Address PR #401 R5 review (1 P2 transactional set_params)

R5 verdict was "Looks good" with one P2 code-quality regression
flagged. The previous round's `test_set_params_partial_mutation_recoverable`
documented inconsistent intermediate state as expected behavior:
on a mutex violation, `set_params()` mutated `self` before
validation, so a failed call left both `by_path` and
`paths_of_interest` populated. `fit()` did not re-check the mutex
and `_enumerate_treatment_paths()` silently preferred
`paths_of_interest` when both were set, so subsequent fits ran
with stale selector state unless the caller manually repaired the
object.

Fix: make `set_params()` transactional. Snapshot the pre-call
values for the keys being set, attempt mutation + validation, and
roll back to the snapshot on any exception before re-raising. This
mirrors sklearn-style atomic semantics. The validation block was
extracted into a private `_validate_invariants()` helper so
__init__ and set_params can share the rules without code
duplication risking drift.

Replaces `test_set_params_partial_mutation_recoverable` (which
asserted the buggy partial-mutation behavior) with
`test_set_params_failed_validation_is_transactional`, asserting
that:
1. After a failed `set_params()`, `get_params()` returns
   bit-identical to the pre-call snapshot.
2. Subsequent valid `set_params()` succeeds against the rolled-
   back state.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/chaisemartin_dhaultfoeuille.py

@@ -779,14 +779,30 @@ def set_params(self, **params: Any) -> "ChaisemartinDHaultfoeuille":
         """
         Set estimator parameters (sklearn-compatible).
 
-        Re-runs the same validation rules as ``__init__`` so invalid
-        parameter combinations cannot be introduced after construction.
+        **Transactional**: validation runs after the candidate mutations,
+        and if any rule fails the estimator state is rolled back to its
+        pre-call values before the exception is re-raised. Callers can
+        therefore retry with corrected params on the same instance
+        without repairing inconsistent intermediate state.
         """
-        for key, value in params.items():
+        # Snapshot current values for the keys we are about to set so
+        # we can roll back on validation failure (transactional semantics).
+        for key in params:
             if not hasattr(self, key):
                 raise ValueError(f"Unknown parameter: {key}")
-            setattr(self, key, value)
+        snapshot = {key: getattr(self, key) for key in params}
+        try:
+            for key, value in params.items():
+                setattr(self, key, value)
+            self._validate_invariants()
+        except Exception:
+            for key, value in snapshot.items():
+                setattr(self, key, value)
+            raise
+        return self
 
+    def _validate_invariants(self) -> None:
+        """Run the post-mutation validation rules. Mirrors `__init__`."""
         # Re-run __init__ validation rules so the post-set state is valid.
         if self.rank_deficient_action not in ("warn", "error", "silent"):
             raise ValueError(
@@ -834,7 +850,6 @@ def set_params(self, **params: Any) -> "ChaisemartinDHaultfoeuille":
                 f"clustering is reserved for a future phase. See REGISTRY.md "
                 f"ChaisemartinDHaultfoeuille section for the full contract."
             )
-        return self
 
     # ------------------------------------------------------------------
     # fit

tests/test_chaisemartin_dhaultfoeuille.py

@@ -8519,18 +8519,28 @@ def test_set_params_re_validates_mutex_poi_added(self):
         with pytest.raises(ValueError, match="mutually exclusive"):
             est.set_params(paths_of_interest=[(0, 1, 1, 1)])
 
-    def test_set_params_partial_mutation_recoverable(self):
-        """After mutex set_params raises mid-mutation, recovery is possible."""
+    def test_set_params_failed_validation_is_transactional(self):
+        """A failed `set_params()` must leave estimator state unchanged
+        (regression for R5 P2 finding: prior implementation mutated
+        before validation, leaving both selectors populated when the
+        mutex check raised, which `fit()` then silently consumed)."""
         est = ChaisemartinDHaultfoeuille(paths_of_interest=[(0, 1, 1, 1)])
-        # set_params mutates self before validation; mutex raise leaves both set
+        # Capture pre-call state.
+        before = est.get_params()
+        # Mutex violation: by_path AND paths_of_interest both non-None.
         with pytest.raises(ValueError, match="mutually exclusive"):
             est.set_params(by_path=2)
-        # User can recover by clearing one of them.
-        est.set_params(by_path=None)
-        # Instance now valid; subsequent calls don't raise.
+        # Post-failure state is rolled back to pre-call.
+        after = est.get_params()
+        assert after == before, (
+            f"set_params() rollback failed: by_path={after['by_path']}, "
+            f"paths_of_interest={after['paths_of_interest']}"
+        )
+        # Subsequent valid set_params() succeeds against rolled-back state.
+        est.set_params(by_path=2, paths_of_interest=None)
         params = est.get_params()
-        assert params["by_path"] is None
-        assert params["paths_of_interest"] == [(0, 1, 1, 1)]
+        assert params["by_path"] == 2
+        assert params["paths_of_interest"] is None
 
     def test_get_params_includes_paths_of_interest(self):
         est = ChaisemartinDHaultfoeuille(