fix(polymarket_polygon): is_taker_side flag + apply filter in ohlcv to fix volume double-counting

[los-xyz]

Summary

Polymarket's CLOB contracts emit two OrderFilled events per match (one for the maker leg, one for the taker leg), each carrying the full fill amount. Naive SUM(amount) over polymarket_polygon.market_trades therefore over-reports volume by ~2x. This is the double-counting documented by Paradigm in Dec 2025.

This PR:

• market_trades_raw + market_trades: adds a new column is_taker_side BOOLEAN. TRUE on the taker leg of each match. All rows preserved (no row-count change), so users who want the raw OrderFilled stream still have it.
• ohlcv_hourly: filters to is_taker_side = true at the base CTE. volume_usd, volume_contracts, trade_count, vwap halve to Polymarket-canonical values. open / high / low / close unchanged (price is per-trade, symmetric across both legs).

Predicate: taker IN (V1 CTF 0x4bfb...82e, V1 NegRisk 0xc5d5...80a, V2 CTF 0xe111...996b, V2 NegRisk 0xe222...0f59).

Validation

Cell-by-cell match against the Polymarket-authored reference query 6899861 across 17 months (Jan 2025 to May 2026). Deltas are floating-point noise (1e-7 range). Sample:

Month	Reference query (Polymarket)	New is_taker_side filter	Delta
2025-01	$711,931,402.49	$711,931,402.49	0
2025-12	$2,377,961,354.40	$2,377,961,354.40	4.77e-7
2026-03	$4,984,078,339.64	$4,984,078,339.64	-9.54e-7
2026-04	$4,212,805,014.74	$4,212,805,014.74	-1.43e-6
2026-04 notional (shares)	9.0086B	9.0086B	0

The 9.0086B April 2026 notional matches the publicly-cited "Polymarket posted $9 billion in notional terms" figure exactly.

Verification query: 7538562.

Downstream impact

Spell / consumer	Effect
polymarket_polygon.market_trades	New column. Existing columns unchanged. Row count unchanged.
polymarket_polygon.ohlcv_hourly	volume_usd, volume_contracts, trade_count halve. Prices unchanged.
prediction_markets.ohlcv_hourly (curated-data)	Auto-corrects on rebuild. Same shape change.
prediction_markets.trades (curated-data)	Will be updated in a follow-up PR to surface is_taker_side. All Kalshi rows will carry is_taker_side = TRUE (Kalshi has no double-counting) so the filter can be applied uniformly across venues.
External dashboards reading market_trades.amount without filter	Continue to show 2x. Methodology change worth a comms note (Paradigm + Allium + DefiLlama already moved to this convention).

Test plan

• dbt parse --warn-error-options error:all clean locally (matches CI's exact flag set).
• dbt compile clean for market_trades_raw, market_trades, ohlcv_hourly.
• Side-by-side match against Polymarket's reference methodology query 6899861 (17 months, see table above).
• Notional shares cross-check against publicly-stated April 2026 number.
• CI green on this PR.

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Changes core Polymarket volume/trade-count aggregation by filtering to a new is_taker_side flag; downstream dashboards/metrics may shift (~50%) even though trade rows are preserved.

Overview
Fixes Polymarket CLOB volume double-counting by introducing a new boolean is_taker_side on market_trades_raw/market_trades, computed via taker IN (known exchange contract addresses) to identify the canonical taker-leg event per match.

Updates ohlcv_hourly to filter its base trade set to is_taker_side, bringing volume_usd, volume_contracts, and trade_count in line with one-sided (non-duplicated) methodology while leaving price OHLC logic unchanged. Documentation in _schema.yml is updated to describe the new column and recommended filtering.

^{Reviewed by Cursor Bugbot for commit 119aaa1. Configure here.}