feat(openai): forward image blocks on the Responses (Codex) path

[Destynova2]

The bug behind "grob has no vision support"

grob does translate image content blocks — on the Chat Completions path
(image_url), Gemini (inline_data) and Anthropic (native). But the Responses
transform used by the Codex / ChatGPT-OAuth path silently dropped them.

push_blocks_as_items matched Text, ToolUse, ToolResult and sent everything
else to a _ => {} arm. Its own doc comment said "text and image blocks collapse
into message items" — but Image was never in the match. So a vision request
through /codex/responses reached the model with the image removed: it saw only
the text, and being the Codex code agent it responded agentically
("let me look at the current folder… {cmd: bash -lc 'rg --files'}") instead of
describing the picture. That's the exact symptom reported.

Fix

Bring the Responses path to parity with Chat Completions. image blocks become
input_image parts — base64 → data: URI, URL sources pass through — interleaved
with input_text in the same message item, so a "describe this" + image turn
stays one message.

The Responses API takes image_url as a bare string (unlike Chat Completions,
where it's an object) — handled.

Verified live (gpt-5.5 over Codex OAuth, a solid-magenta 16×16 PNG)

before (0.36.79, image dropped):  model answered "Unknown"
after:                            model answered "Magenta"

Outbound /codex/responses body now carries:

{ "type": "input_image", "image_url": "data:image/png;base64,iVBORw…" }

Unit tests: base64 image → input_image data URI interleaved after the text
part; URL source passes through. cargo test --lib providers::openai::transform
— 36 passed. Full local prek gate — clean.

Not in this PR (the reporter's other two points)

• gpt-5.5 not routable / "not configured": grob normalizes inbound model
  names dots→dashes (gpt-5.5 → gpt-5-5) for lookup but matches [[models]]
  names verbatim, so a dotted config name never matches. Workaround: name the
  model entry gpt-5-5. A real ergonomic fix (normalize config names too) is a
  separate change.
• The dev config's routable models are code-oriented (Codex): that's a config
  choice, not a grob limitation — point a model/tier at a vision-first backend
  (Claude, or Pixtral via an OpenAI-compatible endpoint) for clean image
  description. With this PR, gpt-5.5 over Codex also describes images correctly.

🤖 Generated with Claude Code