Replace CD stacking with cloning + national block assignment

[MaxGhenis]

Summary

Replace the current CD-level stacking approach with a simpler architecture:

Current approach:

1. Clone ECPS records per congressional district
2. Assign block within each CD
3. Calibrate per-CD weights

Proposed approach:

1. Take full national ECPS (drop geographic identifiers)
2. Clone it N times to get the same total records as current approach
3. Assign each record to a population-weighted random census block nationally
4. Derive CD (and all other geography) from the block
5. Use calibration to hit CD-level targets post-facto

Rationale

1. Simpler architecture - no CD-specific cloning logic needed
2. Block-first design - CD becomes just another derived geography
3. Natural population weighting - CDs have roughly equal population (~760k each), so pop-weighted national block assignment naturally gives equal representation per CD
4. Already have the infrastructure - block_crosswalk.csv.gz + block_cd_distributions.csv.gz provide all needed data

Implementation notes

• May need to create a national P(block) distribution file (simple: just use 2020 Census block populations)
• Calibration already handles fine-tuning to CD targets
• Could simplify data pipeline significantly

Related

• PR Add census block-level geographic assignment with comprehensive lookups #484: Add census block-level geographic assignment

🤖 Generated with Claude Code