Unknown-column warning hides removed model inputs behind raw source columns (ten-name cap) #609

Follow-up to #530 / #531 (the aggregated unknown-column warning), found while removing two inputs from policyengine-us (PolicyEngine/policyengine-us#9605).

Problem

The warning names only the first ten unknown columns, sorted, plus a count. On a dataset that also carries raw source columns, those ten slots are taken by names that were never model variables, so a column for a removed model input is never named.

Measured 2026-09-25 by loading the certified US default dataset (populace_us_2024.h5@populace-us-2024-spm-20260909, sha256 6496cc43…) through policyengine_us.Microsimulation on the #9605 branch with policyengine-core 3.32.7. The single warning was:

The dataset contains 191 column(s) that do not match any variable in the tax-benefit system and were ignored: ACTC_CRD, AGI, ANN_VAL, A_AGE, A_ENRLW, A_EXPRRP, A_FAMREL, A_FAMTYP, A_FNLWGT, A_FTPT, … (181 more)

That branch deletes first_home_mortgage_interest and second_home_mortgage_interest, and the dataset carries both, but neither appears: uppercase survey columns sort first. The warning fires on this file whether or not a model input was removed, so it carries no signal there. On a dataset with no raw source columns it works as intended (policyengine-us test tests/microsimulation/data/test_removed_mortgage_interest_inputs.py asserts both names appear).

Options

1. Removed-input registry (preferred). Let a tax-benefit system declare removed or renamed inputs (name → replacement or guidance). At dataset load, emit a separate warning naming every dataset column that matches the registry, independent of the ten-name cap. This is the only option that tells a user what to do.
2. Full list at a lower level. Keep the capped WARNING and also log the complete sorted list at INFO or DEBUG.
3. Ordering. List snake_case names (the model's convention) before other names, so removed inputs outrank raw survey columns. Cheap, but a heuristic.

Acceptance

• A test that builds a dataset with more than ten unknown columns, one of them a registered removed input, and asserts the warning output names that input and its guidance.
• The existing Warn when a dataset carries columns that match no variable #531 tests still pass (known columns load unchanged; clean datasets stay quiet).

The data-side counterpart (certified files not shipping raw source columns or stale inputs) is tracked in PolicyEngine/microcosm#673.