Add advanced signal tracking and injury/matchup context to backtest engine

[GodingWal]

Summary

This PR enhances the backtest engine with new signal types and improved context building by adding injury tracking, matchup history extraction, defender matchup detection, and CLV tracking capabilities.

Key Changes

Frontend (client/src/pages/backtest.tsx)

• Added 6 new signal display configurations:
  • clv_tracker: Closing line value tracking (purple)
  • defender_matchup: Primary defender impact (teal)
  • matchup_history: Head-to-head performance (orange)
  • line_movement: Sharp money detection (cyan)
  • fatigue: Schedule & travel load (red)
  • referee: Referee tendency adjustment (pink)

Backend (server/nba-prop-model/src/evaluation/backtest_engine.py)

• Data Loading: Added two new data loaders:

  • _load_injuries(): Loads injury history from database to track OUT/doubtful players by team and date
  • _load_team_rosters(): Loads player rosters with positions for defender matching

• Context Building: Enhanced _build_context() with new features:

  • Injury alpha: Identifies injured teammates affecting player opportunity
  • Matchup history: Extracts head-to-head game history vs specific opponents
  • CLV tracking: Computes model direction (OVER/UNDER) by comparing baseline to line
  • Blowout risk: Estimates game spread from team defensive ratings
  • Defender matchup: Identifies primary defender on opponent team

• Helper Methods:

  • _find_injured_teammates(): Locates injured players on a team for a given date
  • _find_primary_defender(): Matches player position with known elite/weak defenders
  • _estimate_spread(): Calculates estimated spread using defensive rating differential
  • _extract_vs_team_history(): Parses recent_games JSONB to find head-to-head matchups

• Improved B2B Detection: Enhanced _detect_b2b() with fallback logic:

  • Primary: Database query for games on previous day
  • Fallback: Checks recent_games JSONB data if DB unavailable

Implementation Details

• Injury and roster data are cached in memory during initialization for performance
• Context building gracefully handles missing data and falls back to heuristics
• Defender matching uses position categories (G/F/C) for flexible matching
• Spread estimation incorporates home court advantage (~3 points)
• All new methods include error handling and logging for debugging

https://claude.ai/code/session_01RXgC7zzvzoDMkf7nJjecaR