tracepipe 0.2.0__py3-none-any.whl → 0.3.1__py3-none-any.whl

tracepipe/instrumentation/indexer_capture.py

@@ -0,0 +1,813 @@
+# tracepipe/instrumentation/indexer_capture.py
+"""
+loc/iloc instrumentation for TracePipe.
+
+Operations tracked:
+| Pattern                      | Type       | Completeness |
+|------------------------------|------------|--------------|
+| df.loc[mask]                 | Filter     | FULL         |
+| df.loc[mask, 'col']          | Filter     | FULL         |
+| df.iloc[0:5]                 | Filter     | FULL         |
+| df.iloc[[1,3,5]]             | Filter     | FULL         |
+| df.loc[mask, 'col'] = val    | Transform  | FULL         |
+| df.iloc[0:5, 0] = val        | Transform  | FULL         |
+| df.loc[mask] = other_df      | Transform  | PARTIAL      |
+
+Key insight: We wrap the indexer's __getitem__ and __setitem__, not DataFrame.loc itself.
+"""
+
+import warnings
+from typing import Any, Optional, Union
+
+import numpy as np
+import pandas as pd
+
+from ..context import get_context
+from ..core import ChangeType, CompletenessLevel
+from ..safety import TracePipeWarning, get_caller_info
+
+
+class _CallableLocIndexer:
+    """
+    Wrapper for the indexer returned by df.loc(axis=...).
+
+    This is used internally by pandas methods like dropna.
+    We skip tracking when _filter_op_depth > 0 to avoid double-counting
+    since the outer operation (e.g., dropna) will track the drops.
+    """
+
+    def __init__(self, indexer, parent_df: pd.DataFrame):
+        self._indexer = indexer
+        self._parent_df = parent_df
+
+    def __getitem__(self, key):
+        """Pass through to underlying indexer and propagate RIDs."""
+        ctx = get_context()
+        result = self._indexer[key]
+
+        if not ctx.enabled:
+            return result
+
+        # Skip step tracking if we're inside a filter operation (parent will track)
+        # But still propagate RIDs for the result
+        if isinstance(result, pd.DataFrame):
+            try:
+                row_mgr = ctx.row_manager
+                source_df = self._parent_df
+
+                source_rids = row_mgr.get_ids_array(source_df)
+                if source_rids is None:
+                    row_mgr.register(source_df)
+                    source_rids = row_mgr.get_ids_array(source_df)
+
+                # Propagate RIDs to result
+                if hasattr(key, "dtype") and key.dtype == bool:
+                    mask = key.values if isinstance(key, pd.Series) else np.asarray(key)
+                    row_mgr.propagate_by_mask(source_df, result, mask)
+                else:
+                    row_mgr.register(result)
+
+                # Only create step if NOT inside a filter operation
+                if ctx._filter_op_depth == 0:
+                    store = ctx.store
+                    if hasattr(key, "dtype") and key.dtype == bool:
+                        mask = key.values if isinstance(key, pd.Series) else np.asarray(key)
+                        dropped_rids, _ = row_mgr.compute_dropped_with_positions(source_rids, mask)
+                        completeness = CompletenessLevel.FULL
+                    else:
+                        result_rids = row_mgr.get_ids_array(result)
+                        dropped_rids = row_mgr.compute_dropped_ids(source_rids, result_rids)
+                        completeness = CompletenessLevel.PARTIAL
+
+                    code_file, code_line = get_caller_info(skip_frames=4)
+                    step_id = store.append_step(
+                        operation="DataFrame.loc(axis)[]",
+                        stage=ctx.current_stage,
+                        code_file=code_file,
+                        code_line=code_line,
+                        params={},
+                        input_shape=source_df.shape,
+                        output_shape=result.shape,
+                        completeness=completeness,
+                    )
+
+                    if len(dropped_rids) > 0:
+                        store.append_bulk_drops(step_id, dropped_rids)
+            except Exception as e:
+                if ctx.config.strict_mode:
+                    raise
+                warnings.warn(f"TracePipe: loc(axis)[] capture failed: {e}", TracePipeWarning)
+
+        return result
+
+    def __setitem__(self, key, value):
+        """Pass through setitem."""
+        self._indexer[key] = value
+
+
+class TrackedLocIndexer:
+    """
+    Wrapper around pandas _LocIndexer that captures lineage.
+
+    Usage (internal - user never sees this):
+        df.loc  # Returns TrackedLocIndexer wrapping the real _LocIndexer
+    """
+
+    def __init__(self, indexer, parent_df: pd.DataFrame):
+        self._indexer = indexer
+        self._parent_df = parent_df
+
+    def __getattr__(self, name):
+        """Proxy any other attribute access to the underlying indexer."""
+        return getattr(self._indexer, name)
+
+    def __call__(self, axis=None):
+        """
+        Support for df.loc(axis=...) callable form used internally by pandas.
+
+        This is used by dropna and other methods that call self.loc(axis=axis)[mask].
+        """
+        # Return a callable-aware indexer that wraps the result of calling the original
+        return _CallableLocIndexer(self._indexer(axis), self._parent_df)
+
+    def __getitem__(self, key) -> Union[pd.DataFrame, pd.Series, Any]:
+        """
+        Capture filter operations via loc[].
+
+        Handles:
+        - df.loc[mask] -> DataFrame (filter)
+        - df.loc[mask, 'col'] -> Series (filter + column select)
+        - df.loc[mask, ['a', 'b']] -> DataFrame (filter + column select)
+        - df.loc['label'] -> Row (single row access)
+        """
+        ctx = get_context()
+
+        # Always run original first
+        result = self._indexer[key]
+
+        if not ctx.enabled:
+            return result
+
+        try:
+            self._capture_loc_getitem(key, result, ctx)
+        except Exception as e:
+            if ctx.config.strict_mode:
+                raise
+            warnings.warn(f"TracePipe: loc[] capture failed: {e}", TracePipeWarning)
+
+        return result
+
+    def __setitem__(self, key, value) -> None:
+        """
+        Capture transform operations via loc[] = value.
+
+        Handles:
+        - df.loc[mask, 'col'] = scalar
+        - df.loc[mask, 'col'] = array
+        - df.loc[mask, ['a', 'b']] = values
+        - df.loc[mask] = other_df (PARTIAL - complex assignment)
+        """
+        ctx = get_context()
+
+        # Capture before state for watched columns
+        before_values = None
+        affected_cols = None
+        if ctx.enabled and ctx.watched_columns:
+            before_values, affected_cols = self._capture_before_state(key, ctx)
+
+        # Always run original
+        self._indexer[key] = value
+
+        if not ctx.enabled:
+            return
+
+        try:
+            self._capture_loc_setitem(key, value, before_values, affected_cols, ctx)
+        except Exception as e:
+            if ctx.config.strict_mode:
+                raise
+            warnings.warn(f"TracePipe: loc[]= capture failed: {e}", TracePipeWarning)
+
+    def _capture_loc_getitem(self, key, result, ctx) -> None:
+        """Capture filter via loc[]."""
+        if not isinstance(result, pd.DataFrame):
+            return  # Series or scalar - not a filter operation
+
+        row_mgr = ctx.row_manager
+        source_df = self._parent_df
+
+        source_rids = row_mgr.get_ids_array(source_df)
+        if source_rids is None:
+            row_mgr.register(source_df)
+            source_rids = row_mgr.get_ids_array(source_df)
+
+        # Derive kept mask from key
+        kept_mask, completeness = self._derive_loc_mask(key, source_df)
+
+        # Always propagate RIDs
+        if kept_mask is not None:
+            row_mgr.propagate_by_mask(source_df, result, kept_mask)
+        else:
+            row_mgr.register(result)
+
+        # Skip step tracking if we're inside a filter operation (parent will track)
+        if ctx._filter_op_depth > 0:
+            return
+
+        store = ctx.store
+
+        if kept_mask is not None:
+            dropped_rids, _ = row_mgr.compute_dropped_with_positions(source_rids, kept_mask)
+        else:
+            completeness = CompletenessLevel.PARTIAL
+            result_rids = row_mgr.get_ids_array(result)
+            dropped_rids = row_mgr.compute_dropped_ids(source_rids, result_rids)
+
+        code_file, code_line = get_caller_info(skip_frames=4)
+        step_id = store.append_step(
+            operation="DataFrame.loc[]",
+            stage=ctx.current_stage,
+            code_file=code_file,
+            code_line=code_line,
+            params={"key_type": type(key).__name__},
+            input_shape=source_df.shape,
+            output_shape=result.shape,
+            completeness=completeness,
+        )
+
+        if len(dropped_rids) > 0:
+            store.append_bulk_drops(step_id, dropped_rids)
+
+    def _derive_loc_mask(
+        self, key, df: pd.DataFrame
+    ) -> tuple[Optional[np.ndarray], CompletenessLevel]:
+        """
+        Derive boolean mask from loc key.
+
+        Key types:
+        - Boolean array/Series -> mask directly
+        - Slice -> convert to positional mask
+        - List of labels -> index.isin()
+        - Single label -> single row mask
+        - Tuple (row_key, col_key) -> handle row_key
+        """
+        row_key = key[0] if isinstance(key, tuple) else key
+
+        # Boolean mask
+        if hasattr(row_key, "dtype") and row_key.dtype == bool:
+            mask = row_key.values if isinstance(row_key, pd.Series) else np.asarray(row_key)
+            if len(mask) == len(df):
+                return mask, CompletenessLevel.FULL
+
+        # List of labels
+        if isinstance(row_key, (list, np.ndarray)) and not (
+            hasattr(row_key, "dtype") and row_key.dtype == bool
+        ):
+            mask = df.index.isin(row_key)
+            return (
+                mask.to_numpy() if hasattr(mask, "to_numpy") else np.asarray(mask),
+                CompletenessLevel.FULL,
+            )
+
+        # Slice
+        if isinstance(row_key, slice):
+            try:
+                # Get positional indices from label slice
+                start_idx = df.index.get_loc(row_key.start) if row_key.start is not None else 0
+                stop_idx = df.index.get_loc(row_key.stop) if row_key.stop is not None else len(df)
+                # loc slice is inclusive on both ends
+                if isinstance(start_idx, int) and isinstance(stop_idx, int):
+                    mask = np.zeros(len(df), dtype=bool)
+                    mask[start_idx : stop_idx + 1] = True
+                    return mask, CompletenessLevel.FULL
+            except (KeyError, TypeError):
+                pass
+
+        # Single label
+        if not isinstance(row_key, (list, np.ndarray, slice, pd.Series)):
+            try:
+                mask = df.index == row_key
+                return (
+                    mask.to_numpy() if hasattr(mask, "to_numpy") else np.asarray(mask),
+                    CompletenessLevel.FULL,
+                )
+            except Exception:
+                pass
+
+        return None, CompletenessLevel.PARTIAL
+
+    def _capture_before_state(self, key, ctx) -> tuple[Optional[dict], Optional[list]]:
+        """Capture values before assignment for watched columns."""
+        col_key = key[1] if isinstance(key, tuple) and len(key) > 1 else None
+
+        # Determine affected columns
+        if col_key is None:
+            affected_cols = list(ctx.watched_columns & set(self._parent_df.columns))
+        elif isinstance(col_key, str):
+            affected_cols = [col_key] if col_key in ctx.watched_columns else []
+        elif isinstance(col_key, list):
+            affected_cols = [c for c in col_key if c in ctx.watched_columns]
+        else:
+            affected_cols = []
+
+        if not affected_cols:
+            return None, None
+
+        # Derive affected rows
+        mask, _ = self._derive_loc_mask(key, self._parent_df)
+        if mask is None:
+            return None, affected_cols
+
+        # Capture values (vectorized per column)
+        rids = ctx.row_manager.get_ids_array(self._parent_df)
+        if rids is None:
+            return None, affected_cols
+
+        before = {}
+        affected_positions = np.where(mask)[0]
+        for col in affected_cols:
+            before[col] = {
+                "rids": rids[affected_positions].copy(),
+                "values": self._parent_df[col].values[affected_positions].copy(),
+            }
+
+        return before, affected_cols
+
+    def _capture_loc_setitem(self, key, value, before_values, affected_cols, ctx) -> None:
+        """Capture transform via loc[] = value."""
+        if before_values is None or not affected_cols:
+            return
+
+        store = ctx.store
+
+        code_file, code_line = get_caller_info(skip_frames=4)
+        step_id = store.append_step(
+            operation="DataFrame.loc[]=",
+            stage=ctx.current_stage,
+            code_file=code_file,
+            code_line=code_line,
+            params={"columns": affected_cols[:3]},
+            input_shape=self._parent_df.shape,
+            output_shape=self._parent_df.shape,
+            completeness=CompletenessLevel.FULL,
+        )
+
+        from ..utils.value_capture import values_equal
+
+        for col in affected_cols:
+            if col not in before_values:
+                continue
+
+            rids = before_values[col]["rids"]
+            old_vals = before_values[col]["values"]
+
+            # Get current positions for these rids
+            mask, _ = self._derive_loc_mask(key, self._parent_df)
+            if mask is None:
+                continue
+
+            new_vals = self._parent_df[col].values[np.where(mask)[0]]
+
+            # Vectorized diff detection
+            for rid, old_val, new_val in zip(rids, old_vals, new_vals):
+                if not values_equal(old_val, new_val):
+                    store.append_diff(
+                        step_id=step_id,
+                        row_id=int(rid),
+                        col=col,
+                        old_val=old_val,
+                        new_val=new_val,
+                        change_type=ChangeType.MODIFIED,
+                    )
+
+
+class TrackedILocIndexer:
+    """
+    Wrapper around pandas _iLocIndexer that captures lineage.
+
+    Similar to TrackedLocIndexer but uses positional indexing.
+    """
+
+    def __init__(self, indexer, parent_df: pd.DataFrame):
+        self._indexer = indexer
+        self._parent_df = parent_df
+
+    def __getattr__(self, name):
+        """Proxy any other attribute access to the underlying indexer."""
+        return getattr(self._indexer, name)
+
+    def __getitem__(self, key) -> Union[pd.DataFrame, pd.Series, Any]:
+        """Capture filter via iloc[]."""
+        ctx = get_context()
+        result = self._indexer[key]
+
+        if not ctx.enabled:
+            return result
+
+        try:
+            self._capture_iloc_getitem(key, result, ctx)
+        except Exception as e:
+            if ctx.config.strict_mode:
+                raise
+            warnings.warn(f"TracePipe: iloc[] capture failed: {e}", TracePipeWarning)
+
+        return result
+
+    def __setitem__(self, key, value) -> None:
+        """Capture transform via iloc[] = value."""
+        ctx = get_context()
+
+        before_values = None
+        affected_cols = None
+        if ctx.enabled and ctx.watched_columns:
+            before_values, affected_cols = self._capture_before_state(key, ctx)
+
+        self._indexer[key] = value
+
+        if not ctx.enabled:
+            return
+
+        try:
+            self._capture_iloc_setitem(key, value, before_values, affected_cols, ctx)
+        except Exception as e:
+            if ctx.config.strict_mode:
+                raise
+            warnings.warn(f"TracePipe: iloc[]= capture failed: {e}", TracePipeWarning)
+
+    def _capture_iloc_getitem(self, key, result, ctx) -> None:
+        """Capture filter via iloc[]."""
+        if not isinstance(result, pd.DataFrame):
+            return
+
+        row_mgr = ctx.row_manager
+        source_df = self._parent_df
+
+        source_rids = row_mgr.get_ids_array(source_df)
+        if source_rids is None:
+            row_mgr.register(source_df)
+            source_rids = row_mgr.get_ids_array(source_df)
+
+        # Derive positions from key
+        positions = self._derive_iloc_positions(key, source_df)
+
+        # Always propagate RIDs
+        if positions is not None:
+            row_mgr.propagate_by_positions(source_df, result, positions)
+        else:
+            row_mgr.register(result)
+
+        # Skip step tracking if we're inside a filter operation (parent will track)
+        if ctx._filter_op_depth > 0:
+            return
+
+        store = ctx.store
+
+        if positions is not None:
+            kept_mask = np.zeros(len(source_df), dtype=bool)
+            kept_mask[positions] = True
+            dropped_rids, _ = row_mgr.compute_dropped_with_positions(source_rids, kept_mask)
+            completeness = CompletenessLevel.FULL
+        else:
+            completeness = CompletenessLevel.PARTIAL
+            result_rids = row_mgr.get_ids_array(result)
+            dropped_rids = row_mgr.compute_dropped_ids(source_rids, result_rids)
+
+        code_file, code_line = get_caller_info(skip_frames=4)
+        step_id = store.append_step(
+            operation="DataFrame.iloc[]",
+            stage=ctx.current_stage,
+            code_file=code_file,
+            code_line=code_line,
+            params={"key_type": type(key).__name__},
+            input_shape=source_df.shape,
+            output_shape=result.shape,
+            completeness=completeness,
+        )
+
+        if len(dropped_rids) > 0:
+            store.append_bulk_drops(step_id, dropped_rids)
+
+    def _derive_iloc_positions(self, key, df: pd.DataFrame) -> Optional[np.ndarray]:
+        """Derive position array from iloc key."""
+        row_key = key[0] if isinstance(key, tuple) else key
+        n = len(df)
+
+        # Integer
+        if isinstance(row_key, int):
+            pos = row_key if row_key >= 0 else n + row_key
+            return np.array([pos], dtype=np.int64)
+
+        # Slice
+        if isinstance(row_key, slice):
+            indices = range(*row_key.indices(n))
+            return np.array(list(indices), dtype=np.int64)
+
+        # List/array of integers
+        if isinstance(row_key, (list, np.ndarray)):
+            arr = np.asarray(row_key, dtype=np.int64)
+            # Handle negative indices
+            arr = np.where(arr < 0, n + arr, arr)
+            return arr
+
+        # Boolean array
+        if hasattr(row_key, "dtype") and row_key.dtype == bool:
+            return np.where(row_key)[0].astype(np.int64)
+
+        return None
+
+    def _capture_before_state(self, key, ctx):
+        """Capture values before assignment for watched columns."""
+        col_key = key[1] if isinstance(key, tuple) and len(key) > 1 else None
+
+        # Determine affected columns by position
+        if col_key is None:
+            affected_cols = list(ctx.watched_columns & set(self._parent_df.columns))
+        elif isinstance(col_key, int):
+            col_name = self._parent_df.columns[col_key]
+            affected_cols = [col_name] if col_name in ctx.watched_columns else []
+        elif isinstance(col_key, (list, np.ndarray)):
+            col_names = [self._parent_df.columns[i] for i in col_key]
+            affected_cols = [c for c in col_names if c in ctx.watched_columns]
+        elif isinstance(col_key, slice):
+            col_names = self._parent_df.columns[col_key].tolist()
+            affected_cols = [c for c in col_names if c in ctx.watched_columns]
+        else:
+            affected_cols = []
+
+        if not affected_cols:
+            return None, None
+
+        positions = self._derive_iloc_positions(key, self._parent_df)
+        if positions is None:
+            return None, affected_cols
+
+        rids = ctx.row_manager.get_ids_array(self._parent_df)
+        if rids is None:
+            return None, affected_cols
+
+        before = {}
+        for col in affected_cols:
+            before[col] = {
+                "rids": rids[positions].copy(),
+                "values": self._parent_df[col].values[positions].copy(),
+                "positions": positions.copy(),
+            }
+
+        return before, affected_cols
+
+    def _capture_iloc_setitem(self, key, value, before_values, affected_cols, ctx) -> None:
+        """Capture transform via iloc[] = value."""
+        if before_values is None or not affected_cols:
+            return
+
+        store = ctx.store
+
+        code_file, code_line = get_caller_info(skip_frames=4)
+        step_id = store.append_step(
+            operation="DataFrame.iloc[]=",
+            stage=ctx.current_stage,
+            code_file=code_file,
+            code_line=code_line,
+            params={"columns": affected_cols[:3]},
+            input_shape=self._parent_df.shape,
+            output_shape=self._parent_df.shape,
+            completeness=CompletenessLevel.FULL,
+        )
+
+        from ..utils.value_capture import values_equal
+
+        for col in affected_cols:
+            if col not in before_values:
+                continue
+
+            rids = before_values[col]["rids"]
+            old_vals = before_values[col]["values"]
+            positions = before_values[col]["positions"]
+            new_vals = self._parent_df[col].values[positions]
+
+            for rid, old_val, new_val in zip(rids, old_vals, new_vals):
+                if not values_equal(old_val, new_val):
+                    store.append_diff(
+                        step_id=step_id,
+                        row_id=int(rid),
+                        col=col,
+                        old_val=old_val,
+                        new_val=new_val,
+                        change_type=ChangeType.MODIFIED,
+                    )
+
+
+# Store original properties for restore
+_original_loc = None
+_original_iloc = None
+_original_at = None
+_original_iat = None
+
+
+class TrackedAtIndexer:
+    """
+    Wrapper around pandas _AtIndexer that captures scalar assignments.
+
+    .at is optimized for scalar access by label.
+    """
+
+    def __init__(self, indexer, df):
+        self._indexer = indexer
+        self._df = df
+
+    def __getitem__(self, key):
+        return self._indexer[key]
+
+    def __setitem__(self, key, value) -> None:
+        """Capture scalar assignment via at[row, col] = value."""
+        ctx = get_context()
+        if not ctx or not ctx.enabled:
+            self._indexer[key] = value
+            return
+
+        row_label, col = key
+        col_str = str(col)
+
+        # Check if column is watched
+        should_track = col_str in ctx.watched_columns if ctx.watched_columns else False
+
+        # Capture before state
+        old_val = None
+        if should_track:
+            try:
+                old_val = self._df.at[row_label, col]
+            except (KeyError, IndexError):
+                pass
+
+        # Execute original
+        self._indexer[key] = value
+
+        # Capture after state
+        if should_track and ctx.store:
+            from .pandas_inst import get_caller_info
+
+            code_file, code_line = get_caller_info(skip_frames=2)
+            step_id = ctx.store.append_step(
+                operation="DataFrame.at[]=",
+                stage=ctx.current_stage,
+                code_file=code_file,
+                code_line=code_line,
+                params={"row": str(row_label), "col": col_str},
+                input_shape=self._df.shape,
+                output_shape=self._df.shape,
+            )
+
+            # Get row_id for this position
+            try:
+                row_pos = self._df.index.get_loc(row_label)
+                if isinstance(row_pos, int):
+                    rids = ctx.row_manager.get_ids_array(self._df)
+                    if rids is None:
+                        ctx.row_manager.register(self._df)
+                        rids = ctx.row_manager.get_ids_array(self._df)
+                    if rids is not None and row_pos < len(rids):
+                        from ..core import ChangeType
+
+                        row_id = int(rids[row_pos])
+                        ctx.store.append_diff(
+                            step_id=step_id,
+                            row_id=row_id,
+                            col=col_str,
+                            old_val=old_val,
+                            new_val=value,
+                            change_type=ChangeType.MODIFIED,
+                        )
+            except (KeyError, TypeError):
+                pass
+
+
+class TrackedIAtIndexer:
+    """
+    Wrapper around pandas _iAtIndexer that captures scalar assignments.
+
+    .iat is optimized for scalar access by integer position.
+    """
+
+    def __init__(self, indexer, df):
+        self._indexer = indexer
+        self._df = df
+
+    def __getitem__(self, key):
+        return self._indexer[key]
+
+    def __setitem__(self, key, value) -> None:
+        """Capture scalar assignment via iat[row, col] = value."""
+        ctx = get_context()
+        if not ctx or not ctx.enabled:
+            self._indexer[key] = value
+            return
+
+        row_pos, col_pos = key
+        col_str = self._df.columns[col_pos] if col_pos < len(self._df.columns) else str(col_pos)
+
+        # Check if column is watched
+        should_track = col_str in ctx.watched_columns if ctx.watched_columns else False
+
+        # Capture before state
+        old_val = None
+        if should_track:
+            try:
+                old_val = self._df.iat[row_pos, col_pos]
+            except (KeyError, IndexError):
+                pass
+
+        # Execute original
+        self._indexer[key] = value
+
+        # Capture after state
+        if should_track and ctx.store:
+            from .pandas_inst import get_caller_info
+
+            code_file, code_line = get_caller_info(skip_frames=2)
+            step_id = ctx.store.append_step(
+                operation="DataFrame.iat[]=",
+                stage=ctx.current_stage,
+                code_file=code_file,
+                code_line=code_line,
+                params={"row": row_pos, "col": col_str},
+                input_shape=self._df.shape,
+                output_shape=self._df.shape,
+            )
+
+            # Get row_id for this position
+            try:
+                rids = ctx.row_manager.get_ids_array(self._df)
+                if rids is None:
+                    ctx.row_manager.register(self._df)
+                    rids = ctx.row_manager.get_ids_array(self._df)
+                if rids is not None and row_pos < len(rids):
+                    from ..core import ChangeType
+
+                    row_id = int(rids[row_pos])
+                    ctx.store.append_diff(
+                        step_id=step_id,
+                        row_id=row_id,
+                        col=col_str,
+                        old_val=old_val,
+                        new_val=value,
+                        change_type=ChangeType.MODIFIED,
+                    )
+            except (KeyError, TypeError):
+                pass
+
+
+def instrument_indexers():
+    """
+    Install tracked indexers for loc, iloc, at, iat.
+
+    Monkey-patches DataFrame.loc, DataFrame.iloc, DataFrame.at, DataFrame.iat properties.
+    """
+    global _original_loc, _original_iloc, _original_at, _original_iat
+
+    if _original_loc is not None:
+        # Already instrumented
+        return
+
+    _original_loc = pd.DataFrame.loc.fget
+    _original_iloc = pd.DataFrame.iloc.fget
+    _original_at = pd.DataFrame.at.fget
+    _original_iat = pd.DataFrame.iat.fget
+
+    @property
+    def tracked_loc(self):
+        return TrackedLocIndexer(_original_loc(self), self)
+
+    @property
+    def tracked_iloc(self):
+        return TrackedILocIndexer(_original_iloc(self), self)
+
+    @property
+    def tracked_at(self):
+        return TrackedAtIndexer(_original_at(self), self)
+
+    @property
+    def tracked_iat(self):
+        return TrackedIAtIndexer(_original_iat(self), self)
+
+    pd.DataFrame.loc = tracked_loc
+    pd.DataFrame.iloc = tracked_iloc
+    pd.DataFrame.at = tracked_at
+    pd.DataFrame.iat = tracked_iat
+
+
+def uninstrument_indexers():
+    """Restore original loc/iloc/at/iat."""
+    global _original_loc, _original_iloc, _original_at, _original_iat
+
+    if _original_loc is not None:
+        pd.DataFrame.loc = property(_original_loc)
+        _original_loc = None
+    if _original_iloc is not None:
+        pd.DataFrame.iloc = property(_original_iloc)
+        _original_iloc = None
+    if _original_at is not None:
+        pd.DataFrame.at = property(_original_at)
+        _original_at = None
+    if _original_iat is not None:
+        pd.DataFrame.iat = property(_original_iat)
+        _original_iat = None