skill-compressorlisted

# Skill Compressor Optimize behavior per loaded token, not file length. A smaller candidate wins only when it preserves or improves every required quality cell. Treat plausible wording changes as hypotheses until execution evidence supports them. ## Compression loop 1. **Freeze:** preserve the exact baseline, promotion bank, graders, settings, thresholds, staged evidence budget, and stop rules. Include real failures. 2. **Measure:** inventory description, always-loaded body, each routed reference, full surface, and expected loaded tokens. Use `scripts/measure_skill.py`; provide observed route frequencies when available. 3. **Map:** create a behavioral ledger: stable rule ID, decision changed, activation condition, owning file, dependent rules, positive case, and failure if lost. Separate catalog routing from post-load execution behavior. 4. **Hypothesize:** consider unchanged control, removal, and the shortest replacement per seam. Prefer one independent variable. Reject noncompetitive variants by inspection; generating hypotheses does not require executing them. 5. **Screen:** run deterministic checks, then one observation per live variant on the most discriminating known failure. Reuse a condition-identical control observation. One hard regression may reject; one clean screen cannot promote. 6. **Falsify:** before replication, challenge survivors on a different failure mechanism, archetype, or grader. Run another trial only when its result c