Captions were being baked into the image by the FLUX model, causing
text to get cut off at the edges. Now the LLM generates scene and
caption separately — the scene drives image gen (with explicit
"no text in the image" instruction), and the caption is inserted
as a styled <figure>/<figcaption> HTML block at the top of the post.
Changes:
- rewrite_via_llm() returns array{scene, caption} instead of string
- build_article_prompt() and build_source_prompt() return {prompt, caption}
- Style suffix v4: removes "caption text" instruction, adds "no text" directive
- Image pipeline destructures prompt/caption, stores caption in attachment meta
- prepend_caption_to_post() inserts figure+figcaption HTML into post content
- Deduplicated generate_image_a/b into shared generate_single_image()
- One-time v4 migration auto-updates existing style suffix on deploy
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
214f57bff2