Protobuf wire format mismatch

A bad deploy of the pricing service is rolled back at 16:05 after a latency regression. The rollback uses the same rolling strategy and is reported 'complete' at 16:09. But errors don't fully clear, they settle at ~4% with proto: cannot parse invalid wire-format data on the quote RPC, intermittent and seemingly random across requests. Dashboards: the deploy tool shows 100% of pricing pods on the old (rolled-back) version; but the caller, checkout, was deployed an hour earlier and is still on its new version. The new pricing had added a field to the protobuf response (field number 9, a new nested message); checkout's new version started reading field 9. After rollback, old pricing no longer sends field 9, but ~4% of in-flight gRPC connections are sticky/long-lived. How do you triage and mitigate?