[possibly-used-before-assignment] Model using assert_never (#9716)

Clarify "exact same test" exception

doc/data/messages/p/possibly-used-before-assignment/details.rst

@@ -1,15 +1,49 @@
-If you rely on a pattern like:
+You can use ``assert_never`` to mark exhaustive choices:
+
+.. sourcecode:: python
+
+    from typing import assert_never
+
+    def handle_date_suffix(suffix):
+        if suffix == "d":
+            ...
+        elif suffix == "m":
+            ...
+        elif suffix == "y":
+            ...
+        else:
+            assert_never(suffix)
+
+    if suffix in "dmy":
+        handle_date_suffix(suffix)
+
+Pylint currently allows repeating the same test like this, even though this
+lets some error cases through, as pylint does not assess the intervening code:
 
 .. sourcecode:: python
 
     if guarded():
         var = 1
 
+    # what if code here affects the result of guarded()?
+
     if guarded():
-        print(var)  # emits possibly-used-before-assignment
+        print(var)
+
+But this exception is limited to the repeating the exact same test.
+This warns:
+
+.. sourcecode:: python
+
+    if guarded():
+        var = 1
+
+    if guarded() or other_condition:
+        print(var)  # [possibly-used-before-assignment]
 
-you may be concerned that ``possibly-used-before-assignment`` is not totally useful
-in this instance. However, consider that pylint, as a static analysis tool, does
-not know if ``guarded()`` is deterministic or talks to
-a database. (Likewise, for ``guarded`` instead of ``guarded()``, any other
-part of your program may have changed its value in the meantime.)
+If you find this surprising, consider that pylint, as a static analysis
+tool, does not know if ``guarded()`` is deterministic or talks to
+a database. For constants (e.g. ``guarded`` versus ``guarded()``),
+this is less of an issue, so in this case,
+``possibly-used-before-assignment`` acts more like a future-proofing style
+preference than an error, per se.