OpenTelemetry Java Agent Traceparent Extraction Boundary in Custom Middleware

OpenTelemetry Java instrumentation automatically extracts the W3C Tracecontext traceparent header from incoming HTTP requests and injects it into outbound calls at the server and client boundaries. When a service introduces custom middleware between the servlet container and application handlers, or when request processing moves to an async or multi-threaded executor, the automatic propagation boundary may not align with the actual execution path. The goal is to ensure downstream spans join the same trace without manually parsing headers in every custom component.

The uncertainty is where the SDK expects the propagation extension to be active. Version-specific changes in the propagation library can alter how headers are handled, and the collector configuration may need explicit propagation settings. Without the tracecontext extension enabled, spans remain isolated to the originating process, but it is unclear which configuration layer takes precedence when custom middleware intercepts the request before auto-instrumentation runs.

What is the minimal SDK or agent configuration required to make custom middleware and async handlers participate in automatic trace context propagation? Does the Java agent enforce a specific extraction point that custom middleware must not bypass, or must the propagation extension be registered separately for non-standard execution paths? Which version-specific behaviors in the propagation library should be verified before relying on automatic header handling in a multi-threaded environment?