fix(extraction): treat the en-dash as a clause boundary too, and pin hyphen non-boundary

src/surreal_memory/extraction/keywords.py

@@ -554,12 +554,13 @@
 )
 
 # Clause boundary: bigrams never pair words that cross one of these.
-# Em-dash is included — it sets off a parenthetical/separate clause the same as a
-# comma (live-proven: "keywords.py — bigramy" otherwise lets the "py" filename
-# fragment glue onto the next clause's first word). Plain ASCII hyphen is
+# Em-dash (—) and en-dash are both included — each sets off a parenthetical /
+# separate clause the same as a comma (live-proven: "keywords.py — bigramy"
+# otherwise lets the "py" filename fragment glue onto the next clause's first
+# word; editors and OSes emit the en-dash just as often). Plain ASCII hyphen is
 # deliberately NOT included: it's mid-word in compounds like "write-gate", which
 # must still yield the useful bigram "write gate".
-_CLAUSE_BOUNDARY = re.compile(r"[.,;:!?\n\r—]+")
+_CLAUSE_BOUNDARY = re.compile(r"[.,;:!?\n\r—–]+")  # noqa: RUF001 (literal en-dash intended)
 
 
 def _get_stop_words(language: str, text: str) -> frozenset[str]:

tests/unit/test_polish_keywords.py

@@ -70,6 +70,25 @@ def test_no_bigram_across_em_dash(self) -> None:
         assert "py" in unigrams
         assert "bigramy" in unigrams
 
+    def test_no_bigram_across_en_dash(self) -> None:
+        # The en-dash (U+2013) is a clause boundary too, not only the em-dash
+        # (—, U+2014): editors and OSes emit it for ranges and asides, so without
+        # it the same cross-clause junk bigram (a filename fragment glued onto the
+        # next clause) slips through.
+        text = "blad w keywords.py – bigramy nie powinny przecinac granic"  # noqa: RUF001
+        bigrams = _bigrams(text)
+        assert "py bigramy" not in bigrams
+        unigrams = _unigrams(text)
+        assert "py" in unigrams
+        assert "bigramy" in unigrams
+
+    def test_hyphen_is_not_a_clause_boundary(self) -> None:
+        # A hyphen inside a compound term must NOT split a clause: the two halves
+        # are adjacent content words and should still pair as a bigram (contrast
+        # with the en-/em-dash, which do split).
+        bigrams = _bigrams("modul cache-aside dziala szybko")
+        assert "cache aside" in bigrams
+
     def test_phase1_junk_bigrams_eliminated(self) -> None:
         junk = {
             "reguly czy",