registry: add 5 held-out generalization markets (Milk, Oats, KC Wheat, MidCap, Nikkei)

[mspinola]

What

Phase 1 of the held-out generalization workstream: register 5 markets that were never part of the CMR deploy selection, so the strategy can be scored on genuinely out-of-selection data. This is the only source of clean GENERAL evidence for a COT-positioning book whose investable universe it otherwise fully consumed (see docs/analysis_tools.md in npf).

Class	Symbol	Market	CFTC code
Dairy (new class)	DC	Class III Milk	052641
Grains	ZO	Oats	004603
Grains	KE	KC HRW Wheat	001612
Equities	EMD	E-mini S&P MidCap 400	33874A
Equities	NKD	Nikkei 225 ($, CME)	240741

All codes verified against the raw cached CFTC source (dea_fut_xls_2026.zip).

Done in this PR

• registry.yaml: 5 entries (new Dairy class; ZO/KE under Grains; EMD/NKD under Equities → is_equity=True).
• Legacy COT pulled + verified into the store (scoped cftc.update(codes=…), cache-parsed, existing 42 untouched). Comm/NonComm positioning legs present; history: Milk 1997→, Oats/KC-Wheat 1998→, MidCap 2002→, Nikkei 1990→ (1,233–1,688 weeks each).
• Golden registry tests updated for the intentional +5 (count 42→47, Equities 4→6, new Dairy class, identity checksum). 51 passed.

Remaining (Norgate/Windows — you run it)

Prices + contract specs need the Norgate Data Updater:

cotdata-update --prices --metadata --symbols DC ZO KE EMD NKD

Until those land, the CMR strategy can't build frames for these markets (no OHLC).

Notes

• NKD legacy currently ends 2026-03-03 (vs 2026-07-07 for the others) — the $-denominated Nikkei may be winding down at CME. 36y of history regardless; flagging for awareness.
• CotSymbolCodeMap has stale codes for ZR (Rough Rice) and NIY (Nikkei ¥) — those two were held back from this batch pending the code fix.

Next

Phase 2 (cot-analyzer): role: heldout field + collect/plot decoupling + Norgate-Group→class map. Phase 3 (npf): the frozen pooled held-out generalization run.

🤖 Generated with Claude Code