fix: normalize pool depth to ETH for path scoring

[tvinagre]

Summary

• Normalize pool depth from raw token_in units to ETH-equivalent in DepthAndPrice::from_sim_and_derived, using TokenGasPrices already available in DerivedData
• Fix min(depths) comparison across hops — previously meaningless when comparing different token denominations (e.g., WETH 18 decimals vs USDC 6 decimals)
• Guard against zero numerator/denominator in token price

Problem

MostLiquidAlgorithm::try_score_path computes score = product(spot_prices) × min(depths) to prioritize which paths to simulate. Depth was stored in raw token_in units, so a WETH edge depth of 1e18 (1 WETH) appeared 10^12x larger than a USDC depth of 1e9 (1000 USDC). This caused misordered simulation priority, and with max_routes truncation or timeout, genuinely good paths could be permanently excluded.

Changes

Single file change: fynd-core/src/algorithm/most_liquid.rs

Production code:

• from_sim_and_derived: after reading raw BigUint depth, converts to ETH using depth_eth = raw_depth * denominator / numerator from TokenGasPrices
• Returns None (skips edge) if token price is missing or has zero numerator/denominator
• Updated depth field doc comment and computation_requirements() comment

Tests (6 new/updated):

• 3 existing tests updated to provide token prices in DerivedData
• test_from_sim_and_derived_missing_token_price_returns_none
• test_from_sim_and_derived_normalizes_depth_to_eth (integer price ratio)
• test_from_sim_and_derived_normalizes_depth_fractional_price (fractional price ratio)

Benchmark Results (500 requests, fynd-benchmark compare)

Coverage:
  Both found route:          371
  Only main found route:       0
  Only normalized-depth:       0
  Neither found route:       129

Head-to-head (amount out before gas):
  normalized-depth wins:  65  (13.0%)
  main wins:              29  ( 5.8%)
  Ties:                  406  (81.2%)
  Avg diff:           +0.50 bps

Head-to-head (net of gas, server-side):
  normalized-depth wins:  96  (19.2%)
  main wins:              26  ( 5.2%)
  Ties:                  378  (75.6%)
  Avg diff:           +0.67 bps

Solve Time:
  main:             avg=34ms  p50=35  p95=87   p99=103
  normalized-depth: avg=34ms  p50=34  p95=88   p99=102

Route Depth:
  main avg swaps: 1.6  (max 3)
  normalized-depth avg swaps: 1.7  (max 3)
  Identical routes: 415/500

Key takeaways:

• Zero coverage regression — no routes lost from skipping edges without token prices
• 2:1 win ratio pre-gas (65 vs 29), improving to ~4:1 net-of-gas (96 vs 26)
• Top outliers show +18-20 bps on multi-hop routes (WETH → small tokens) where the fix matters most
• Solve time unchanged — normalization adds negligible overhead
• Worst single trade: -11.9 bps; best: +19.9 bps; average consistently positive

Test plan

• All 324 fynd-core tests pass
• Clippy clean (-D warnings)
• Formatting clean
• Benchmark comparison against main (500 requests)

🤖 Generated with Claude Code

[tvinagre]

Please note that I chose to not change how pool_depth was calculated. It's still denominated in sell_token, we just normalize here for this algorithm usage.

Fixing the calculation would change a lot, as it would create another dependency layer (pool_depth needs token_prices). IMO this is fine to be done here - we can move to the derived_data layer if preferred.

[louise-poole]

Two minor comments, otherwise looks good! Interesting that pre-normalising results had some better paths it found... I expected with normalising we'd find the same or better always. What max_routes value did you run these benchmarks with?

[propellerci]

This PR is included in version 0.44.1 🎉