[litellm] [Bug]: Gemini errors on streaming /v1/chat/completions bypass exception mapping (raw bytes message, no ContextWindowExceededError)

## Observed symptom [Bug]: Gemini errors on streaming /v1/chat/completions bypass exception mapping (raw bytes message, no ContextWindowExceededError) On `/v1/chat/completions` with `stream: true` against a `gemini/` model, an upstream error bypasses LiteLLM's exception mapping. For a Gemini context window 400, the proxy returns a 400 whose message is the raw upstream body as a Python bytes repr, with no mapped exception class. The same error without streaming, or on `/v1/responses` with streaming, is mapped to `ContextWindowExceededError`: `is_error_str_context_window_exceeded` matches Gemini's message (after #18282), but the chat streaming path never calls it. 429 and 503 behave the same way: the status is correct, but the message is the raw bytes repr. ## Literal error text ``` {"error": {"code": 400, "message": "The input token count (1200000) exceeds the maximum number of tokens allowed (1048576).", "status": "INVALID_ARGUMENT"}} ``` ## Environment GitHub BerriAI/litellm #43014 ## Trigger / context Public GitHub report in BerriAI/litellm (FRESH_14D). ## Source evidence - https://github.com/BerriAI/litellm/issues/43014 (2026-09-25T08:30:11Z) — evidence_class: PRIMARY_ISSUE_REPORT — #43014 ## Freshness - classification: FRESH_14D - retrieved_at: 2026-10-02T10:50:53.329738Z - scout_id: Ca9fb44614c ## 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.