fix: use text injection for Gemini models to avoid thought signature errors

[Tarquinen]

Summary

• Detect Gemini models and inject as text parts instead of tool parts
• Avoids thought signature validation errors with Gemini 3+ models

Problem

Gemini 3+ models have strict validation requiring thoughtSignature on functionCall parts. When the plugin injects synthetic assistant messages with tool parts, providers without robust signature handling fail with 400 errors:

Function call is missing a thought_signature in functionCall parts.

Solution

Check if the model is Gemini (via modelID.includes("gemini")) and inject the prunable tools list as a text part instead of a tool part. This avoids the thought signature requirement entirely while maintaining the same functionality.

Model	Injection Type
Gemini	Text part
All others	Tool part (unchanged)