feat: full restructuring of venue-experts

[yk4r2]

venue-expert restructuring done — 31 files/9K lines → 48/12.4K. 50 research questions integrated from primary sources. what matters:

• fixed cancel rate 5%→20-40%, every queue model was wrong by 4-8x
• cross-venue matching matrix (30 product-venue combos), now agents know CME SR3=Allocation, Chinese exchanges=no native modify
• session overlap matrix with arb windows pre-computed (DCE-SGX 8hrs, SHFE-COMEX 12.5hrs, INE-ICE 11.5hrs)
• data quirks table: 17 gotchas preventing silent pipeline bugs (CME implied depth=2 not 10, CZCE timestamps=0, ICE at Basildon not
  LD4)
• CME deep doc from scratch (442 lines), no docs existed for world's biggest futures venue
• 16 factual errors fixed (CFFEX sessions, SPAN formula, DCE-SGX overlap)
• regime change DBs: US 2005-2025, China 2022-2025 with backtesting impact
• 6/9 research agents went from score 1-2 to 7-8 on venue questions
• ~10% over-invested (GFEX got CME-level treatment for 5 illiquid products). gaps remain: LME, options, fee tables