UK spine: refuse malformed FRS claimant roles directly; check every supplied role reaches the engine

[MaxGhenis]

WIP (draft): the H2 parity fixture regeneration, the full test run and an independent review are still to come.

Follow-up to #1095's person types (6900190), which supplies is_claimant_or_partner as FRS adult-table membership (uk-data#524).

What was wrong

The spine and the stage spec both said that the UC claimant mask "already refuses a benefit unit without one or two adult records". It does not. The mask counts head, or adult with dependants, or any member of a unit whose DEPCHLDB is 0. Where DEPCHLDB is 0 beside child-table records:

• A unit with no adult record builds. Its child becomes the claimant, and is_claimant_or_partner is false for everyone in the unit.
• A lone adult's child becomes a UC claimant while is_claimant_or_partner says dependant. policyengine-uk 2.107.0 and later derive is_uc_claimant from is_claimant_or_partner, so with both supplied, UC would see a couple and every other programme a single adult.

Both cases were executed against main's own spine on the fabricated test fixture before the fix: both built without error.

What this changes

• frs_claimant_or_partner (in frs_spine.py) refuses a benefit unit with zero or three or more adult-table records itself, mirroring uk-data#524's derive_is_claimant_or_partner_from_frs_microdata.
• The spine refuses any person whose supplied is_uc_claimant differs from is_claimant_or_partner.
• The stage spec's text is corrected, and the source pins in the coverage manifest are regenerated.
• The policyengine-uk adapter now checks that is_claimant_or_partner and is_hbai_dependent_child reach the model as unchanged inputs, as it already did for is_uc_claimant. microcosm locks policyengine-uk 2.122.2, which reads both, so a loader that dropped either would silently bring back age inference.

Impact

None on the pinned FRS 2024-25. The aggregate-only measurement on the raw tabs (34,966 people in 18,850 benefit units) found:

• 9,986 units with one adult-table record and 8,864 with two;
• DEPCHLDB equal to the child-record count in every unit;
• is_uc_claimant equal to adult-table membership for every person.

So no built column changes. The refusals guard the next vintage.

Invariants (tested)

• Range: every benefit unit has one or two claimant-or-partner roles. Refused otherwise.
• Children: a child-table record never holds the role. The role does not come from UPERSON: a dependant can be UPERSON 2 under a single head.
• Adults: everyone aged 20 or over holds the role (FRS dependants are 19 at most), and so does every head.
• Determinism: record order does not change anyone's role.
• Differential: where the dependant count equals the child records, the UC mask and adult-table membership name the same people (Hypothesis).
• Agreement: the supplied is_uc_claimant equals is_claimant_or_partner, enforced at build time.

axiom: n/a: microsim data-pipeline change.

🤖 Generated with Claude Code