Experiment Registry

v0.71 adds src/closed_world_lm/experiment_registry.py so training evidence starts with an explicit intent instead of a loose command.

Every experiment intent records:

• version and run id;
• component under test;
• hypothesis;
• allowed data sources;
• planned artifacts;
• training recipe id;
• acceptance gates;
• failure criteria;
• notes;
• decision.

Self-improvement answer cycles write experiment_intent.json as soon as an attempt directory exists, then close that intent with the promotion-gate result. The final intent is copied into the attempt report and the latest run report.

Transformer answer-training runs also write experiment_intent.json before training. They record baseline/final snapshot gates, closed-world data checks, no-pretrained-weight/tokenizer/embedding checks, and direct-answer branch screen gates when applicable. From v0.77 onward, transformer runs close through the constraint-first promotion report: quality metrics can affect promotion only after the closed-world constraints pass.

The registry is intentionally small: JSON artifacts, pure validation, and no hidden promotion behavior.