix a silent failure?

Only partly, and only if you already have a guess about where the failure lives. Stratified sampling raises your review rate inside a slice you choose, a specific tool, a specific customer segment, a specific step in the flow, so a failure concentrated there gets caught sooner than uniform random sampling would catch it. That’s a real improvement when you have a hypothesis to test.

It doesn’t help with the failure you don’t have a hypothesis about yet. One that doesn’t correlate with any dimension you thought to stratify on hides in whatever slice you left thin, same as it would under plain random sampling. Stratification changes where your review budget concentrates; it doesn’t raise the total share of production you’re reading. The only way around that ceiling is reading more of production, or reading all of it.