Memory: ingest images into the brain (description and text from PNG/JPEG/WebP/HEIC) #7066

Problem

Adding an image to the brain fails today. OpenHuman's document converter chain (crates/openhuman-core/src/memory/convert.rs, TinyMemory's ConverterChain) has converters for plain text, HTML and Markdown, and, since #7034, PDF, Word, PowerPoint and Excel. No converter claims an image, so the chain answers UnsupportedFormat, and the image is not stored.

#6718 listed blob_ref extraction for images ("the chat model for images") as an ingest path. It was not built, and this issue tracks it on its own so #6718 can close.

What is needed

• A converter for common image formats (PNG, JPEG, WebP, HEIC at least) that produces text the brain can index:
  • a description of the image;
  • any text in it (screenshots, scanned pages, whiteboards).
• A decision on how that text is produced:
  • a vision model through the existing inference path, which costs credits per image and adds latency; or
  • local OCR, which is cheaper but has no description, and adds a native dependency. That must stay out of the always-on kernel floor (scripts/kernel-floor.limits).
• Long work goes off the request path, like the Office converter does (spawn_blocking / a background job).
• A clear error for formats that are still unsupported, shown to the user rather than a silent drop.
• Tests: a fixture per format, a failure case, and a check that the cost of each ingest is known.

Open questions

• Should a scanned PDF (image-only pages) go through the same path?
• What happens to PII redaction on text taken out of images?
• What are the size and dimension limits per image?

Related

• Stabilize the CortexDB-based memory engine in OpenHuman: validate hosted recall, cut over from tinycortex, then prove imports, context, efficiency and production readiness #6718, which is closing; this is the one ingest path it moves out.
• fix(memory): read PDF, Word, PowerPoint and Excel into the brain (#7023) #7034, which added PDF and Office ingest.
• Measure hosted memory's per-turn cost and stability: ingest paths, pack budget, prompt cache, PII, soak #7023, which measures per-ingest cost.