feat(trade): extend Trade FBS/overlay to 2018–2024; export 1:m uses Detail_Supply T007

[bl-young]

cc:
Closes: #727
Related: #729

What changed? Why?

Nowcast years 2018-2024 use the same Trade FBS + overlay path as 2017. Extract YAMLs already declare Census_USATrade / BEA_IEA through 2024; methods, MDTY/MADJ helpers, and nowcast gates were 2017-only.

Export 1:m Crosswalk splits use same-year domestic commodity output T007: each Trade_Exports_<year> caches Detail_Supply_<year> and runs trade.attribution.collapse_detail_supply_to_t007 before proportional attribution (#729). Imports stay on frozen 2017 Supply MCIF.

Methods

• Trade_Exports_common.yaml / Trade_Imports_common.yaml hold shared bodies; thin Trade_{Exports,Imports}_{2017..2024}.yaml stubs set year.
• Export stubs cache Detail_Supply_<year> with clean_fbs: collapse_detail_supply_to_t007 and use that cache key as attribution_source.
• Census and IEA sources inherit the method year. Exports and imports remain separate methods (different FlowName, IEA Crosswalk, attribution column).

Helpers + nowcast wire

• mdty_detail_usd / madj_detail_usd work for every overlay year (no year != 2017 guard).
• MADJ level: 2017 ->’ published Supply MADJ; later years ->’ that year's mapped GEN_CHA_YR magnitude with the published Supply MADJ sign. Destination shares stay frozen 2017 Supply MADJ.
• TRADE_OVERLAY_YEARS = range(2017, 2025) in eeio/nowcast.py for F04000 / MCIF / MDTY / MADJ. Inventories stay 2017-only.
• S00900 / F04000 identity keeps frozen 2017 Supply T016[S00900] for every trade-overlay year.
• Overlay stays unscaled mapped Trade mass (Step 1B: revisit wholesale ITA G+S scale vs raw Census + BEA IEA #647 Option B).

bedrock/transform/trade/README.md documents the multi-year layout and attribution rule. score_2017_trade_detail_baseline.csv is advanced for T007 export-attribution leaf moves (49 commodities; no Crosswalk change).

GCS extract inputs (staged; not in git)

Raw Census / IEA dumps for 2018-2024 are on gs://cornerstone-default/extract/input-data/:

• Census_USATrade/{2018..2024}/Census_USATrade_<year>_{exports,imports}.csv
• BEA_IEA/{2018..2024}/BEA_IEA_<year>_{Exports,Imports}.csv

Detail_Supply_{2018..2024} FBS parquets live under gs://cornerstone-default/transform/output_data/ (not flowsa/FlowBySector/ -> download locally before generate). With extract_data_from_raw_sources: False, FBA generation loads Census/IEA via gcs_fxn without API keys.

Outstanding: Census NAICS 2022+ -> Detail Crosswalk

Sector_Crosswalk_Census_USATrade is still a 2017-vintage Activity list. Smoke vs Census_USATrade FBAs:

Year	Export mass join	Import mass join	Notes
2017-2021	~100% / ~97%	~100% / ~97%	Import gap almost entirely intentional 980000
2022	~99.7%	~97%	Small dig-6 orphans + residual rename 11211X->1121XX; OK to ship
2023-2024	~92%	~82%	NAICS-2022 leaves (e.g. 336110) drop; FBS Detail row count 305->290

This PR does not extend the Crosswalk. Follow-on under #670 / goods-allocation: inventory orphan Activities by year -> concordance to BEA 2017 Detail -> append Crosswalk rows (keep 980000 omitted) -> regen Trade_*_{2022..2024}. Cheap first cut: alias 1121XX to the same Detail targets as 11211X (~0.5 B exports).

Out of scope

• Full NAICS-2022 Crosswalk extension -> see Outstanding.
• Import total-uses attribution (Step 1B: Trade 1:m attribution — commodity output (exports) and total uses (imports) #729 Phase C).

Testing

Regenerated Trade_Exports/Imports_{2017..2024} locally (Census/IEA FBAs local; Detail_Supply_{2018..2024} pulled from GCS transform/output_data/).

$env:Path = "$env:USERPROFILE\.local\bin;$env:Path"
foreach ($y in 2017..2024) {
  uv run python -c "from bedrock.transform.flowbysector import FlowBySector as F; e=F.generateFlowBySector(f'Trade_Exports_$y', download_sources_ok=True); i=F.generateFlowBySector(f'Trade_Imports_$y', download_sources_ok=True); print($y, e['FlowAmount'].sum()/1e6, i['FlowAmount'].sum()/1e6)"
}
uv run python -c "from bedrock.transform.eeio.nowcast import derive_initial_Y_pur, derive_initial_supply_bridge; y=derive_initial_Y_pur(2022); b=derive_initial_supply_bridge(2022); print(y['F04000'].sum()/1e6, b['MCIF'].sum()/1e6, b['MDTY'].sum()/1e6, b['MADJ'].sum()/1e6)"
uv run python -m bedrock.analysis.nowcasting.trade_data.score_2017_trade_detail

Results (local):

⚠️ Vintage note. These totals were generated before Step 4a rewired
Detail_Supply and commodity output for 2022–2024. Export 1:m splits use same-year
T007 as proportional attribution weights, which redistribute within a 1:m group
and preserve its total — so the totals below still hold, but the per-commodity
composition for 2022–2024 exports moves once Trade_Exports_{2022..2024} is
regenerated against the updated commodity output. Imports are unaffected: they stay
on frozen 2017 Supply MCIF.

Year	Exports M	Imports M
2017	2,132,637	2,679,578
2018	2,271,469	2,870,643
2019	2,278,811	2,835,013
2020	2,037,942	2,665,519
2021	2,443,060	3,209,061
2022	2,810,816	3,723,030
2023	2,643,809	3,092,093
2024	2,763,107	3,309,118

2017 scorecard (after T007; baseline updated): F040 national +12.20%, MCIF +1.14%; Pearson non-special 0.92 / 0.85; no regressions vs prior baseline (49 leaf moves). Re-run without --update-baseline should report no changes.