From 9a200aa58d14986b2cc7d11ae4be7dc4f397d543 Mon Sep 17 00:00:00 2001
From: Uyanide <pywang0608@foxmail.com>
Date: Fri, 3 Apr 2026 16:12:02 +0200
Subject: [PATCH] feat: add hard duration filter to avoid fake positive matches

---
 lrx_cli/fetchers/selection.py | 22 ++++++++++++++++++----
 pyproject.toml                |  2 +-
 tests/test_selection.py       | 25 +++++++++++++++++++++++++
 uv.lock                       |  2 +-
 4 files changed, 45 insertions(+), 6 deletions(-)

diff --git a/lrx_cli/fetchers/selection.py b/lrx_cli/fetchers/selection.py
index cb2f76d..7ed99a5 100644
--- a/lrx_cli/fetchers/selection.py
+++ b/lrx_cli/fetchers/selection.py
@@ -80,7 +80,8 @@ def _score_candidate(
       - Title:    40
       - Artist:   30
       - Album:    10
-      - Duration: 10
+      - Duration: 10 (only when reference track has duration; hard mismatch is
+                      pre-filtered before scoring)
     """
     raw = 0.0
     available_weight = 0.0
@@ -115,10 +116,15 @@ def _score_candidate(
                 normalize_for_match(ref_album), normalize_for_match(c.album)
             )
 
-    # Duration
-    if ref_length_ms is not None or c.duration_ms is not None:
+    # Duration — only counted when the reference track has duration.
+    # If the candidate also has duration, it contributes positively when matching
+    # (hard mismatch is already filtered upstream in select_best).
+    # If the candidate lacks duration, it contributes 0 to raw but still counts
+    # in available_weight (penalty for missing verifiable info).
+    # If the reference has no duration, duration is excluded entirely (neutral).
+    if ref_length_ms is not None:
         available_weight += _W_DURATION
-        if ref_length_ms is not None and c.duration_ms is not None:
+        if c.duration_ms is not None:
             diff = abs(c.duration_ms - ref_length_ms)
             if diff <= DURATION_TOLERANCE_MS:
                 raw += _W_DURATION * (1.0 - diff / DURATION_TOLERANCE_MS)
@@ -157,6 +163,14 @@ def select_best(
     best_score = -1.0
 
     for c in candidates:
+        # Hard duration filter: both sides have duration but they don't match → skip.
+        if (
+            track_length_ms is not None
+            and c.duration_ms is not None
+            and abs(c.duration_ms - track_length_ms) > DURATION_TOLERANCE_MS
+        ):
+            continue
+
         s = _score_candidate(c, title, artist, album, track_length_ms)
         if s > best_score:
             best_score = s
diff --git a/pyproject.toml b/pyproject.toml
index 5a8b665..8e1de1e 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "lrx-cli"
-version = "0.3.2"
+version = "0.3.4"
 description = "Fetch line-synced lyrics for your music player."
 readme = "README.md"
 requires-python = ">=3.13"
diff --git a/tests/test_selection.py b/tests/test_selection.py
index 632a14b..90511b8 100644
--- a/tests/test_selection.py
+++ b/tests/test_selection.py
@@ -101,6 +101,31 @@ def test_score_duration_linear_decay() -> None:
     assert score_edge == 0.0
 
 
+def test_duration_hard_filter_rejects_all_mismatched() -> None:
+    """All candidates outside duration tolerance are filtered before scoring."""
+    candidates = [
+        SearchCandidate(
+            item="wrong", duration_ms=180000.0, title="My Love", artist="Westlife"
+        ),
+        SearchCandidate(
+            item="also-wrong", duration_ms=300000.0, title="My Love", artist="Westlife"
+        ),
+    ]
+    best, _ = select_best(candidates, 232000, title="My Love", artist="Westlife")
+    assert best is None
+
+
+def test_duration_neutral_when_ref_has_no_duration() -> None:
+    """Candidate duration does not penalise when the reference has no duration."""
+    # Candidate A: title only (no duration)
+    c_no_dur = SearchCandidate(item="no-dur", title="My Love")
+    # Candidate B: same title + a duration (ref has none)
+    c_with_dur = SearchCandidate(item="with-dur", title="My Love", duration_ms=232000.0)
+    score_no_dur = _score_candidate(c_no_dur, "My Love", None, None, None)
+    score_with_dur = _score_candidate(c_with_dur, "My Love", None, None, None)
+    assert score_no_dur == score_with_dur
+
+
 def test_score_case_insensitive_title() -> None:
     c = SearchCandidate(item="x", title="my love")
     s1 = _score_candidate(c, "My Love", None, None, None)
diff --git a/uv.lock b/uv.lock
index a93895e..9fc4d6d 100644
--- a/uv.lock
+++ b/uv.lock
@@ -153,7 +153,7 @@ wheels = [
 
 [[package]]
 name = "lrx-cli"
-version = "0.3.2"
+version = "0.3.3"
 source = { editable = "." }
 dependencies = [
     { name = "cyclopts" },