[codex] Stabilize panel-local paper figure contracts

[yiweiqin]

Summary

Stabilizes dense multi-panel paper figures by making panel-local node instances and relations the authoritative scientific contract, while preserving repeated labels across training, inference, and architecture panels.

What changed

• extract concise labels and spatial context from the target overview figure
• compile independent panel_graphs with evidence-backed local nodes and connectors
• treat descriptive dataset headings as groups instead of inventing data-flow arrows
• synchronize prompts, semantic blueprints, overlays, and validators with panel-local instances
• preserve repeated visible labels and group/container titles
• normalize OpenCV HoughLinesP output across (1, 4) and (4,) shapes so editable PPT rebuilding works across OpenCV versions

Why

The previous global graph could merge repeated components across panels, create false cross-panel relations, and overfit labels such as dataset descriptions into operational inputs. The editable rebuild path also failed on newer OpenCV return shapes.

Validation

• 268 repository tests passed with python -m unittest discover -s tests
• python -m compileall -q rfs tests
• git diff --check
• TabPFN hot cached semantic compilation: 1.12 seconds
• TabPFN three-seed Image-2 stability audit: 3/3 production pass, mean 0.9907, worst 0.985