feat(telemetry): propagate W3C trace-context on outbound MCP HTTP calls (#60177)

[webtecnica]

Propagate W3C trace-context on outbound MCP HTTP calls via traceparent header from active OpenTelemetry span. Closes #60177

[teknium1]

Thanks for tackling MCP observability. The gap exists on current main, but this implementation does not yet propagate an agent call's active trace context.

Problems

• tools/mcp_tool.py:4107 sends MCP tool calls through an already-established ClientSession, while this diff injects traceparent only during _run_http() transport setup. A connection-start header cannot carry a fresh span for each tool invocation.
• MCP lifecycle work runs on the daemon mcp-event-loop thread (tools/mcp_tool.py:3731-3744). The scheduler documents that work on that loop receives its own thread context and only restores HERMES_HOME (tools/mcp_tool.py:3801-3811), so trace.get_current_span() at connection time will not normally observe the agent caller's span.
• The PR has no tests for active-span propagation, the cross-thread context boundary, or per-call HTTP headers.
• AGENTS.md:118-121 requires a generic user-facing opt-in before outbound telemetry/identifier tagging; this diff adds none.

Suggested changes

• Establish the approved opt-in/observability ownership first, then capture caller trace context before the MCP loop boundary and inject it per outbound MCP request/tool call. Add recording-transport coverage for span/no-span and configured-header cases.

Automated hermes-sweeper review.