16 households have a negative vehicle count on 1.58.0 (WAS missing-value code in the wealth imputation) #574

What happens

On the released 1.58.0 enhanced FRS (enhanced_frs_2024_25.h5, SHA-256 5ba399ea…60b6), 16 households store a negative num_vehicles. The lowest is −8.

Where it comes from

num_vehicles is imputed from the Wealth and Assets Survey by the wealth QRF. generate_was_table (wealth.py) maps vcarnr8 to num_vehicles without cleaning. Two of the 15,128 WAS training rows carry −8, which looks like a WAS missing-value code (not checked against the WAS codebook). A QRF predicts by sampling training values, so the code reaches the dataset.

The has-fuel model in consumption.py already clips the same column at zero for its own training (was["num_vehicles"].fillna(0).clip(lower=0)), so it is unaffected there. Anything that reads num_vehicles from the built dataset sees the negatives.

Other WAS training columns with negatives (num_bedrooms in 14,485 rows, is_renter in 11,548) look like the same missing-value codes and are worth checking at the same time; net_financial_wealth and wealth can be negative legitimately.

Ask

• Treat WAS negative codes as missing in generate_was_table before training.
• Add a built-dataset test that num_vehicles is non-negative.

Found while tracing #565 (#572), which is about VAT and does not touch this.