fix(sdk): normalize read slices after windowing (#3888)

Changes `slice_read_response` to split first and normalize line endings
only across the requested read window.

This avoids rewriting the full file contents for paginated reads while
preserving LF-normalized output for downstream edit and grep behavior.

### Benchmarks

Focused local microbenchmarks compared the original combined branch
against `main` on Python 3.14.2:

- `slice_read_response` on large CRLF content: **2.55x faster** for both
first-window and late-window reads (~60.8% improvement), with **~29.2%
lower peak memory**.

libs/deepagents/deepagents/backends/utils.py

@@ -291,15 +291,12 @@ def slice_read_response(
     if not content or content.strip() == "":
         return content
 
-    # Normalize line endings to LF before slicing. State/Store backends may
-    # carry CRLF or CR content as written; downstream tooling (edit match,
-    # grep, format) assumes LF.
-    content = content.replace("\r\n", "\n").replace("\r", "\n")
-
     # `splitlines(keepends=True)` retains each line's terminator, including
     # the absence of one on the final line. Joining with `""` therefore
     # round-trips the trailing-newline state of the file faithfully —
-    # required so `edit()` can report EOF-newline mismatches accurately.
+    # required so `edit()` can report EOF-newline mismatches accurately. It
+    # also splits on CR / CRLF, so line indexing matches the LF-normalized
+    # form without first rewriting the whole (potentially huge) string.
     lines = content.splitlines(keepends=True)
     start_idx = offset
     end_idx = min(start_idx + limit, len(lines))
@@ -307,7 +304,10 @@ def slice_read_response(
     if start_idx >= len(lines):
         return ReadResult(error=f"Line offset {offset} exceeds file length ({len(lines)} lines)")
 
-    return "".join(lines[start_idx:end_idx])
+    # Normalize line endings to LF, but only across the requested window.
+    # State/Store backends may carry CRLF or CR content as written;
+    # downstream tooling (edit match, grep, format) assumes LF.
+    return "".join(lines[start_idx:end_idx]).replace("\r\n", "\n").replace("\r", "\n")
 
 
 def perform_string_replacement(

libs/deepagents/tests/unit_tests/backends/test_utils.py

@@ -343,6 +343,12 @@ class TestSliceReadResponse:
         result = slice_read_response(self._file("a\nb\nc\nd\n"), offset=1, limit=2)
         assert result == "b\nc\n"
 
+    def test_partial_window_normalizes_crlf(self) -> None:
+        """An internal CRLF slice is LF-normalized even though only the window is rewritten."""
+        result = slice_read_response(self._file("a\r\nb\r\nc\r\nd\r\n"), offset=1, limit=2)
+        assert result == "b\nc\n"
+        assert "\r" not in result
+
     def test_partial_window_ending_on_unterminated_last_line(self) -> None:
         """A window covering the last line keeps that line's missing-terminator state."""
         result = slice_read_response(self._file("a\nb\nc"), offset=2, limit=1)