Bug: Detailed anatomy prompt templates in prompts/anatomy/*.txt are never loaded

[Liohtml]

Summary

The repository ships rich, structured anatomy prompt templates at src/medcheck/prompts/anatomy/{knee,shoulder,spine}.txt, but build_prompt() in vision_analysis.py only uses the short hardcoded _ANATOMY_HINTS dict and never reads those files. The detailed templates are effectively dead code, and the Vision-LLM receives only a one-line hint instead of the systematic checklist authored for each region.

Location

• src/medcheck/pipeline/vision_analysis.py — _ANATOMY_HINTS dict and build_prompt()
• src/medcheck/prompts/anatomy/*.txt — detailed templates that are never read

Impact

• The carefully written knee/shoulder/spine checklists do not influence model output.
• Adding a new *.txt template (cf. Add hip anatomy prompt template #10 hip, Add ankle/foot anatomy prompt template #11 ankle/foot, Add wrist anatomy prompt template #12 wrist) has no effect, because nothing loads them.
• README advertises abdomen support, but there is neither a hint entry nor a template for it.

Suggested Fix

1. Add a loader that reads prompts/anatomy/<region>.txt when present and uses its content as the anatomy-specific instruction block.
2. Fall back to the existing _ANATOMY_HINTS dict, then to a generic instruction, when no template file exists (keeps hip/ankle/wrist/brain working).
3. Add an abdomen hint so the README's advertised regions are all covered.
4. Cache loaded templates to avoid re-reading on every call.

Category

Bug / Enhancement

Severity

Medium

Effort Estimate

~1 h