feat: add builtin file_read tool for multimodal files

[khj809]

Summary

Adds a file_read builtin tool that lets the LLM autonomously read files from the runenv(both local and sandbox) — both text and image files — within the agent loop. This enables multimodal file analysis (charts, screenshots, photos) without the caller needing to pre-load content into the message history.
The implementation mostly follows Claude Code's FileReadTool with limited file types.

New: BuiltinToolProvider::FileRead / file_read tool (src/tool_impl/builtins/file_read.rs)

• Detects file type by extension: PNG/JPEG/GIF/WEBP return Part::Image (embedded base64); all other readable files return Part::Text
• Unsupported binary types (PDF, ZIP, EXE, WASM, …) return a structured error
• Non-UTF-8 text files return a structured error
• Supports offset and limit parameters for windowed text reads (character-based)
• Reads via RunEnv::read() so it works in both local and sandbox environments
• Registered as BuiltinToolProvider::FileRead {} in src/tool/provider.rs

Provider marshal fixes for Part::Image in Role::Tool messages

Previously all providers only used contents[0] of a tool-result message and only as plain text/JSON. Now each provider correctly handles images in tool results:

Provider	Fix
Anthropic (anthropic.rs)	tool_result.content is now the full contents array; Part::Value is wrapped as {"type":"text","text":"..."} content block (required by the Anthropic API)
Gemini (gemini.rs)	Images go as sibling inline_data parts in the outer parts array alongside functionResponse — the FunctionResponse proto has no parts field, so nested functionResponse.parts was wrong
OpenAI Responses (openai.rs)	function_call_output.output becomes a content array with input_image blocks; image_url is a plain string (not {"url":"..."} object)
ChatCompletion (chat_completion.rs)	Images in tool results are substituted with [image: mime_type] / [image at url] text placeholders, since ChatCompletion backends don't reliably accept inline images in tool results

Gemini thinking model support (gemini.rs)

Gemini 3 thinking models return a thoughtSignature that must be replayed verbatim in subsequent turns. Added:

• Capture thoughtSignature from thought text parts and from functionCall parts (sibling field) during unmarshal
• Include thoughtSignature back in thought text parts and alongside functionCall parts during marshal
• marshal_messages now always passes include_thinking = true so signatures are never dropped mid-conversation

Shared utility (src/util/truncate.rs)

Extracted middle_truncate into src/util/truncate.rs, removing local copies from bash.rs and sandbox.rs.

Test plan

• cargo test --lib tool_impl::builtins::file_read — 8 unit tests covering text, image, offset/limit, UTF-8 error, unsupported extension, missing path, nonexistent path
• cargo test --lib lang_model_impl::api::gemini::tests::test_tool_result_with_image — 2-turn Gemini 3 interaction: step 1 gets real functionCall + thoughtSignature, step 2 sends JPEG tool result and verifies text response (model correctly describes Jensen Huang photo)
• cargo test --lib lang_model_impl::api — all 13 provider unit tests pass; test_tool_result_with_image for Anthropic and OpenAI skipped without API keys (require compile-time test_with::env)
• Manual smoke test with Anthropic/OpenAI/Gemini test_tool_result_with_image (requires API keys at compile time)

🤖 Generated with Claude Code

[jhlee525]

I'll merge file_read tool & read tool in #395