Fix: Prevent geometric mean underflow for small values

[sunlight7777777]

Problem

The current implementation of geometric_mean computes the product of all values before taking the nth root. For many small positive values (e.g. [1e-200] * 1000), this causes floating-point underflow to 0.0, producing incorrect results.

Root Cause

Iterative multiplication underflows in IEEE 754 double precision:

product *= v
Solution

Replaced the product-based computation with a numerically stable log-space formulation:

exp(sum(log(v)) / n)

This avoids underflow while preserving mathematical correctness.

Impact
Fixes incorrect outputs for small-value inputs
Improves numerical stability of scoring functions
No breaking changes
Example

Before:

geometric_mean([1e-200] * 1000) → 0.0

After:

→ ~1e-200