Impute below-threshold student loan holders

[MaxGhenis]

Summary

• add SLC liable-to-repay targets for Plan 2 and Plan 5 alongside the above-threshold targets
• fix the SLC parser to read the Higher education total rows instead of the first matching above-threshold row
• impute missing below-threshold England student loan holders into the FRS base dataset using the liable-to-repay shortfall against the observed repayer stock

Details

This keeps the active-repayer stock anchored to observed PAYE deductions, then fills the missing England tertiary-education borrower stock needed for the base dataset and future plan uprating.

The checked-in SLC snapshot implies these missing borrower stocks before weighting against the live dataset:

• Plan 2 shortfall: 4.955m in 2025, rising to 5.57m in 2028
• Plan 5 shortfall: 10k in 2025, rising to 2.165m in 2030

Testing

• uvx ruff check policyengine_uk_data/targets/sources/slc.py policyengine_uk_data/datasets/imputations/student_loans.py policyengine_uk_data/targets/compute/other.py policyengine_uk_data/targets/compute/__init__.py policyengine_uk_data/targets/build_loss_matrix.py policyengine_uk_data/tests/test_student_loan_targets.py policyengine_uk_data/tests/test_student_loan_plan.py
• uv run --python 3.13 pytest -q policyengine_uk_data/tests/test_student_loan_targets.py policyengine_uk_data/tests/test_student_loan_plan.py

Closes #281.

[vahid-ahmadi]

Entity-level mismatch bug in compute functions

Both compute_student_loan_plan and compute_student_loan_plan_liable in other.py mix person-level and household-level arrays:

plan = ctx.pe_person("student_loan_plan")       # person-level (n_persons)
repayments = ctx.pe_person("student_loan_repayments")  # person-level (n_persons)
on_plan = (plan == plan_value) & (ctx.country == "ENGLAND") & (repayments > 0)

ctx.country is sim.calculate("country").values which returns a household-level array (confirmed: country entity is household). This will fail at runtime with real data due to array length mismatch.

The old code handled this correctly by explicitly mapping region to person level:

region = ctx.sim.calculate("region", map_to="person").values

The test doesn't catch this because DummyCtx uses same-length arrays for everything and household_from_person is identity.

Fix: replace ctx.country == "ENGLAND" with something like ctx.sim.calculate("country", map_to="person").values == "ENGLAND" in both functions. Note that the imputation code in student_loans.py already does this correctly (sim.calculate("country", map_to="person").values).

Minor: Plan 2 below-threshold eligible pool is broader than Plan 2 cohort

plan_2_eligible uses plan_2_age_band (ages 21–55) but doesn't require the Plan 2 cohort filter (uni start >= 2012). In 2025, anyone aged ~31+ started uni pre-2012 and would be Plan 1 cohort, yet they're in the Plan 2 eligible pool. This likely doesn't affect the total count (the probabilistic assignment still targets the right number), but it distributes Plan 2 holders across a wider age range than reality. Worth considering whether to tighten the age band or add the cohort filter.

[MaxGhenis]

Addressed the review in 21b2243. The fix switches the SLC compute path back to person-level mapping instead of the household-level , and it tightens the Plan 2 below-threshold imputation pool to the estimated Plan 2 cohort as well as the age band. Local validation after the change: E902 No such file or directory (os error 2)
--> ...:1:1

Found 1 error. and ============================= test session starts ==============================
platform darwin -- Python 3.13.9, pytest-8.4.2, pluggy-1.6.0
rootdir: /Users/maxghenis/worktrees/policyengine-uk-data-fix-281
configfile: pyproject.toml
plugins: anyio-4.11.0
collected 11 items

policyengine_uk_data/tests/test_student_loan_targets.py ...... [ 54%]
policyengine_uk_data/tests/test_student_loan_plan.py ..... [100%]

============================== 11 passed in 9.23s ============================== ().

[MaxGhenis]

Addressed the review in 21b2243.

Changes:

• switched the SLC compute path back to person-level country mapping instead of the household-level ctx.country
• tightened the Plan 2 below-threshold imputation pool to the estimated Plan 2 cohort as well as the age band

Local validation after the change:

• uvx ruff check ...
• uv run --python 3.13 pytest -q policyengine_uk_data/tests/test_student_loan_targets.py policyengine_uk_data/tests/test_student_loan_plan.py (11 passed)

[MaxGhenis]

Follow-up: a post-merge review found that the SLC parser is still dropping literal zero target values, so the Plan 5 2025 above-threshold zero is not actually enforced in calibration. I’m fixing that in a small follow-up PR now.