Revise the hmrc SPI band exclusions per the #807 support re-measurement

[juaristi22]

Closes #807. Part of the #796 spine-deferral epic.

What moved

Class B — lifted (3 entries). hmrc/state_pension_income_band_50_000_to_70_000, hmrc/self_employment_income_income_band_50_000_to_70_000, and hmrc/private_pension_income_count_income_band_100_000_to_150_000 return to the calibrated surface. The "sparse cell" premise fails on measurement: 77–1,158 in-band carrier records (up to 535k weighted carriers), cell-for-cell identical on spine-i and spine-j (receipts: uk-candidate-eval out/excluded_cell_support_spine_{i,j}.json). Their v9-era adjudication predates #803 — excluding a band then silently widened its neighbour — and their v9 fits were overshoots (+27% to +34%), not support gaps.

Class A — reworded (11 entries). The 1m+ zero-support reasons now name the income concept — total income (ITA 2007 s.23, which excludes capital gains) — and cite the measured support (0 records for nine families; 0 SPI / 2 proxy for self-employment), so the premise can no longer drift into a "no wealth over £1m" claim. The spine carries capital gains to £20.4m and 6,925 records above £1m wealth; these cells stand because income excluding gains has no support there.

Class C — reworded (3 entries). The thin 500k–1m cells keep their exclusion (2–16 records) but cite per-cell measured support instead of the now-false "six-cell set" phrase, with re-measure-before-renewal noted.

Register census pins updated: 44 entries, 39 in the 2026-08-26 tranche, the three lifted cells asserted non-excluded. The compiled register surface (408 specs) is untouched — exclusions prune at apply time — so the frozen scoring-register pin and the UK spec digest both stand.

Measurement: the v12 seam run validates the lift

spine-j (sha 481fa7ff…) through tools/calibrate_uk_national_dataset.py on this branch (code pin 92da31ce), 1500 epochs / family_equal, register bound at 364 calibrated targets (was 361). Receipts: runs/uk-623-first-calibrated/spine-assessment-v12/.

All three lifted cells fit inside the 25% uk_target_fit band:

lifted cell	v12 fit
hmrc/state_pension_income_band_50_000_to_70_000	+22.2%
hmrc/self_employment_income_income_band_50_000_to_70_000	+18.5%
hmrc/private_pension_income_count_income_band_100_000_to_150_000	+23.3%

The one-sided sibling distortions collapse at the source, exactly as #807 predicted:

sibling	v11	v12
hmrc/self_employment_income_count_income_band_50_000_to_70_000	+38.6% (blocking)	+0.1%
hmrc/state_pension_count_income_band_50_000_to_70_000	+23.5%	+3.0%
hmrc/private_pension_income_income_band_100_000_to_150_000	−0.3%	−0.1%

Nothing else moves: the UC family is flat (caseload −20.8% → −20.6%), obr.capital_gains_tax slides just inside the band (−25.01% → −24.9%), headline loss 0.0297 → 0.0300 while binding three more targets, within-10% 90.6% → 89.8%.

uk_target_fit now blocks on 6 cells, all UC composition/two-child-limit (−56.5% to −26.5%) — the uk-data#452 residual beyond mechanism M2, i.e. the #736 adjudication input. The hmrc SPI band surface is fully green.

Note on the engine: the v11/v12 assessment runs resolved policyengine-uk 2.89.0 (lock floor >=2.88). Verified insensitive for the measured surface — ons.land.* initial estimates are identical across the split (corporate land does not flow through corporate_wealth), and the UC capital test reads corporate_wealth alone under both 2.89 and 2.92. The >=2.92 floor bump (#750's remaining microcosm-side item) follows in its own PR and must precede the certified cut.

🤖 Generated with Claude Code