fix(langgraph): prevent duplicate streamed replies and tool results

[davidmckayv]

Summary

With ChatGPT sign-in, the LangGraph bot displayed streamed tool results and replies twice when the final history snapshot arrived. Preserve the provider response ID announced in the initial metadata-only chunk across later chunks, so streamed messages and the snapshot refer to the same IDs. Track IDs per model invocation and lane, with cleanup at invocation/run completion.

Fixes FOR-309.

Validation

• Extended the real SDK/AG-UI protocol regression to assert that streamed text IDs and tool parent IDs exist in the final snapshot. It failed before the fix and passed afterward.
• All 60 LangGraph harness tests passed.
• Actual Mac desktop with the patched harness: one Gmail read tool result and one summary, matching the audit record.
• The selected bot also opened Example Domain with its browser tool and rendered a real graphical bar chart without duplicate output.
• Rebased onto current main; the entire Python harness tree is unchanged by the rebase. Diff/working-tree checks passed.

The desktop check used a locally patched running container. Shipping this fix requires publishing the updated harness image; no desktop frontend change is required.

[linear-code]

FOR-309