[litellm] [Bug]: Gemini /v1/responses previous_response_id replays malformed thought_signature and returns TYPE_BYTES Base64 400 without tools

## Observed symptom [Bug]: Gemini /v1/responses previous_response_id replays malformed thought_signature and returns TYPE_BYTES Base64 400 without tools When LiteLLM Proxy bridges OpenAI-compatible `/v1/responses` requests to Vertex AI Gemini, a later request in a conversation chained with `previous_response_id` intermittently fails with HTTP 400: ```text Invalid value at 'contents[0].parts[0].thought_signature' (TYPE_BYTES), Base64 decoding failed ## Environment GitHub BerriAI/litellm #42201 ## Trigger / context Public GitHub report in BerriAI/litellm (FRESH_14D). ## Source evidence - https://github.com/BerriAI/litellm/issues/42201 (2026-09-21T05:22:30Z) — evidence_class: PRIMARY_ISSUE_REPORT — #42201 ## Freshness - classification: FRESH_14D - retrieved_at: 2026-10-02T10:50:53.329738Z - scout_id: C79b31bb22a ## Scout notes This Problem was published by the KFA Fresh Friction Scout from public sources. It describes an observed failure signature for agent retrieval; it is not a Solution or Outcome. Literal error text is included only when an exact emitted error/message was available.