Fix tool result images and optional parameters for GPT, Gemini and Ollama

[stippi]

Fixes two problems the browser tools had with GPT models over the OpenAI Responses providers (seen with "GPT-6.1 Sol (ChatGPT)" on the ChatGPT subscription WebSocket provider). Images in tool results were also dropped by the Gemini and Ollama providers; this PR fixes those too.

Images in tool results never reached the model

Both Responses providers sent only the text of a ToolResult as function_call_output, so screenshots were dropped. output is now a list of input_text / input_image when the result carries images (the Responses API accepts an array there); text-only results stay a plain string. Verified live against the ChatGPT subscription WebSocket endpoint: the model described the screenshot correctly. The WebSocket provider also keeps an image's media type instead of labelling every image as PNG.

The same loss happened in two more providers, fixed natively without extra user messages:

• Gemini (vertex.rs): images travel as functionResponse.parts[].inlineData (multimodal function responses). Verified live with Gemini 3.5 Flash.
• Ollama: the tool message carries the images in its images field. Verified live with qwen3.8:27b.

In each live check, the model described the probe screenshot correctly with the fix. Without the fix it made up the content.

GPT filled optional tool parameters with placeholders

Without an explicit strict, the Responses API normalizes tool schemas into strict mode, which makes every property required. The model then sent tab_id: "", coordinate: [], duration: 0, and even a made-up coordinate: [0, 0] next to a ref.

• Opt out of strict mode: both Responses providers render tools through one helper that sets "strict": false. Live, the same click request came back as just {"ref": "ref_7"}. Every request (including side requests like compaction) renders tools the same way, so the prompt-cache prefix stays identical; running sessions lose their cached prefix once.
• Tool layer: the schema coercion pass drops "", [] and null from properties the schema does not list as required (0, false, {} stay). This also covers other models and MCP tools. browser_batch steps now go through the same coercion against the inner tool's schema.

Behaviour change: get_session_content with tool_names: [] used to hide all tool calls and now shows all of them, which matches the documented "omit for all tools".

Not covered

openai.rs (Chat Completions, also used by Groq, Cerebras, Mistral, Moonshot, OpenRouter and Z.ai) still drops images in tool results. Its spec allows only text parts in tool messages, so a fix would need a separate user message; that is left for later.

Testing

• New red/green unit tests for all four providers, the coercion pass and batch step parsing
• cargo test --release, cargo clippy --all-targets --all-features -- -D warnings, cargo fmt --check pass
• a_hung_page_fails_fast failed once under full-suite load (timing-based, unrelated) and passed on rerun
• Not yet exercised end to end in the app