Health take-up wiring + certified-stack computability gate (real residuals from #604)

Two scoped items surviving the #604 investigation (its $0 premise didn't reproduce — $880B Medicaid / $7.2B CHIP / $236.9B ACA PTC on the certified pair; receipts there):

1. Wire Medicaid/CHIP take-up from stored facts. takes_up_medicaid_if_eligible defaults True and isn't stored, so medicaid_enrolled ≡ is_medicaid_eligible (both exactly 72,457,786 on Build P at 1.764.6). The artifact already stores has_medicaid_health_coverage_at_interview (8.7% of person rows) and siblings — the reported facts exist; they're just not connected to take-up. Design question (mirrors the #591 class-1/class-2 lens): populate take-up flags in the data from the at-interview facts + eligibility, or give the engine a take-up formula seeded from them. Related: #591 remainder (receives_medicaid, gated on policyengine-us#9206), #454 (ESI class).

2. Certified-stack computability battery. Nothing asserts "every variable pe.py reports is either stored in the artifact or computes nonzero-where-expected on the pinned engine." #604-class scares (and the real holes it pattern-matched to: #361 sparse zeroing, #454) should be a CI table, not an investigation. Sketch: for pe.py's US output set, on each certified candidate compute totals + nonzero counts, assert against per-variable expectations (nonzero / zero-allowed / range), publish the table as a cert artifact.

🤖 Generated with Claude Code

Start by reviewing the referenced #604 investigation and #591 remainder, then inspect the US output set in pe.py and the stored at-interview coverage facts. Done means take-up values are connected to those facts or a seeded formula, and a CI certification table checks totals, nonzero counts, zero-allowed cases, and ranges for each reported variable.