Fix `in_nyc` before enforcing no-formula exports

[MaxGhenis]

Summary

in_nyc is formula-owned in PolicyEngine-US and should not be persisted as a final input, but it cannot be fixed by simply dropping the stored column from the current populace artifact. The current artifact has in_nyc and state_fips, but not county / county_str; if the stored in_nyc is removed as-is, the PE-US formula recomputes NYC as zero.

Evidence from the local audit

• Stored in_nyc aggregate in the published US artifact: about 5.94M weighted people.
• One-column removal/recompute: formula aggregate becomes 0 because county data are missing.
• Non-NY rows were not the problem; the issue is the missing county input surface needed to recompute NYC inside NY.

Fix direction

• Do not treat this as an automatic drop.
• Either export the proper county/county_str input needed by the PE-US in_nyc formula, or make an explicit PE-US/data compatibility decision if in_nyc itself is intended to be a source input.
• Add an aggregate regression so removing the stored in_nyc does not collapse NYC to zero.
• Keep this consistent with the no-formula rule from Certified populace-us persists 18 formula-owned variables eCPS doesn't (formula_owned_export gate never wired into the build) #24/Formula-owned export must supersede eCPS parity matching (drop formula-owned vars even when eCPS stores them) #25: final formula-owned outputs should not be persisted unless there is a documented model-side compatibility exception.

Related: #18, #24, #25.

[MaxGhenis]

PE-US-data backlog migration note from #46: keep this tied to the broader export-surface lessons from PolicyEngine/policyengine-us-data#444 and related formula-owned export issues. For in_nyc, the fix should be the upstream geography input surface needed for PE-US to compute the formula, or an explicit compatibility exception; not an eCPS-parity stored-output workaround.

[MaxGhenis]

The failure mode this issue predicted is now live in the certified artifact. Verified on the sparse 57k release (populace.data.load("us", 2024), policyengine-us 1.752-era):

• The artifact stores state_fips but no in_nyc, county, county_str, county_fips, or place_fips — so the stored-in_nyc era ended (per Certified populace-us persists 18 formula-owned variables eCPS doesn't (formula_owned_export gate never wired into the build) #24/Formula-owned export must supersede eCPS parity matching (drop formula-owned vars even when eCPS stores them) #25) without the county input surface this issue said had to land first.
• Recomputed in_nyc is false for every household across repeated fresh simulations (NY-state household share is 3.79%; the true NYC share should be roughly 2.5–3%).
• nyc_income_tax therefore totals $0.0B nationally (actual NYC PIT is roughly $15–16B), and county_fips/place_fips calculate empty for every household.

Repro:

from populace.data import load
from policyengine_us import Microsimulation
sim = Microsimulation(dataset=load("us", 2024))
sim.calculate("in_nyc", 2026).mean()        # 0.0
sim.calculate("nyc_income_tax", 2026).sum() # 0.0

Blast radius beyond NYC PIT: any county/place-conditional variable is a dead branch (local taxes and geography-gated programs), household net income is overstated for would-be NYC households, and downstream consumers cannot scope below state — concretely, axiom-oracles just flipped its US oracle population to populace-us and had to keep the legacy eCPS per-city NYC file as its one remaining eCPS dependency (TheAxiomFoundation/axiom-oracles#74).

Fix direction is unchanged from the issue body, now with a natural home: the clone-and-assign geography spine (#204's local-area architecture; the Belgian commune spine in #263 generalizes the same operator) can assign county within state as spine columns, which restores in_nyc computably and unlocks place-scoped filtering. The aggregate regression this issue asked for ("removing stored in_nyc must not collapse NYC to zero") would have caught this at release time and still seems worth adding to the gates.

[MaxGhenis]

#275 now carries the general fix: county arrives as a rung of the block-anchored geography ladder (block → county via geoid prefix), which restores in_nyc/nyc_income_tax computably. The aggregate regression this issue asked for is in #275's acceptance.