Add certification cadence, BBCE, and Medicare premium band features

[MaxGhenis]

The Eric/Ben feature round

Implements the three feature families from the 2026-08-09 meeting as 17 additive model features, wired through the official-error classifier, both hurdle stages, and every distributional model:

• F1 certification cadence — LASTCERT/CERTMTH semantics verified against the FY2024 techdoc (the codebook's definitions, not the prompt's guess): months-since-cert, period length, final-two-months indicator, elderly/disabled interaction, structural-inconsistency and sentinel-missing flags. No state elderly-cadence registry: the 24–48 month techdoc passages are SSI-CAP case-identification rules, not statewide policy.
• F2 BBCE — state registry from the official FNS State Options Reports (16th ed. FY2024: 44 adopters; nonadopters AK AR KS MS MO SD TN UT WY, locked by a committed-data test), with historical edition mapping disclosed per model year. CAT_ELIG stays excluded from features (codes 2/3 don't uniquely identify BBCE; cross-check in report).
• F3 Medicare Part B premium bands (Ben Molin's insight) — elderly households' reconstructed gross medical expense (FSMEDEXP + $35 floor) vs the sample-month's CMS premium: premium_only (±$5) and just_above_premium (+$5..+$50]. All 36 elderly Colorado $165-censored SMD cases land in the just-above band after reconstruction (a naive literal comparison misclassifies 6) — the Colorado standard sits in Ben's zone exactly as he predicted.

Results (additive, same frozen protocol; not causal)

Metric	Committed	New
FY2024 weighted ROC AUC	0.766626	0.767947
PR-AUC	0.355304	0.358093
Hurdle stage-1 AUC	0.835617	0.839738
Equal-state unfactored MAE	1.827pp	1.714pp

Distributional coverage is mixed-to-worse (mean abs gap 4.38→4.64pp) and retained honestly. Full report: analysis/FEATURES_REPORT.md.

Gates

• Byte-identical full-pipeline rerun 7/7; then a post-gate rerun after pinning the official State Options PDF SHA-256 (96b1e5c5…, cover-verified) showed provenance-only drift — 4 leaf diffs, none numeric; app base_model.sha256 pin intact.
• 151 tests, ruff clean (CI-exact commands in report).
• Cross-model review: sol lane (6 passes) + fable gate review of feature builders, band predicates, and boundary logic.
• ASSET_V bumped 20260809b→c so clients refetch the rolled model JSONs.

🤖 Generated with Claude Code