feat(genai): add support for thought_signature in image content blocks

[Maciej Nachtygal (nhtgl)]

Description

Gemini image generation models return a thought_signature alongside generated images. This signature must be sent back in subsequent requests for multi-turn image refinement — without it, the API rejects the request with INVALID_ARGUMENT: Image part is missing a thought_signature.

Two bugs prevented this from working:

1. thought_signature on image Parts was not exposed through AIMessage.content_blocks (it was being dropped during parsing)
2. Even when extras.signature was present on a content block, _convert_to_parts had a pass stub instead of actually setting thought_signature on the outgoing Part

Both paths are now fixed: inline base64 images and file URI images correctly round-trip their thought_signature through content_blocks.

Relevant issues

#1837

Type

🐛 Bug Fix

Changes(optional)

• _compat.py: preserve thought_signature from extras.signature when converting content blocks back to Google GenAI format (both inline inline_data and file_data paths)
• chat_models.py: wire thought_signature into Part kwargs for image blocks (was a pass stub)
• Tests: roundtrip tests via content_blocks for both base64 and file URI image blocks; direct _convert_to_parts unit tests for both image block types

Testing(optional)

New unit tests in tests/unit_tests/test_chat_models.py:

• test_image_signature_roundtrip_via_content_blocks
• test_image_file_uri_signature_roundtrip_via_content_blocks
• test_convert_to_parts_image_base64_thought_signature
• test_convert_to_parts_image_url_thought_signature
  Integration tested manually against gemini-3-pro-image using the repro script from the issue — Turn 2 now succeeds.

Note(optional)

The issue also mentions skip_thought_signature_validator as a fallback. This PR does not add that — if the signature is present in content_blocks, it will be sent; if not (e.g. older messages), the caller is responsible. Addressing the validator flag is a separate concern.

AI involvement: This PR was developed with the assistance of Claude Code (Anthropic).