feat(#554): multi-window trend engine with percentile comparison and explanation

[amberly-d]

Closes #554

Description

Replaces the spike-based network_condition_adjustment with a multi-window percentile comparison engine. This provides smoother, more grounded fee adjustments by comparing recent fee trends against the full history window.

Changes

• Shared analytics module (packages/core/src/analytics/): trend linear regression, nearest-rank percentile, volatility metrics
• Multi-window comparison: short window (5 min) vs long window (60 min) percentile comparison replaces spike detection
• RecommendationExplanation: Every recommendation now includes explanation data — short/long window medians, percentile shift, adjustment factor, and human-readable reason
• Cold-start fallback: When sample count is below min_sample_count, returns a configurable fallback fee with no explanation
• Bug fix: ETag values now properly quoted (RFC 7231 compliant)
• Devkit clippy fixes: Dead code, manual range contains, manual arithmetic check

Test Results

• 223 tests pass, 0 failures (all pre-existing failures fixed)
• Clippy clean (-D warnings)
• Format clean
• Regression tests verify: rising fees → higher adjustment, falling → lower, stable → no change, cold start → no explanation