Looking for feedback on narrow backtrack causes approach

src/resolvelib/resolvers.py

@@ -267,6 +267,34 @@ def _attempt_to_pin_criterion(self, name):
         # end, signal for backtracking.
         return causes
 
+    def _narrow_causes(self, causes):
+        """Return a narrowed causes list
+        """
+        # Causes are often duplicates, first dedup them
+        deduped_causes = list({id(c): c for c in causes}.values())
+
+        # For each cause check if it actually contradicts with another cause
+        # and put them both in "narrowed causes", or otherwise disregard it
+        narrowed_causes = []
+        while deduped_causes:
+            cause = deduped_causes.pop()
+            for i, alternative_cause in enumerate(deduped_causes):
+                if cause.requirement.name != alternative_cause.requirement.name:
+                    continue
+
+                specifier = (
+                    alternative_cause.requirement.get_candidate_lookup()[1].specifier
+                )
+                alternative_specifier = (
+                    alternative_cause.requirement.get_candidate_lookup()[1].specifier
+                )
+                specifier_intersection = specifier and alternative_specifier
+                if not str(specifier_intersection):
+                    narrowed_causes.append(cause)
+                    narrowed_causes.append(deduped_causes.pop(i))
+
+        return narrowed_causes
+
     def _backjump(self, causes):
         """Perform backjumping.
 
@@ -315,7 +343,7 @@ def _backjump(self, causes):
                     broken_state = self._states.pop()
                     name, candidate = broken_state.mapping.popitem()
                 except (IndexError, KeyError):
-                    raise ResolutionImpossible(causes) from None
+                    raise ResolutionImpossible(causes)
                 current_dependencies = {
                     self._p.identify(d)
                     for d in self._p.get_dependencies(candidate)
@@ -432,7 +460,7 @@ def resolve(self, requirements, max_rounds):
                 # an unpinned state, so we can work on it in the next round.
                 self._r.resolving_conflicts(causes=causes)
                 success = self._backjump(causes)
-                self.state.backtrack_causes[:] = causes
+                self.state.backtrack_causes[:] = self._narrow_causes(causes)
 
                 # Dead ends everywhere. Give up.
                 if not success: