[Repo Assist] feat(gcm): add random_seed to interventional_samples, counterfactual_samples, average_causal_effect

[github-actions]

🤖 This PR was created by Repo Assist, an automated AI assistant.

Problem

interventional_samples(), counterfactual_samples(), and average_causal_effect() in dowhy/gcm/whatif.py have no random_seed parameter, making reproducible results across independent Python sessions unnecessarily fragile. Users must currently call set_random_seed() globally before calling these functions — which is order-sensitive, easy to miss, and inconsistent with other GCM functions.

Reported in #1307 (multiple users experiencing non-reproducible results across sessions). A previous fix PR #1489 was closed as stale in July 2026 before being merged.

Change

Add an optional random_seed: Optional[int] = None parameter to:

• interventional_samples()
• counterfactual_samples()
• average_causal_effect()

When provided, set_random_seed(random_seed) is called at function entry, following the existing GCM convention used in falsify_graph, model_evaluation, etc.

# Before — fragile global seed required
from dowhy.gcm.util.general import set_random_seed
set_random_seed(42)
samples = gcm.interventional_samples(model, {"X": lambda x: 1}, num_samples_to_draw=500)

# After — explicit, self-contained
samples = gcm.interventional_samples(model, {"X": lambda x: 1}, num_samples_to_draw=500, random_seed=42)

Trade-offs

• Backwards compatible: parameter is optional with default None (existing call sites unaffected).
• Global side effect: set_random_seed affects numpy/torch global state, consistent with the rest of the GCM module's approach.

Test Status

Added 3 new unit tests in tests/gcm/test_whatif.py:

• test_given_random_seed_when_draw_interventional_samples_then_results_are_reproducible — same seed → identical samples; different seeds → different samples.
• test_given_random_seed_when_draw_counterfactual_samples_then_results_are_reproducible — same seed → identical counterfactual frames.
• test_given_random_seed_when_estimate_average_causal_effect_then_results_are_reproducible — same seed → identical ACE; different seeds → different ACE.

19 passed in 9.04s

All 19 tests in tests/gcm/test_whatif.py pass. Format/lint: black and isort show no changes; 3 pre-existing E501 lint warnings in the file (unrelated to this PR).

Closes #1307

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76