Add census block-level geographic assignment with comprehensive lookups

[MaxGhenis]

Summary

• Add census block-level geographic assignment that replaces county-only methodology
• Assign a census block to each household, then derive ALL geography from it for consistency
• Look up comprehensive Census geographies from block GEOID:
  • County, tract, state (from GEOID structure)
  • CBSA/metro area (via NBER county crosswalk)
  • SLDU/SLDL (state legislative districts upper/lower)
  • Place/City FIPS (incorporated places and CDPs)
  • PUMA (Public Use Microdata Area)
  • VTD (Voting Tabulation District)

Data files

• block_cd_distributions.csv.gz (25MB): P(block|CD) from 2020 Census populations + 119th Congress BEF
• block_crosswalk.csv.gz (21MB): 8.1M blocks mapped to SLDU, SLDL, Place, VTD, PUMA from Census BAFs

Test plan

• Tests for block assignment shape and determinism
• Tests for geography lookup (county, tract, state from block GEOID)
• Tests for CBSA lookup from county FIPS
• Tests for SLDU/SLDL lookups
• Tests for Place/City lookups
• Tests for PUMA lookups
• Tests for VTD lookups
• Tests for integrated assignment returning all geography fields
• Tests for County enum integration (backwards compatibility)

🤖 Generated with Claude Code

[baogorek]

This looks very clean and well-structured. Approved, but based on the scope of this PR, and the fact that tests are failing due to no fault of this PR, I think it makes sense to finish #473 first. I will need a review after I make a few more changes.

[baogorek]

For my own reference:

  PR #484 Summary: Data Enhancement, NOT Architectural Change                                                                                                 
                                                                                                                                                              
  Core Change                                                                                                                                                 
                                                                                                                                                              
  This PR improves the geographic assignment granularity within the existing local area calibration system. It's an enhancement to data quality, not a        
  structural change.                                                                                                                                          
                                                                                                                                                              
  Before: Counties were assigned to households using county_assignment.py with P(county|CD) distributions.                                                    
                                                                                                                                                              
  After: Census blocks (15-digit GEOIDs) are assigned using block_assignment.py with P(block|CD) distributions, and ALL other geography is derived from the   
  block for consistency.                                                                                                                                      
                                                                                                                                                              
  What the block assignment provides                                                                                                                          
                                                                                                                                                              
  From a single block GEOID, the system now derives:                                                                                                          
  - From GEOID structure directly: state, county, tract                                                                                                       
  - Via crosswalks: CBSA/metro area, SLDU, SLDL, Place/City, PUMA, VTD                                                                                        
                                                                                                                                                              
  Changes to stacked_dataset_builder.py                                                                                                                       
                                                                                                                                                              
  The changes are minimal and preserve the existing flow:                                                                                                     
  # Before                                                                                                                                                    
  from .county_assignment import assign_counties_for_cd                                                                                                       
  county_indices = assign_counties_for_cd(cd_geoid, n_households, seed)                                                                                       
  cd_sim.set_input("county", time_period, county_indices)                                                                                                     
                                                                                                                                                              
  # After                                                                                                                                                     
  from .block_assignment import assign_geography_for_cd                                                                                                       
  geography = assign_geography_for_cd(cd_geoid, n_households, seed)                                                                                           
  cd_sim.set_input("county", time_period, geography["county_index"])  # backwards compat                                                                      
  cd_sim.set_input("block_geoid", time_period, geography["block_geoid"])                                                                                      
  # ... etc for other geography variables                                                                                                                     
                                                                                                                                                              
  New data files                                                                                                                                              
                                                                                                                                                              
  - block_cd_distributions.csv.gz (26MB) - P(block|CD) from Census population                                                                                 
  - block_crosswalk.csv.gz (22MB) - Maps 8.1M blocks to other geographies                                                                                     
                                                                                                                                                              
  Unrelated changes also in this PR                                                                                                                           
                                                                                                                                                              
  - Adds partnership_se_income variable from PUF (Schedule K-1 Box 14)                                                                                        
  - Pandas 3.0 compatibility fixes                                                                                                                            
                                                                                                                                                              
  Bottom line                                                                                                                                                 
                                                                                                                                                              
  This is data improvement - more granular geographic data within the existing CD-stacked dataset architecture. The system still builds CD-stacked datasets   
  the same way; it just has richer geographic information.         

[MaxGhenis]

Let's add ZIP from block too