fix(openai): preserve tool-result images on the Responses (Codex) path

[Destynova2]

The fidelity loss (targeted spot-check from the competitive analysis)

The analysis names "tool arguments / images lost in translation" as the deadliest
gateway defect class — HTTP 200 + valid stream + visible text, but the agent
turn is semantically corrupt. I spot-checked grob's Responses/Codex path and
found one:

A tool that returns an image (screenshot, computer-use, an MCP vision tool) puts
an image block in its tool_result. grob built the function_call_output with
content.to_string(), and ToolResultContent's Display renders an image block
as the literal text [Image]. The model gets no pixels — and hallucinates.

Verified live, end to end, through grob (gpt-5.5 over Codex OAuth)

Two-turn agent loop: user asks the model to call get_pixel, the tool returns a
solid-magenta PNG in its tool_result:

before (0.36.84, tool image → "[Image]"):   model answered "White"   ← hallucinated
after  (this PR, image → input_image part):  model answered "Magenta" ← saw it

The Codex backend accepts function_call_output.output as an array of
input_text/input_image parts — I probed the backend directly before writing
the fix (both pure-image and mixed text+image accepted). Isolated the test daemon
under its own GROB_HOME so the live check never touched the running daemon.

Change

• push_blocks_as_items now serializes a tool result via responses_tool_output:
  a bare string for text-only results (wire shape unchanged, prompt cache
  undisturbed), an array of parts when an image is present — text and image in
  original order.
• OpenAIResponsesItem::FunctionCallOutput.output widened String → serde_json::Value
  to carry either shape (only construction site is this transform).
• Unknown tool-result blocks are folded into text instead of silently dropped.

Tests

• tool_result_text_stays_a_bare_string — text path is byte-for-byte unchanged.
• tool_result_image_survives_as_input_image — image → input_image data URI, no [Image].
• cargo test --lib providers::openai::transform — 38 passed.

Same class as the message-content vision fix (#476); this closes the tool-output
half. Feeds the loss-matrix work in #484.

🤖 Generated with Claude Code