Summary of Changes

This pull request restores and explicitly defines the optimal default top_p threshold (default=0.95) across key configuration classes in google/genai/types.py (GenerateContentConfig, GenerationConfig, Model, and LiveConnectConfig).

Motivation & Problem Statement

In recent updates, the omission of explicit default top_p nucleus sampling thresholds coupled with default temperature: 1.0 causes notable sampling entropy, particularly when using thinking models on lower/medium reasoning budgets (ThinkingLevel.MEDIUM).

Empirical testing on real-world multimodal benchmarks (fine-grained OCR under motion blur, complex regulatory logic, and noisy audio processing) demonstrated that:

1. When top_p is left unbounded at temperature: 1.0, the model occasionally samples low-probability tokens in initial reasoning steps, causing cascading hallucinations (e.g., misreading ambiguous digits or misinterpreting domain-specific rules).
2. Setting top_p=0.95 by default restores the deterministic stability previously experienced in earlier stable releases (such as the workflow prior to recent upstream schema changes), while still preserving creative exploration when explicitly requested by developers.

Classes Updated

• GenerateContentConfig: top_p: Optional[float] = Field(default=0.95, ...)
• GenerationConfig: top_p: Optional[float] = Field(default=0.95, ...)
• Model: top_p: Optional[float] = Field(default=0.95, ...)
• LiveConnectConfig: top_p: Optional[float] = Field(default=0.95, ...)

Verification & Testing

• Built and validated wheel package locally on Debian 12 (Bookworm).
• Verified types.py serialization and schema compatibility with google/genai/models.py.
• Contributor License Agreement (CLA) signed via Google CLA portal.