test(02-02): add failing reading-override tests

- test_hitoke_override, test_override_table_is_well_formed, test_apply_overrides_is_idempotent

tests/test_units.py

@@ -288,3 +288,76 @@ def test_build_units_is_pure_over_dicts():
     assert build_units(ms) == units  # deterministic
     assert build_units([]) == []
     assert all(type(u) is dict for u in units)
+
+
+# --- Task 3: the committed reading-override table -------------------------------------------
+
+
+def test_hitoke_override():
+    from japanese_avatar.nlp.overrides import OVERRIDES_PATH, apply_overrides
+
+    assert OVERRIDES_PATH.name == "reading_overrides.json"
+    assert OVERRIDES_PATH.is_file()
+
+    units = units_of("人気のない通り")
+    assert surfaces(units) == ["人気", "の", "ない", "通り"]
+    assert units[0]["reading"] == "にんき"  # what Sudachi core gives (measured)
+    before = [dict(u) for u in units]
+
+    replaced = apply_overrides(units)
+    assert replaced == 1
+    assert units[0]["reading"] == "ひとけ"
+    # Only the reading of the matched unit changes; everything else is untouched.
+    assert {k: v for k, v in units[0].items() if k != "reading"} == {
+        k: v for k, v in before[0].items() if k != "reading"
+    }
+    assert units[1:] == before[1:]
+
+    units = units_of("人気がある")
+    assert apply_overrides(units) == 0
+    assert units[0]["reading"] == "にんき"
+
+    units = units_of("この店は人気だ")
+    assert apply_overrides(units) == 0
+    assert [u["reading"] for u in units if u["surface"] == "人気"] == ["にんき"]
+
+    # 人気 at the very end of the text has no following unit: next_surface cannot match.
+    units = units_of("人気")
+    assert apply_overrides(units) == 0
+    assert units[0]["reading"] == "にんき"
+
+
+def test_override_table_is_well_formed():
+    import json
+    import re
+
+    from japanese_avatar.nlp.overrides import OVERRIDES_PATH, load_overrides
+
+    rows = load_overrides()
+    assert isinstance(rows, tuple) and rows
+    hiragana = re.compile(r"^[ぁ-ゖー]+$")
+    for row in rows:
+        assert set(row) == {"surface", "reading_from_sudachi", "next_surface", "reading", "why"}
+        assert hiragana.match(row["reading"]), row
+        assert hiragana.match(row["reading_from_sudachi"]), row
+        assert row["reading"] != row["reading_from_sudachi"], row
+        assert row["next_surface"] is None or isinstance(row["next_surface"], str)
+        assert isinstance(row["why"], str) and row["why"].strip(), row
+    assert rows[0]["surface"] == "人気"
+    assert rows[0]["reading"] == "ひとけ"
+    assert rows[0]["next_surface"] == "の"
+    # The committed file is what load_overrides serves, and it is plain JSON.
+    raw = json.loads(OVERRIDES_PATH.read_text(encoding="utf-8"))
+    assert tuple(raw) == rows
+    assert load_overrides() is rows  # cached: loaded once
+
+
+def test_apply_overrides_is_idempotent():
+    from japanese_avatar.nlp.overrides import apply_overrides
+
+    once = units_of("人気のない通り")
+    assert apply_overrides(once) == 1
+    twice = [dict(u) for u in once]
+    assert apply_overrides(twice) == 0
+    assert twice == once
+    assert apply_overrides([]) == 0