[Kamilbenkirane]

Google image models never received the unified grounding image parameter added in #513: no Google images mapper wrote it, so grounding=True raised UnsupportedParameterWarning ("Parameter 'grounding' is not supported by model 'gemini-3.1-flash-image' and will be ignored.") and was dropped.

Google documents Search grounding for image generation:

• Nano Banana image generation (Interactions): "Use the Google Search tool to generate images based on real-time information, such as weather forecasts, stock charts, or recent events", with "tools": [{"type": "google_search"}], and "Grounding with Google Search is not supported by the gemini-3.1-flash-lite-image model."
• The model pages mark Search grounding supported for gemini-3-pro-image, gemini-3.1-flash-image and gemini-nano-banana-2.1, and not supported for gemini-3.1-flash-lite-image and gemini-2.5-flash-image.
• Interactions API reference (POST /v1beta/interactions): GoogleSearch, "A tool that can be used by the model to search Google", type: google_search, optional search_types.
• Grounding with Google Search on Vertex AI lists Gemini Nano Banana 2.1, Gemini 3.1 Flash Image and Gemini 3 Pro Image as supported models; the request carries a googleSearch tool. The GenerateContent reference defines Tool.googleSearch ("If not set, web search is enabled by default").

Fix:

• src/celeste/providers/google/interactions/parameters.py and src/celeste/providers/google/generate_content/parameters.py: a GoogleSearchMapper per wire. When the value is true it appends the existing WebSearchMapper().map_tool(WebSearch()) output to tools: {"type": "google_search"} on Interactions, {"google_search": {}} on GenerateContent, the same shape the text modality sends for Tool.WEB_SEARCH. None and False write nothing, because Google's default is no search tool.
• src/celeste/modalities/images/providers/google/parameters.py: InteractionsGroundingMapper and VertexGroundingMapper bind ImageParameter.GROUNDING, appended to GOOGLE_INTERACTIONS_PARAMETER_MAPPERS and GOOGLE_VERTEX_PARAMETER_MAPPERS.
• src/celeste/modalities/images/providers/google/models.py: ImageParameter.GROUNDING: Bool() on the three documented models, the same constraint flux-3-image uses. The two unsupported models stay unconstrained, so the value passes through for Google to reject, by design.
• tests/unit_tests/test_parameter_wire_contracts.py: a True and a False wire row for each backend list.
• CHANGELOG.md: one line under Features.

Not in this PR:

• True enables web search only. Image Search (search_types: ["web_search", "image_search"]) exists only on 3.1 Flash Image and Nano Banana 2.1, so mapping it would make grounding model-dependent. It stays reachable through extra_body.
• ImageOutput.grounding stays None. The search steps (google_search_result.search_suggestions) and groundingMetadata.searchEntryPoint are in the response but not surfaced: metadata["raw_response"] drops steps / candidates as content fields. Google requires displaying the search suggestions for grounded results (terms). Parsing them for images, reusing map_grounding_interactions / map_grounding_vertex, is a parsing change that needs your approval first.

Validation:

• make ci passes (767 unit tests).
• Baseline on main: _build_request(..., grounding=True) for gemini-3.1-flash-image emits UnsupportedParameterWarning and the body has no tools.
• Live, Interactions (POST /v1beta/interactions, gemini-3.1-flash-image, quality="512"), through client.generate with the raw response captured at _make_request. Prompt: "Visualize the current weather forecast for the next 5 days in San Francisco as a clean, modern weather chart with the dates and temperatures."
  • grounding=True: request tools: [{"type": "google_search"}], no warning. The response has google_search_call (queries: ["current 5 day weather forecast in San Francisco", ...], search_type: web_search) and google_search_result (with search_suggestions) steps, and usage.grounding_tool_count: [{"type": "google_search", "count": 2}]. The image is dated "Monday, November 9, 2026" (today) through Friday, Nov 13, with Veterans Day on Nov 11.
  • Omitted: no tools, steps model_output/thought only, no grounding_tool_count. The image shows stale dates, "Oct 26 – Oct 30".
• Live, the Vertex backend body against the GenerateContent wire: GoogleVertexImagesClient with an API key routes to Gemini API POST /v1beta/models/gemini-3.1-flash-image:generateContent, the same wire Vertex serves.
  • grounding=True: request tools: [{"google_search": {}}]. The candidate carries groundingMetadata with webSearchQueries ("current weather forecast San Francisco 5 days", ...), groundingChunks (weather.gov, accuweather.com) and searchEntryPoint. The image is dated Nov 10 – Nov 14, 2026.
  • Omitted: no groundingMetadata.
• No live Vertex call: the local ADC project lacks aiplatform.endpoints.predict (403 on publishers/google/models/gemini-3.1-flash-image).