Visit GST grow branches in descending score order (BOSS/GRaSP)

[sanghack81]

Summary

Fixes the grow-branch sort direction in causallearn/search/PermutationBased/gst.py so the Grow-Shrink Tree visits higher-scoring grow branches first, matching TETRAD's GrowShrinkTree (discussed in #264).

Root cause

BOSS uses a higher-is-better score. In TETRAD's edu.cmu.tetrad.search.utils.GrowShrinkTree, GSTNode.compareTo() orders by ascending growScore, but grow() sorts with this.branches.sort(Collections.reverseOrder()), so branches are visited in descending growScore order.

In causal-learn, GSTNode.__lt__ is ascending and grow() called self.branches.sort() — i.e. ascending, the opposite direction.

Why it matters (not just presentation order)

GST.trace descends into the first prefix-matching branch and returns its shrink score, so the sort direction changes the returned score and parent set for a given prefix. With the deterministic mock scores from #264 and prefix [1, 2]:

ascending  -> score 2.0837, parents {1}
descending -> score 4.3568, parents {2}

Visiting the best improving add first follows the more promising path — a better local approximation that also prunes more of the tree.

Change

# grow()
self.branches.sort(reverse=True)

I kept __lt__ ascending and reversed at the sort site, mirroring TETRAD's compareTo + Collections.reverseOrder() (equivalent to inverting the comparator, the other form proposed in #264). This GST is shared by both boss and grasp, so both now use the descending (TETRAD-style) traversal.

Scope

This PR changes only the branch traversal order. Other existing behavior is intentionally left as-is: grow branches are still included on strict improvement (score > self.grow_score; TETRAD uses >=), and background knowledge (forbidden/required) remains unimplemented as before. Aligning those with TETRAD is out of scope here.

Performance (from #264; synthetic linear-Gaussian, BIC-from-cov, avg_degree=2, lambda=2)

p=50, n=1000:

seed	current ascending	TETRAD-style descending	speedup	CPDAG edges
101	1.376s	0.269s	5.12x	48 / 48
102	0.806s	0.236s	3.42x	44 / 44
103	1.543s	0.315s	4.91x	46 / 46
104	0.714s	0.209s	3.42x	40 / 40

Mean wall time 1.110s → 0.257s (4.32x).

p=100, n=1000:

seed	current ascending	TETRAD-style descending	speedup	CPDAG edges
101	3.603s	1.107s	3.25x	86 / 86
102	10.686s	1.816s	5.89x	111 / 110
103	13.717s	1.753s	7.82x	108 / 108
104	6.377s	1.344s	4.75x	101 / 102

Mean wall time 8.596s → 1.505s (5.71x). As noted in #264, in some seeds the returned CPDAG edge count also changes.

Test

Adds tests/TestGST.py, a focused, deterministic regression test over the #264 mock scores: it asserts the root grow branches are ordered by descending score and that tracing prefix [1, 2] selects the higher-scoring parent {2}. The test fails on the previous ascending behavior and passes after this change.

Closes #264

[kunwuz]

Thanks so much for your contribution!