fix(reward): ignore git extended headers and mode lines in diff norma…

src/repo2rlenv/reward.py

@@ -31,6 +31,11 @@
 _FILE_HEADER_RE = re.compile(r"^(?:---|\+\+\+) ")
 _INDEX_LINE_RE = re.compile(r"^index ")
 _DIFF_GIT_RE = re.compile(r"^diff --git ")
+_SIMILARITY_LINE_RE = re.compile(r"^(?:similarity index|dissimilarity index)\b")
+_MODE_OR_RENAME_RE = re.compile(
+    r"^(?:old mode|new mode|new file mode|deleted file mode|copy from|copy to|rename from|rename to)\b"
+)
+_NO_NEWLINE_RE = re.compile(r"^\\ No newline at end of file")
 
 
 @dataclass(slots=True)
@@ -42,22 +47,54 @@ class DiffRewardMetadata:
     parse_error: str | None = None
 
 
-def _normalize_diff(diff: str) -> list[str]:
-    """Strip volatile metadata (hunk line numbers, indices, file headers context)."""
-    lines: list[str] = []
+def _split_diff_sections(diff: str) -> list[list[str]]:
+    sections: list[list[str]] = []
+    current: list[str] = []
     for line in diff.splitlines():
         if _DIFF_GIT_RE.match(line):
-            continue
-        if _INDEX_LINE_RE.match(line):
-            continue
-        if _HUNK_HEADER_RE.match(line):
-            lines.append("@@")  # keep as a separator but drop line numbers
-            continue
-        if _FILE_HEADER_RE.match(line):
-            # Keep filename markers but normalize whitespace
-            lines.append(line.split("\t")[0].strip())
-            continue
-        lines.append(line)
+            if current:
+                sections.append(current)
+            current = [line]
+        else:
+            current.append(line)
+    if current:
+        sections.append(current)
+    return sections
+
+
+def _normalize_diff(diff: str) -> list[str]:
+    """Strip volatile metadata while preserving mode-only and rename-only changes."""
+    lines: list[str] = []
+    for section in _split_diff_sections(diff):
+        has_hunks = any(_HUNK_HEADER_RE.match(line) for line in section)
+        has_file_headers = any(_FILE_HEADER_RE.match(line) for line in section)
+
+        for line in section:
+            if _DIFF_GIT_RE.match(line):
+                # Keep diff --git if there are no ---/+++ markers (e.g. mode-only or rename-only changes)
+                if not has_file_headers:
+                    lines.append(line.strip())
+                continue
+            if _INDEX_LINE_RE.match(line):
+                continue
+            if _SIMILARITY_LINE_RE.match(line):
+                continue
+            if _NO_NEWLINE_RE.match(line):
+                continue
+            if _MODE_OR_RENAME_RE.match(line):
+                # Extended headers are volatile when content hunks are present,
+                # but essential to preserve when there are mode-only or rename-only changes.
+                if not has_hunks:
+                    lines.append(line.strip())
+                continue
+            if _HUNK_HEADER_RE.match(line):
+                lines.append("@@")  # keep as a separator but drop line numbers
+                continue
+            if _FILE_HEADER_RE.match(line):
+                # Keep filename markers but normalize whitespace
+                lines.append(line.split("\t")[0].strip())
+                continue
+            lines.append(line)
     return lines
 
 

tests/test_reward.py

@@ -77,3 +77,87 @@ def test_partial_match_scores_in_between():
     b = SAMPLE_DIFF.replace("hello, world", "goodbye")
     reward, _ = calculate_diff_similarity_reward(a, b)
     assert 0.5 < reward < 1.0
+
+
+def test_normalization_ignores_git_extended_headers_and_mode_changes():
+    """Diffs that differ only in file modes or extended git headers should score 1.0."""
+    oracle = """diff --git a/script.py b/script.py
+new file mode 100755
+index 0000000..abcdef1
+--- /dev/null
++++ b/script.py
+@@ -0,0 +1,2 @@
++#!/usr/bin/env python3
++print("run")
+\\ No newline at end of file
+"""
+    predicted = """diff --git a/script.py b/script.py
+index 0000000..abcdef1
+--- /dev/null
++++ b/script.py
+@@ -0,0 +1,2 @@
++#!/usr/bin/env python3
++print("run")
+"""
+    reward, meta = calculate_diff_similarity_reward(oracle, predicted)
+    assert reward == 1.0
+    assert meta.parse_error is None
+
+
+def test_mode_only_patch_scores_one():
+    """An identical mode-only patch (e.g. 100644 to 100755) should score 1.0."""
+    diff = """diff --git a/script.sh b/script.sh
+old mode 100644
+new mode 100755
+"""
+    reward, meta = calculate_diff_similarity_reward(diff, diff)
+    assert reward == 1.0
+    assert meta.parse_error is None
+
+
+def test_mode_only_patch_mismatch():
+    """A mode-only patch with differing target mode should score less than 1.0."""
+    oracle = """diff --git a/script.sh b/script.sh
+old mode 100644
+new mode 100755
+"""
+    predicted = """diff --git a/script.sh b/script.sh
+old mode 100644
+new mode 100644
+"""
+    reward, _ = calculate_diff_similarity_reward(oracle, predicted)
+    assert reward < 1.0
+
+
+def test_rename_only_patch_scores_one():
+    """An identical rename-only patch should score 1.0."""
+    oracle = """diff --git a/old.py b/new.py
+similarity index 100%
+rename from old.py
+rename to new.py
+"""
+    predicted = """diff --git a/old.py b/new.py
+similarity index 100%
+rename from old.py
+rename to new.py
+"""
+    reward, meta = calculate_diff_similarity_reward(oracle, predicted)
+    assert reward == 1.0
+    assert meta.parse_error is None
+
+
+def test_multi_file_diff_with_mode_only_change():
+    """Diff with both code hunks and mode-only changes should score 1.0 when identical."""
+    diff = """diff --git a/foo.py b/foo.py
+--- a/foo.py
++++ b/foo.py
+@@ -1,2 +1,2 @@
+-print(1)
++print(2)
+diff --git a/run.sh b/run.sh
+old mode 100644
+new mode 100755
+"""
+    reward, meta = calculate_diff_similarity_reward(diff, diff)
+    assert reward == 1.0
+    assert meta.parse_error is None