fix: use log-space computation in geometric_mean to prevent float underflow

[Grizouforever]

Summary

• Fixes Bug: geometric_mean underflows to 0.0 with many small positive values #346 — geometric_mean returned 0.0 incorrectly when iterative multiplication of many small positive values underflowed in IEEE 754
• Replaces the product loop with a numerically stable log-space computation: exp((1/n) * Σ log(vᵢ))
• Preserves all existing behavior: zero/negative inputs still return 0.0, empty list returns 0.0, the smoothed-epsilon path is unchanged

Root Cause

# Before — underflows to 0.0 for e.g. [1e-200] * 1000
product = 1.0
for v in values:
    product *= v
return product ** (1.0 / n)

IEEE 754 doubles underflow to 0.0 when the running product drops below ~5e-324. With 1000 values of 1e-200, the true product is 1e-200000, which is far below that floor.

Fix

# After — numerically stable across full positive float range
try:
    log_mean = sum(math.log(v) for v in values) / n
    return math.exp(log_mean)
except (ValueError, OverflowError):
    return 0.0

Test plan

• geometric_mean([2.0, 8.0]) still returns 4.0
• geometric_mean([1e-200] * 1000) returns ~1e-200 instead of 0.0
• geometric_mean([0.5, 0.0, 0.5]) still returns 0.0
• geometric_mean([]) still returns 0.0
• Existing test suite passes (pre-existing import errors in tests/ are unrelated to this module)

🤖 Generated with Claude Code