feat: add Codex CLI image generation bridge for non-Codex agents

[tznthou]

Closes #7

What

Adds scripts/call-codex-imagegen.sh — a shell bridge that lets non-Codex agents (Claude Code, Cursor, etc.) call Codex CLI's built-in image_gen to generate images, then hand off to the existing postprocess scripts unchanged.

Why

The skills currently require Codex's built-in image_gen, making them Codex-only. This bridge takes a different approach from #2: instead of calling the OpenAI/Gemini API directly from Python, it wraps Codex CLI itself as the image generation backend. This means:

• Same model, same output — built-in image_gen uses gpt-image-2, so generated images are identical to native Codex runs. No prompt adjustments needed.
• No new SDK dependencies — requires only Codex CLI (npm i -g @openai/codex), no openai or google-genai Python packages.
• Existing scripts untouched — generate2dsprite.py, extract_prop_pack.py, and compose_layered_preview.py work as-is.

This is complementary to #2, not a replacement — #2 covers the "no Codex CLI" scenario (pure API), while this covers the "has Codex CLI but orchestrating from another agent" scenario.

How

Any AI agent / CLI
  │
  │  call-codex-imagegen.sh -o output.png --bg "#FF00FF" "art prompt"
  │
  ├─ Builds meta-prompt → codex exec --sandbox workspace-write
  │                         └─ image_gen (gpt-image-2)
  │                         └─ Saves PNG + prints IMAGE_PATH: marker
  │
  ├─ 4-layer fallback path extraction
  │   (marker → regex → output check → find in $CODEX_HOME)
  │
  └─ stdout: absolute path of generated image

Features

• -i <file> reference image support (passes through to codex exec -i)
• -o <path> output path with auto mkdir -p
• --size, --quality, --bg options
• Soft timeout with heartbeat logging (default 300s)
• Structured error classification (TIMEOUT / RATE_LIMIT / AUTH_ERROR / NETWORK / SANDBOX)
• --dry-run for prompt inspection without API calls
• Observability via log file (tag: [codex-imagegen])

Verified

End-to-end pipeline tested on macOS (Codex CLI 0.142.3):

1. call-codex-imagegen.sh → gpt-image-2 generates 2×2 pixel art idle sprite (1024×1024, #FF00FF background)
2. generate2dsprite.py process --rows 2 --cols 2 --shared-scale --component-mode largest → chroma-key → frame extraction → QC
3. Result: 4 clean frames (92×108), transparent sheet, animated GIF, edge_touch_frames: []

Pipeline output

Step 1 — Raw gpt-image-2 output (via call-codex-imagegen.sh):

Step 2 — Chroma-keyed transparent sheet (via generate2dsprite.py process):

Step 3 — Animated GIF:

Map pipeline output (generate2dmap)

Also verified with the map pipeline — base map generation, 3×3 prop pack extraction (9/9 accepted, 0 rejected, 0 edge touch), and layered preview composition:

Foundation-only base	3×3 prop pack (raw)	Layered preview

Relation to existing issues

• Partially addresses generate2dsprite 生成图片时 Codex 的 image_gen 工具没有把生成图落到可读的本地 PNG 路径 #4 (image_gen path not landing): the bridge script has a 4-layer fallback to locate the generated image, covering the case where image_gen doesn't write to the expected path.
• Orthogonal to 我使用的时候发现有透明通道的部分老是有紫色残留，是不是目前只能处理像素图像 #3 (purple fringe on transparency): this PR doesn't change the chroma-key logic, but the same postprocess scripts that handle it are fully compatible.

Backwards compatibility

• requirements.txt unchanged.
• No changes to generate2dsprite.py, extract_prop_pack.py, compose_layered_preview.py, or either SKILL.md.
• Existing Codex workflows are unaffected — the bridge script is additive.
• .gitignore additions (.venv/, test-output/, .claude/) are for development only.