Adjust pipeline to fit with new sales val data model

[wagnerlmichael]

We need to update the way the pipeline interacts with sales val columns because we recently made significant data model changes that incorporate manual analyst review. Res equivalent.

[jeancochrane]

I reran the ingest and updated the DVC lockfile in f107c88, then kicked off model run 2026-02-03-pensive-iris. The reports are broken for that model run, which is what we expect, but I did a quick check to confirm that predictions are populated as expected:

select meta_pin, meta_card_num, pred_card_initial_fmv
from model.assessment_card
where run_id = '2026-02-03-pensive-iris' and meta_class = '299'
limit 1000

I also checked the training data locally in RStudio to confirm that we see review flags:

training <- read_parquet("input/training_data.parquet")
training %>% count(sv_is_outlier)
training %>% count(sv_outlier_reason) %>% arrange(desc(n))

[jeancochrane]

Same as ccao-data/model-res-avm#433, I'm not sure why my version of DVC wants to break these onto multiple lines, but I figure it doesn't matter too much.

[jeancochrane]

Alright @wrridgeway, this looks good to me, so I'm sending it over to you since I contributed to it.

[wrridgeway]

Thanks guys, this is great. Appreciate the single outlier column that contains so much info.