Store the UI start month as a dataset input instead of hashing person_id

CLAUDE.md

@@ -117,6 +117,7 @@ changelog.d/medicaid-ce-exclusions.md
 - For scale parameters that return integers, avoid using `rate_unit: int` in metadata (use `/1` instead)
 - Use `bool` instead of `int` or `/1` in `rate_unit` for scale parameters when appropriate
 - Program takeup is assigned during microdata construction, not simulation time
+  - So is every other stochastic input (for example the month unemployment compensation receipt starts): the dataset build draws it and stores it as an input variable, and formulas only compare against it. Seeding does not make a draw acceptable, and neither does hashing entity ids. `tests/code_health/test_no_engine_time_randomness.py` fails on the draw, sample and entity-hash patterns it recognises (not a proof that none exists), outside a reviewed exemption list and a ratchet of transitional draws that only shrinks
   - Changes to takeup parameters (SNAP, EITC, etc.) have no effect in the web app
   - These parameters should include `economy: false` in their metadata
 - **Labor Supply Response & Negative Earnings**: Use `max_(earnings, 0)` to prevent sign flips. Negative total earnings should result in zero labor supply responses.

changelog.d/deterministic-ui-start-month.added.md

@@ -0,0 +1 @@
+Add the `unemployment_compensation_start_month` input, the calendar month in which unemployment compensation receipt starts, assigned when the dataset is built. A dataset that stores it must give every person a start month in every year it covers, and no longer has the start month hashed from `person_id`; one that does not, as no Populace US release does yet, keeps the hash for now and gives unchanged results. Household situations that supply no `person_id` keep their start months; ones that supply it now get the same household-position months. A code-health check fails on the draw, sample and entity-hash patterns it recognises in the model, apart from reviewed exemptions and a ratchet of transitional draws that only shrinks.

policyengine_us/parameters/gov/usda/snap/work_requirements/README.md

@@ -21,7 +21,7 @@ land.
 | General work registration age exemptions (under 16, 60 and older) — 7 CFR 273.7(b)(1)(i) | Modeled | `meets_snap_general_work_requirements` |
 | General work registration exemption for working 30 hours weekly or earning the federal minimum wage times 30 hours — 7 U.S.C. 2015(d)(2)(E); 7 CFR 273.7(b)(1)(vii) | Modeled | `is_snap_work_registration_exempt_employed`; usual weekly hours are averaged over weeks worked (`snap_work_requirement_weekly_hours`) and annual earnings over 52 weeks (`snap_work_requirement_weekly_earnings`), compared with `gov.usda.snap.work_requirements.general.weekly_hours_threshold` and that threshold times `gov.dol.minimum_wage`; setting `gov.simulation.snap_work_tests_average_over_year` to false switches to usual hours and earnings per week worked |
 | General exemptions: disability, care of child under 6, care of incapacitated person — 7 CFR 273.7(b)(1)(ii), (iv) | Modeled | Shared between general and ABAWD checks |
-| Non-age work registration exemptions: student enrollment, unemployment compensation receipt — 7 CFR 273.7(b)(1)(v), (viii) | Modeled | `is_snap_work_registration_exempt_non_age`; unemployment compensation exempts only in its months of receipt (`is_receiving_unemployment_compensation`), allocated from `weeks_unemployed` as one block per year with a start month hashed from a supplied `person_id` (datasets or explicit inputs) or otherwise from the person's position in the household, so every axis copy of a household matches; or all 12 months when weeks unemployed are not reported |
+| Non-age work registration exemptions: student enrollment, unemployment compensation receipt — 7 CFR 273.7(b)(1)(v), (viii) | Modeled | `is_snap_work_registration_exempt_non_age`; unemployment compensation exempts only in its months of receipt (`is_receiving_unemployment_compensation`), allocated from `weeks_unemployed` as one block per year, or all 12 months when weeks unemployed are not reported. The block starts in `unemployment_compensation_start_month`: a dataset stores it (one without the column still hashes `person_id`, for now), and a household situation may supply it or otherwise follows the person's position in the household, so every axis copy of a household matches |
 | Non-age work registration exemptions: TANF-complying, drug or alcohol treatment participants, unemployment compensation applicants not yet receiving — 7 CFR 273.7(b)(1)(iii), (v), (vi) | Partially modeled | `is_snap_work_registration_exempt_non_age` consumes the input variables `is_complying_with_tanf_work_requirements` (gated on TANF enrollment), `is_in_substance_use_treatment_program` and `has_applied_for_unemployment_compensation`; they default to false unless supplied, and no Populace US stage produces them (data parity is tracked in [PolicyEngine/populace#248](https://github.com/PolicyEngine/populace/issues/248)) |
 | ABAWD work requirement of 20 hours a week averaged monthly, defined as 80 hours a month — 7 CFR 273.24(a)(1)(i) | Modeled | `meets_snap_abawd_work_requirements` compares average weekly hours (worked plus work program) times 52 / 12 with `gov.usda.snap.work_requirements.abawd.monthly_hours_threshold`; hours worked are averaged over weeks worked (`snap_work_requirement_weekly_hours`) |
 | ABAWD age exemption brackets, including the HR1 change of the upper bound from 55 to 65 — 7 U.S.C. 2015(o)(3)(A) | Modeled | `gov.usda.snap.work_requirements.abawd.age_threshold.exempted` |
@@ -69,6 +69,10 @@ supplied inputs:
 - `weeks_unemployed`: weeks spent looking for work; sets the months of
   unemployment compensation receipt. When it is 0 (not reported, as on
   ACS-based rows), any receipt exempts the person in all 12 months.
+- `unemployment_compensation_start_month`: the calendar month in which that
+  block of months starts. The dataset build assigns it, so it is an allocation
+  device, not observed timing. A dataset without the column keeps the start
+  month hashed from `person_id` for now.
 
 Population-level data parity is tracked separately in
 [PolicyEngine/populace#248](https://github.com/PolicyEngine/populace/issues/248).