docs: align tracked files to single coherent SoT

Contradictions resolved:

- AGENTS.md: add 'Known stale PRD lines' section (FR17, FR19, FR40, NFR, FR42)
  so PRD's cross-reference resolves
- AGENTS §6: remove 'tag sets' from Call 1 body description (Call 1 only uses
  species and attribute lists)
- PROMPTS.md: shipped defaults path corrected to backend/app/prompts/ (matches
  DEPLOYMENT repo layout)
- PROMPTS.md: Call 1 notes removed feat-* and fig-* from taxonomy list (Call 1
  returns free-text observations mapped to feat-* tags by server, not fig-*)
- PROMPTS.md: Call 2 template field names aligned with CONTENT-WOO defaults
  (width_avg_in, width_min_in, width_max_in, length_in_remainder)
- UX.md §112: remove 'tag lists' from Call 1 body description
- DATA-MODEL: prompt files location corrected from /data/prompts/ to
  backend/app/prompts/ (bind-mounted read-only to /app/prompts/)
- DEPLOYMENT: backup script updated (prompts are version-controlled in repo,
  no separate backup needed)
- docker-compose.yml: prompts mount corrected to ../backend/app/prompts

docs/UX.md

@@ -109,7 +109,7 @@ values always win and every value stays editable.
 ### What vision fills, and when
 
 When inference is on, Call 1 runs after the calibrated draft uploads. It sends all inventory photos
-downscaled to 1024 plus the store's synced species, attribute, and tag lists, and returns a guess
+downscaled to 1024 plus the store's synced species and attribute lists, and returns a guess
 with a confidence per field. The threshold is 0.7 by default:
 
 - A field at or above 0.7 pre-fills on review. It stays editable; the owner can change or clear it.