diff --git a/Cargo.lock b/Cargo.lock
index 452a556..c91df54 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1871,7 +1871,7 @@ dependencies = [
 
 [[package]]
 name = "tui-textarea-2"
-version = "0.12.1"
+version = "0.12.2"
 dependencies = [
  "arbitrary",
  "crossterm 0.28.1",
diff --git a/src/screen_map.rs b/src/screen_map.rs
index 58cdee8..ea17ee3 100644
--- a/src/screen_map.rs
+++ b/src/screen_map.rs
@@ -225,18 +225,40 @@ mod tests {
     }
 
     fn assert_round_trips(textarea: &TextArea<'_>) {
+        // Word-wrap drops separator whitespace that never fits a row (matching ratatui).
+        // Such positions have no screen representation, so a detached coordinate
+        // round-trip is impossible for them — only the attached (explicit dc) round-trip
+        // is asserted, plus stability of the detached mapping.
+        let screen_lines = textarea.screen_lines.borrow().clone();
         for (row, line) in textarea.lines.iter().enumerate() {
+            let rows_for_line: Vec<WrappedLine> = screen_lines
+                .iter()
+                .filter(|sl| sl.wrapped.row == row)
+                .map(|sl| sl.wrapped)
+                .collect();
             for col in 0..=line.chars().count() {
                 let dc = DataCursor(row, col);
                 let sc = textarea.data_to_screen_cursor(dc);
                 assert_eq!(textarea.screen_to_data_cursor(sc), dc);
 
                 let detached = ScreenCursor { dc: None, ..sc };
-                assert_eq!(textarea.screen_to_data_cursor(detached), dc);
-                assert_eq!(
-                    textarea.data_to_screen_cursor(textarea.screen_to_data_cursor(detached)),
-                    sc
-                );
+                let back = textarea.screen_to_data_cursor(detached);
+                let visible = rows_for_line.iter().any(|w| {
+                    let (lo, hi) = if w.last_in_row {
+                        (w.start_col, w.end_col)
+                    } else {
+                        (w.start_col, w.end_col.saturating_sub(1))
+                    };
+                    lo <= col && col <= hi
+                });
+                if visible {
+                    assert_eq!(back, dc, "detached round-trip failed for {dc:?}");
+                    assert_eq!(textarea.data_to_screen_cursor(back), sc);
+                } else {
+                    // Position inside dropped whitespace: the detached mapping resolves
+                    // to the nearest representable position and must be stable.
+                    assert_eq!(textarea.data_to_screen_cursor(back), sc);
+                }
             }
         }
     }
diff --git a/src/wrap.rs b/src/wrap.rs
index 1557c81..a9a15b7 100644
--- a/src/wrap.rs
+++ b/src/wrap.rs
@@ -57,8 +57,12 @@ pub(crate) fn wrapped_rows(
 
     for (row, line) in lines.iter().enumerate() {
         let ranges = line_ranges(line, mode, width, tab_len);
-        let mut start_col = 0usize;
         for (i, (start_byte, end_byte)) in ranges.iter().copied().enumerate() {
+            // Column bookkeeping is byte-accurate: `start_col` is the data column of the
+            // fragment's first char and `end_col` the column just past its last char.
+            // Whitespace that word-wrap dropped (it never fits a row) is not claimed by
+            // any row, so positions inside it have no screen representation.
+            let start_col = line[..start_byte].chars().count();
             let end_col = start_col + line[start_byte..end_byte].chars().count();
             rows.push(WrappedLine {
                 row,
@@ -69,7 +73,6 @@ pub(crate) fn wrapped_rows(
                 first_in_row: i == 0,
                 last_in_row: i + 1 == ranges.len(),
             });
-            start_col = end_col;
         }
     }
 
@@ -143,6 +146,54 @@ fn wrap_word_chunks(
 
         if seg_end > seg_start {
             out.push((seg_start, seg_end));
+
+            // Separator whitespace that does not fit on the current segment is
+            // handled like ratatui's WordWrapper (trim: false): when the line
+            // is exactly full the first whitespace grapheme is dropped,
+            // otherwise whitespace filling the remaining space is dropped;
+            // any leftover whitespace leads the next row. This prevents
+            // whitespace-only phantom rows (e.g. "abcde fghij" at width 5 must
+            // wrap to 2 rows, not 3).
+            let text = chunk_text(line, chunk);
+            if !text.is_empty() && text.chars().all(char::is_whitespace) {
+                let drop_cells = if seg_width == width {
+                    1
+                } else {
+                    width - seg_width
+                };
+                let kept_start = drop_leading_whitespace(line, chunk, drop_cells, tab_len);
+                if kept_start >= chunk.end {
+                    // Entire separator dropped: start the next row fresh.
+                    seg_start = chunk.end;
+                    seg_end = chunk.end;
+                    seg_width = 0;
+                    i += 1;
+                    continue;
+                }
+                // Leftover whitespace starts the next row.
+                let kept_width = display_width_from(&line[kept_start..chunk.end], 0, tab_len);
+                if kept_width <= width {
+                    seg_start = kept_start;
+                    seg_end = chunk.end;
+                    seg_width = kept_width;
+                    i += 1;
+                    continue;
+                }
+                // Pathological: leftover whitespace wider than the row.
+                if fallback_to_glyph {
+                    split_range_by_grapheme_width(
+                        line, kept_start, chunk.end, width, tab_len, &mut out,
+                    );
+                } else {
+                    out.push((kept_start, chunk.end));
+                }
+                i += 1;
+                seg_start = chunk.end;
+                seg_end = chunk.end;
+                seg_width = 0;
+                continue;
+            }
+
             seg_start = seg_end;
             seg_width = 0;
             continue;
@@ -217,6 +268,29 @@ fn chunk_text(line: &str, chunk: Chunk) -> &str {
     &line[chunk.start..chunk.end]
 }
 
+/// Return the byte offset inside `chunk` after dropping up to `drop_cells`
+/// display cells of leading whitespace (grapheme-atomic). Used to trim
+/// separator whitespace at line breaks like ratatui's `WordWrapper` does.
+fn drop_leading_whitespace(line: &str, chunk: Chunk, drop_cells: usize, tab_len: u8) -> usize {
+    let text = &line[chunk.start..chunk.end];
+    let mut col = 0usize;
+    let mut cells = 0usize;
+    let mut offset = 0usize;
+    for (grapheme_offset, grapheme) in UnicodeSegmentation::grapheme_indices(text, true) {
+        let width = display_width_to(grapheme, col, tab_len).saturating_sub(col);
+        if width == 0 {
+            continue;
+        }
+        if cells + width > drop_cells {
+            break;
+        }
+        cells += width;
+        col += width;
+        offset = grapheme_offset + grapheme.len();
+    }
+    chunk.start + offset
+}
+
 fn display_width_from(text: &str, start_width: usize, tab_len: u8) -> usize {
     display_width_to(text, start_width, tab_len).saturating_sub(start_width)
 }
@@ -276,4 +350,211 @@ mod tests {
         let have = segments("\tX", WrapMode::WordOrGlyph, 2);
         assert_eq!(have, vec!["\t", "X"]);
     }
+
+    #[test]
+    fn separator_whitespace_dropped_when_line_exactly_full() {
+        // One space after an exactly full word must not produce a phantom row.
+        let have = segments("1234567890 ", WrapMode::WordOrGlyph, 10);
+        assert_eq!(have, vec!["1234567890"]);
+        let have = segments("abcde fghij", WrapMode::WordOrGlyph, 5);
+        assert_eq!(have, vec!["abcde", "fghij"]);
+        let have = segments("abcde fghij", WrapMode::Word, 5);
+        assert_eq!(have, vec!["abcde", "fghij"]);
+        let have = segments("abcde fghi", WrapMode::WordOrGlyph, 5);
+        assert_eq!(have, vec!["abcde", "fghi"]);
+    }
+
+    #[test]
+    fn separator_whitespace_overflowing_remaining_space_is_dropped() {
+        // "  " after a 4/5 segment: 1 cell fits the remaining space and is
+        // kept (leading the next row), the second cell is dropped — no
+        // phantom whitespace-only row.
+        let have = segments("abcd  efgh", WrapMode::WordOrGlyph, 5);
+        assert_eq!(have, vec!["abcd", " efgh"]);
+        let have = segments("abcd  efgh", WrapMode::Word, 5);
+        assert_eq!(have, vec!["abcd", " efgh"]);
+    }
+
+    #[test]
+    fn leftover_separator_whitespace_leads_next_row() {
+        // After an exactly full word the first separator grapheme is dropped,
+        // the rest stays visible at the start of the next row.
+        let have = segments("1234567890  ", WrapMode::WordOrGlyph, 10);
+        assert_eq!(have, vec!["1234567890", " "]);
+        let have = segments("abcde  fghij", WrapMode::WordOrGlyph, 5);
+        assert_eq!(have, vec!["abcde", " ", "fghij"]);
+    }
+
+    #[test]
+    fn wrap_matches_ratatui_row_counts() {
+        // Row counts must match ratatui's WordWrapper (trim: false) — the same
+        // algorithm the TUI renders with. `ratatui_reference_rows` is a faithful
+        // port of ratatui's WordWrapper::process_input; `ratatui_buffer_rows`
+        // cross-checks it against the real Paragraph renderer for inputs that do
+        // not end in a whitespace-only row (those are indistinguishable from
+        // empty buffer space).
+        let cases: &[(&str, usize)] = &[
+            ("1234567890", 10),
+            ("1234567890 ", 10),
+            ("1234567890 x", 10),
+            ("1234567890  ", 10),
+            ("1234567890   ", 10),
+            ("hello world", 10),
+            ("hello worldx", 10),
+            ("  padded  text  ", 8),
+            ("abcde fghij", 5),
+            ("abcde fghi", 5),
+            ("abcde fgh", 5),
+            ("abcde  fghij", 5),
+            ("abcd  efgh", 5),
+            ("abcd    efgh", 5),
+            ("a b c d e f g h i j k l", 5),
+            ("ab cd ef gh", 5),
+            ("x y z", 2),
+            ("word word word", 8),
+            ("12345678901234567890", 10),
+            ("123456789012345678901", 10),
+            ("第二行 mixed-width 内容 with English words.", 18),
+            ("ASCII and русский текст can share one buffer.", 18),
+            (
+                "Lorem ipsum dolor sit amet, consectetur adipiscing elit.",
+                24,
+            ),
+        ];
+        for (text, width) in cases {
+            let ours = segments(text, WrapMode::WordOrGlyph, *width).len();
+            let reference = ratatui_reference_rows(text, *width);
+            assert_eq!(
+                ours, reference,
+                "row count mismatch for {text:?} at width {width}: fork={ours} \
+                 ratatui(reference)={reference}"
+            );
+            if !text.ends_with(|c: char| c.is_whitespace()) {
+                // Buffer rendering cannot expose whitespace-only trailing rows.
+                let buffer_rows = ratatui_buffer_rows(text, *width);
+                assert_eq!(
+                    reference, buffer_rows,
+                    "reference diverges from real ratatui render for {text:?} at {width}"
+                );
+            }
+        }
+    }
+
+    /// Faithful port of ratatui's `WordWrapper::process_input` (trim: false).
+    /// Returns the number of wrapped lines ratatui would render.
+    fn ratatui_reference_rows(text: &str, width: usize) -> usize {
+        let width = width as u16;
+        let mut wrapped: Vec<Vec<String>> = Vec::new();
+        let mut line: Vec<String> = Vec::new();
+        let mut line_width = 0u16;
+        let mut pending_word: Vec<String> = Vec::new();
+        let mut word_width = 0u16;
+        let mut pending_ws: Vec<String> = Vec::new();
+        let mut ws_width = 0u16;
+        let mut non_ws_prev = true;
+
+        let symbols: Vec<(String, u16)> =
+            unicode_segmentation::UnicodeSegmentation::graphemes(text, true)
+                .map(|s| {
+                    (
+                        s.to_string(),
+                        unicode_width::UnicodeWidthStr::width(s) as u16,
+                    )
+                })
+                .collect();
+        for (symbol, symbol_width) in symbols {
+            let is_whitespace = symbol.trim().is_empty() && symbol_width > 0;
+            if symbol_width > width {
+                continue;
+            }
+            let word_found = non_ws_prev && is_whitespace;
+            let trimmed_overflow = line.is_empty() && word_width + symbol_width > width;
+            let whitespace_overflow = line.is_empty() && ws_width + symbol_width > width;
+            let untrimmed_overflow =
+                line.is_empty() && word_width + ws_width + symbol_width > width;
+            if word_found || trimmed_overflow || whitespace_overflow || untrimmed_overflow {
+                // Port is used with trim: false only, where ratatui always extends
+                // pending whitespace into the line (`!line.is_empty() || !trim`).
+                line.append(&mut pending_ws);
+                line_width += ws_width;
+                line.append(&mut pending_word);
+                line_width += word_width;
+                pending_ws.clear();
+                ws_width = 0;
+                word_width = 0;
+            }
+            let line_full = line_width >= width;
+            let pending_word_overflow =
+                symbol_width > 0 && line_width + ws_width + word_width >= width;
+            if line_full || pending_word_overflow {
+                let mut remaining = width.saturating_sub(line_width);
+                if !line.is_empty() {
+                    wrapped.push(std::mem::take(&mut line));
+                }
+                line_width = 0;
+                while let Some(first) = pending_ws.first() {
+                    let w = unicode_width::UnicodeWidthStr::width(first.as_str()) as u16;
+                    if w > remaining {
+                        break;
+                    }
+                    ws_width -= w;
+                    remaining -= w;
+                    pending_ws.remove(0);
+                }
+                if is_whitespace && pending_ws.is_empty() {
+                    non_ws_prev = !is_whitespace;
+                    continue;
+                }
+            }
+            if is_whitespace {
+                ws_width += symbol_width;
+                pending_ws.push(symbol);
+            } else {
+                word_width += symbol_width;
+                pending_word.push(symbol);
+            }
+            non_ws_prev = !is_whitespace;
+        }
+        // End of input: ratatui appends pending whitespace (and the pending word)
+        // to the line even when trim is false and the line is empty.
+        line.append(&mut pending_ws);
+        line.append(&mut pending_word);
+        if !line.is_empty() {
+            wrapped.push(line);
+        }
+        wrapped.len().max(1)
+    }
+
+    fn ratatui_buffer_rows(text: &str, width: usize) -> usize {
+        use ratatui::Terminal;
+        use ratatui::backend::TestBackend;
+        use ratatui::text::Text;
+        use ratatui::widgets::{Paragraph, Wrap};
+
+        let backend = TestBackend::new(width as u16, 100);
+        let mut terminal = Terminal::new(backend).expect("terminal should initialize");
+        terminal
+            .draw(|f| {
+                f.render_widget(
+                    Paragraph::new(Text::from(text.to_string())).wrap(Wrap { trim: false }),
+                    f.area(),
+                );
+            })
+            .expect("draw should succeed");
+        let buffer = terminal.backend().buffer().clone();
+        let mut last_non_blank = None;
+        for y in 0..buffer.area.height {
+            let mut line = String::new();
+            for x in 0..buffer.area.width {
+                line.push_str(buffer[(x, y)].symbol());
+            }
+            if !line.trim().is_empty() {
+                last_non_blank = Some(y);
+            }
+        }
+        match last_non_blank {
+            Some(y) => y as usize + 1,
+            None => 1,
+        }
+    }
 }