Test image & detection architecture overhaul

[colefoster]

Problem

The current detector/reader testing system has fundamental limitations:

1. Labels in filenames — can't encode multi-dimensional ground truth (species + detector bools + sprites on one image)
2. 1 image = 1 detector directory — prevents reuse across detectors/readers, forces duplication
3. No comprehensive screen model — detectors added ad-hoc, no shared screen graph for auto-ladder state machine
4. Re-labeling is destructive — new detector/reader means re-labeling everything from scratch

Solution

Replace with a manifest-based, screen-centric architecture:

• screens.yaml — canonical screen graph defining all 13 screen states, transitions, detector/reader registration with typed field schemas
• Per-screen image directories with manifest.json storing all reader labels
• Implicit negative testing — images not in a detector's registered screens are automatic negatives
• Web-based labeler on ash dashboard with OCR pre-population via ColePC job runner API

Key Decisions

Decision	Choice
Image storage	Per-screen directories (13 screens + 1 overlay)
Labels	manifest.json per directory
Negative tests	Implicit from directory membership
Screen graph	screens.yaml with transitions, detectors, readers, field schemas
Singles/doubles	Separate directories for in-battle screens only
C++ test runner	Add yaml-cpp + nlohmann/json, parse new structure
Labeler	Web UI on ash dashboard
OCR suggestions	Job runner API on ColePC (port 8422)
Migration	Rip and replace CommandLineTests/

Full PRD

See plans/test_image_architecture.md for complete specification including:

• Screen inventory and state machine transitions
• screens.yaml and manifest.json schemas
• Directory structure
• Negative test model
• C++ test runner changes
• Web labeler design (inbox, screen, validation views)
• OCR suggestion API spec

[Gin890]

Thank you for writing a plan for improving our test framework! Unfortunately we don't have any bandwidth to refactor it. Would you like to work on it?