Build E: certification-candidate staging build from post-#287 main (fix target_compilation OOM first)

Goal

Run Build E — the certification-candidate US staging build — from current main, which now carries the complete recipe: #266 immigration, #279 input-carry, #280+#287 full-surface chained target aging (default ON, 1,480 targets, ×1.1203 chained TY2022→TY2024; national AGI $15.29T→$16.62T, income tax $2.11T→$2.36T at the target level), #286 stuck-at-default input gate, #283 L2/concentration diagnostics, pinned Ledger bundle enforcement.

Two prerequisites (both builds B and D died before calibration)

1. Fix the target_compilation kill. Builds B (us-2024-immigration-266-b-20260702T114815Z) and D (populace-us-2024-32fcf6b-de00cae57dc7-20260702T145207Z, nonpreemptible, 220 GiB) both died at the same locus: mid-target_compilation on the 224,026-household pool, with 2 of 5 JCT reform materializations complete — D's death was a hard kill, most likely OOM (the driver's own warning puts dense worst-case near ~120 GB; observed exceeds 220 GiB), with the 10h timeout secondary. The JCT reform materialization step (full engine runs over the pool, despite #236's streaming) is the prime suspect. Options: chunk/stream the reform materialization further, checkpoint per-reform (the #227 checkpoint machinery may already support resume — verify why the restart lost the materialization cache in B), or provision >300 GiB.
2. Inputs per #287's provenance contract: the pinned, hash-verified Ledger bundle (f6fa6152…) — do not substitute unpinned facts. Confirm the base pool H5 (populace-us-2024-cd-concept-budget-dbbdcec-512e-b2500-r2-20260627T022640Z, 224,026 hh) carries the #278 input columns so the #286 gate passes with #279's carry (if the base itself lacks them, the base needs a rebuild first).

Guardrails (unchanged)

Staging only (policyengine/populace-us-staging); never touch policyengine/populace-us, its latest.json, or root files; certification is a separate explicit human decision after the results comparison vs the published baseline (2025: net income $14.50T, tax $3.43T, AGI $16.01T, income_tax $2.15T, poverty 13.7%, citizens 100%, IRA/HSA/childcare inputs absent, OBBBA CTC-SSN Δ −$37.1B vs JCT).

Context: failure analysis on #204 (2026-07-03 comment); Modal runner state in the B/D orchestrator's STATE.md (session scratchpad).

Start with the failure analysis on #204 and the B/D orchestrator STATE.md, then inspect the target_compilation step and the #227 checkpoint and #236 streaming context. Done means the target compilation completes without the observed memory failure, the pinned Ledger bundle and #278 input columns pass the #286 gate, and Build E runs in policyengine/populace-us-staging.