v0.20.1-rc2: model/parsers: rework gemma4 tool call handling (#15306)

Ollama, the open-source platform for running large language models locally, has pushed a new release candidate (v0.20.1-rc2) with a critical fix for users of Google's Gemma 4 model. The update specifically targets the model's ability to handle tool or function calls—a core feature for building AI agents that can execute code, call APIs, or take other programmed actions. The fix addresses a parser issue where Gemma 4's unique output format for tool arguments was not being correctly normalized into standard JSON, which could cause agent workflows to fail or behave unpredictably.

The technical core of the patch, referenced in pull request #15306, involves replacing a custom-built 'argument normalizer' with a more robust and stricter conversion process. This new method preserves strings that Gemma 4 outputs with quotes, ensures bare keys are properly quoted, and then unmarshals the final result as valid JSON. This approach maintains correct data types, so a number like `42` from the model stays an integer instead of becoming a string, and boolean values remain true or false. The update also adds test coverage to reject malformed inputs that a reference implementation would deny, increasing overall stability for developers building tool-using applications with the Gemma 4 model on Ollama.