IAM throttling during instance bootstrap

Your fleet scales out fine on normal days, but at your two highest-traffic peaks each week, scale-up stalls: new instances launch but take 4-8 minutes to become ready instead of 60 seconds, and a fraction never come up. Instance dashboards are clean (no vCPU/IP quota errors). Digging into instance boot logs you find the bootstrap is retrying calls to STS/IAM (AssumeRole) and the secrets manager with throttling errors, but only at peak. Your app's runtime traffic doesn't touch those APIs. How do you triage and mitigate the peak-only scale-up failure?