The gemma4 role fix (#5650) only matched Ollama-style names like gemma4:e2b. HF/vLLM/llama.cpp use gemma-4 with a hyphen, e.g. google/gemma-4-26B-A4B, which the substring check missed, causing the wrong tool role on non-Ollama backends.

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: [LiteLLM] wrong tool role for Gemma4 model #6334

2. Or, if no issue exists, describe the change:

Problem:
When using ADK's LiteLLM integration with Gemma 4 models served via vLLM or llama.cpp (using Hugging Face's naming convention, e.g. google/gemma-4-26B-A4B), tool call results are sent with role tool instead of the required tool_responses. This causes the model to hallucinate, produce garbage output, or loop until OOM. The existing fix in #5650 only checked for the substring "gemma4", which matches Ollama-style names (gemma4:e2b) but not the hyphenated Hugging Face naming (gemma-4).

Solution:
Replaced the plain substring check with a regex-based helper, _is_gemma4_model(), using the pattern gemma-?4, so it matches both gemma4 and gemma-4 naming conventions. Applied this helper at both call sites that previously used the substring check.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added test_gemma4_hf_style_naming_uses_tool_responses_role to tests/unittests/models/test_lite_llm_gemma_tool_role.py, verifying that google/gemma-4-26B-A4B (the exact model name from the bug report) resolves to role tool_responses.

Ran the full test file locally:
python -m pytest tests/unittests/models/test_lite_llm_gemma_tool_role.py -v
======= 8 passed, 4 warnings in 8.33s =======

Manual End-to-End (E2E) Tests:
Verified via unit tests above that both Ollama-style (gemma4:e2b) and Hugging Face-style (gemma-4-26B-A4B) model names correctly resolve to tool_responses, and unrelated models still resolve to tool as before (no regression).

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

N/A