State pension: respect recorded amount for 'new SP' cohort (mirror old-SP scaling)

[vahid-ahmadi]

Bug

person_state_pension in src/variables/benefits.rs ignores the recorded person.state_pension value for the new SP cohort (those whose age is below basic_sp_min_age) and always returns the full parameter rate × 52.

The basic SP cohort branch correctly uses scaled-recorded-amount: it returns person.state_pension × (reform_rate / baseline_rate) if recorded, falling back to the parameter rate × 52 only when nothing is recorded.

The new SP branch should follow the same pattern.

Symptom (already measurable)

The parity harness in PR #53 shows a £946 divergence on the pensioner_couple synthetic scenario:

-- pensioner_couple --
  state_pension     rust=23,946  py=23,000  diff=946

Both pensioners have recorded state_pension: 11500 (totalling £23,000). Rust ignores those values and returns 2 × new_state_pension_weekly × 52 = 2 × £230.25 × 52 = £23,946. Python uses the recorded values. £946 = 2 × (£11,973 - £11,500).

Fix

Mirror the existing old-SP pattern, so that:

1. If person.state_pension > 0: return that amount, scaled by (new_state_pension_weekly / baseline_new_sp_weekly) for reform-correctness
2. If person.state_pension == 0: return new_state_pension_weekly × 52

This requires plumbing baseline_new_sp_weekly through Simulation::new and person_state_pension, parallel to the existing baseline_old_sp_weekly field.

Acceptance

• Parity harness shows £0 state-pension diff on pensioner_couple (or matches Python's recorded-amount semantics)
• Existing param_state_pension_new_weekly test still passes (it uses state_pension = 0 so reform behaviour unchanged on that path)
• A new unit test asserts that when state_pension > 0, the recorded amount is returned (×1.0 in baseline) regardless of the new-SP parameter rate

Effort

Small — ~15 LOC + 2-3 unit tests + the YAML parity case.

Surfaces from

PR #53 parity harness — exactly the kind of regression the harness was designed to catch.