Bug: cross-model thoughtSignature deletion breaks Gemini 3 function calls

[BIBIOTA]

Problem

When switching models mid-session (e.g. Claude via Antigravity → Gemini 3), tool calls from the previous model lose their thoughtSignature because transformMessages() strips it when !isSameModel.

Gemini 3 with thinking mode requires thoughtSignature on all function calls. Without it, convertMessages() falls back to [Historical context: a different model called tool...] plain text. The model then mimics that pattern in its own output, polluting the session.

Proposed Fix

1. transform-messages.ts: Stop stripping thoughtSignature on cross-model replay — let each provider's convertMessages() decide
2. google-shared.ts: For tool calls, keep any valid base64 signature regardless of model origin, only fall back to same-model check when invalid

Valid base64 signatures work cross-model for Google APIs. I have a working branch with tests passing.

[BIBIOTA]

⏺ Implementation

Branch: fix/preserve-thought-signatures-cross-model
PR: fix: preserve thoughtSignature on cross-model tool calls for Gemini 3

[kmixter]

The proposed approach of preserving real thoughtSignature cross-model (e.g. Flash → Pro) assumes signatures are compatible within the Gemini family, but the docs don't guarantee this — they only say signatures are valid for the model that generated them.

A safer fix: when a tool call has no valid signature for the target model, use Google's official dummy "skip_thought_signature_validator" instead of downgrading to text. This covers both cases (cross-model Gemini and cross-provider e.g. Claude → Gemini) without relying on undocumented cross-model signature compatibility.

I have a working implementation + test in #1491 (auto-closed by the bot — happy to resubmit once approved).