ECOD aggregates as pyod does, not as the paper's Algorithm 1

Problem

pp.outliers(method="ecod") sums over features the per-feature maximum of the three tail scores, as pyod does. Algorithm 1 of Li et al. (2023) takes the maximum of the three per-row sums instead. The two orders flag different cells. The _core/_ecod.py docstring said the opposite ("the reverse order gives plausible scores that are not ECOD"); that wording is being corrected under #35.

Reproduction

import numpy as np
from scipy.stats import skew
from mantispy._core._ecod import column_ecdf, ecod_scores

X = np.random.default_rng(0).standard_t(3, size=(500, 20))
U_l, U_r = -np.log(column_ecdf(X)), -np.log(column_ecdf(-X))
U_skew = np.where(skew(X, axis=0) < 0, U_l, U_r)
sum_of_max = np.maximum(np.maximum(U_l, U_r), U_skew).sum(axis=1)                          # pyod
max_of_sum = np.maximum(np.maximum(U_l.sum(axis=1), U_r.sum(axis=1)), U_skew.sum(axis=1))  # paper
np.allclose(ecod_scores(X), sum_of_max)  # True
top = lambda s: set(np.argsort(s)[-25:])
len(top(sum_of_max) & top(max_of_sum))   # 7 of 25

Expected

A deliberate choice between pyod's aggregation and the paper's, recorded in the pp.outliers docstring so that users know which ECOD they get.

Read mantispy/_core/_ecod.py and the pp.outliers docstring, then run the reproduction using column_ecdf and ecod_scores. Determine which aggregation is intended, record that deliberate choice in the pp.outliers documentation, and make sure the documented behavior matches the selected ECOD scores.