robin_sparkless_polars/

column.rs

use polars::prelude::{
    DataType, Expr, Field, PolarsError, PolarsResult, RankMethod, RankOptions, SortOptions,
    TimeUnit, WindowMapping, col, lit,
};
use polars_plan::dsl::AggExpr;
use std::ops::Neg;

/// Unwrap UDF result to Column (map() expects Result<Column>, UDFs return Result<Option<Column>>).
#[inline]
pub(crate) fn expect_col(
    r: PolarsResult<Option<polars::prelude::Column>>,
) -> PolarsResult<polars::prelude::Column> {
    r.and_then(|o| o.ok_or_else(|| PolarsError::ComputeError("expected column".into())))
}

/// Convert SQL LIKE pattern (% = any sequence, _ = one char) to regex. Escapes regex specials.
/// When escape_char is Some(esc), esc + any char treats that char as literal (no %/_ expansion).
fn like_pattern_to_regex(pattern: &str, escape_char: Option<char>) -> String {
    let mut out = String::with_capacity(pattern.len() * 2);
    let mut it = pattern.chars();
    while let Some(c) = it.next() {
        if escape_char == Some(c) {
            if let Some(next) = it.next() {
                // Literal: escape for regex
                if "\\.*+?[](){}^$|".contains(next) {
                    out.push('\\');
                }
                out.push(next);
            } else {
                out.push('\\');
                out.push(c);
            }
        } else {
            match c {
                '%' => out.push_str(".*"),
                '_' => out.push('.'),
                '\\' | '.' | '+' | '*' | '?' | '[' | ']' | '(' | ')' | '{' | '}' | '^' | '$'
                | '|' => {
                    out.push('\\');
                    out.push(c);
                }
                _ => out.push(c),
            }
        }
    }
    format!("^{out}$")
}

/// Map PySpark date_trunc/trunc format (year, month, day, hour, etc.) to Polars duration string (1y, 1mo, 1d, 1h).
/// Polars dt.truncate expects a leading integer; PySpark uses unit names without a number.
fn pyspark_trunc_format_to_polars_duration(format: &str) -> String {
    match format.to_lowercase().as_str() {
        "year" | "years" => "1y".to_string(),
        "month" | "months" => "1mo".to_string(),
        "week" | "weeks" | "wk" => "1w".to_string(),
        "day" | "days" => "1d".to_string(),
        "hour" | "hours" => "1h".to_string(),
        "minute" | "minutes" | "min" => "1m".to_string(),
        "second" | "seconds" | "sec" => "1s".to_string(),
        "quarter" | "quarters" | "q" => "1q".to_string(),
        _ => format.to_string(), // already Polars-style (e.g. 1d) or pass through
    }
}

/// Deferred random column: when added via with_column, we generate a full-length series in one go (PySpark-like).
#[derive(Debug, Clone, Copy)]
pub enum DeferredRandom {
    Rand(Option<u64>),
    Randn(Option<u64>),
}

/// Marker for order-sensitive first_value/last_value (PySpark semantics: first/last by window order).
#[derive(Debug, Clone)]
pub struct FirstLastValue {
    /// The expression whose value we take (e.g. col("salary")).
    pub value_expr: Expr,
    /// false = first in order, true = last in order.
    pub is_last: bool,
}

/// Column - represents a column in a DataFrame, used for building expressions
/// Thin wrapper around Polars `Expr`. May carry a DeferredRandom for rand/randn so with_column can produce one value per row.
/// May carry UdfCall for Python UDFs (eager execution at with_column).
#[derive(Debug, Clone)]
pub struct Column {
    name: String,
    expr: Expr, // Polars expression for lazy evaluation
    /// True when this Column was constructed via array() so DataFrame-level logic
    /// can apply PySpark-specific type strictness (e.g., mixed bool + non-bool arrays; issue #1115).
    pub(crate) is_array_expr: bool,
    /// When Some, with_column generates a full-length random series instead of using expr (PySpark-like per-row rand/randn).
    pub deferred: Option<DeferredRandom>,
    /// When Some, with_column executes Python UDF eagerly (name, arg columns).
    pub udf_call: Option<(String, Vec<Column>)>,
    /// When Some, this aggregate (e.g. sum) can use cum_sum for running window when orderBy differs from partitionBy.
    pub source_for_running: Option<String>,
    /// When Some, over_window uses running mean (cum_sum/cum_count) when orderBy is present (#1241).
    pub source_for_running_mean: Option<String>,
    /// When Some, over_window uses order-sensitive first/last (PySpark first_value/last_value semantics; #1145).
    pub first_last_value: Option<FirstLastValue>,
    /// When Some, over_window uses cum_count for running count semantics (PySpark count().over(order); #1218).
    pub source_for_running_count: Option<String>,
}

/// True if the expression is or contains a count_distinct/n_unique aggregate (PySpark rejects distinct window functions).
fn expr_is_or_contains_n_unique(expr: &Expr) -> bool {
    match expr {
        Expr::Agg(AggExpr::NUnique(_)) => true,
        Expr::Cast { expr: inner, .. } => expr_is_or_contains_n_unique(inner.as_ref()),
        Expr::Alias(inner, _) => expr_is_or_contains_n_unique(inner.as_ref()),
        _ => false,
    }
}

impl Column {
    /// Create a new Column from a column name
    pub fn new(name: String) -> Self {
        Column {
            name: name.clone(),
            expr: col(&name),
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Create a Column from a Polars Expr
    pub fn from_expr(expr: Expr, name: Option<String>) -> Self {
        let display_name = name.unwrap_or_else(|| "<expr>".to_string());
        Column {
            name: display_name,
            expr,
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Last aggregate for groupBy.agg() and for .over(window). When used with .over() and orderBy,
    /// over_window uses first_last_value so "last" = current row (PySpark default frame parity).
    pub fn from_last_agg(col: &Column) -> Self {
        let value_expr = col.expr().clone();
        let expr = value_expr.clone().last();
        Column {
            name: "last".to_string(),
            expr,
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: Some(FirstLastValue {
                value_expr,
                is_last: true,
            }),
            source_for_running_count: None,
        }
    }

    /// Create a Column for Python UDF call (eager execution at with_column).
    pub fn from_udf_call(name: String, args: Vec<Column>) -> Self {
        Column {
            name: format!("{name}()"),
            expr: lit(0i32), // dummy, never used
            is_array_expr: false,
            deferred: None,
            udf_call: Some((name, args)),
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Create a Column for rand(seed). When used in with_column, generates one value per row (PySpark-like).
    pub fn from_rand(seed: Option<u64>) -> Self {
        let expr = lit(1i64).cum_sum(false).map(
            move |c| expect_col(crate::udfs::apply_rand_with_seed(c, seed)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Column {
            name: "rand".to_string(),
            expr,
            is_array_expr: false,
            deferred: Some(DeferredRandom::Rand(seed)),
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Create a Column for randn(seed). When used in with_column, generates one value per row (PySpark-like).
    pub fn from_randn(seed: Option<u64>) -> Self {
        let expr = lit(1i64).cum_sum(false).map(
            move |c| expect_col(crate::udfs::apply_randn_with_seed(c, seed)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Column {
            name: "randn".to_string(),
            expr,
            is_array_expr: false,
            deferred: Some(DeferredRandom::Randn(seed)),
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Get the underlying Polars Expr
    pub fn expr(&self) -> &Expr {
        &self.expr
    }

    /// Convert to Polars Expr (consumes self).
    /// Applies the Column's display name as alias so row keys match (PySpark parity #1014, #1017, #1022).
    pub fn into_expr(self) -> Expr {
        self.expr.alias(&self.name)
    }

    /// Get the column name
    pub fn name(&self) -> &str {
        &self.name
    }

    /// If this column is a Python UDF call, return (udf_name, arg_column_names). Used by Python to run UDFs at with_column.
    pub fn udf_call_info(&self) -> Option<(String, Vec<String>)> {
        self.udf_call.as_ref().map(|(name, args)| {
            (
                name.clone(),
                args.iter().map(|c| c.name().to_string()).collect(),
            )
        })
    }

    /// If this column is a Python UDF call, return (udf_name, args slice). Used to materialize expression args before calling the executor.
    pub fn udf_call_with_args(&self) -> Option<(&str, &[Column])> {
        self.udf_call
            .as_ref()
            .map(|(name, args)| (name.as_str(), args.as_slice()))
    }

    /// If this column is a literal expression, return its value as JSON string for Python UDF executor (literal args).
    pub fn literal_as_json_string(&self) -> Option<String> {
        match &self.expr {
            Expr::Literal(lv) => crate::dataframe::literal_value_to_serde_value(lv)
                .and_then(|v| serde_json::to_string(&v).ok()),
            _ => None,
        }
    }

    /// If this column is a Python UDF call, return (udf_name, arg_names, arg_literal_json_strings).
    /// For each arg: None = use row[arg_name]; Some(json) = literal value (not in row).
    pub fn udf_call_info_with_literals(
        &self,
    ) -> Option<(String, Vec<String>, Vec<Option<String>>)> {
        self.udf_call.as_ref().map(|(name, args)| {
            let arg_names: Vec<String> = args.iter().map(|c| c.name().to_string()).collect();
            let literals: Vec<Option<String>> =
                args.iter().map(|c| c.literal_as_json_string()).collect();
            (name.clone(), arg_names, literals)
        })
    }

    /// Alias the column
    pub fn alias(&self, name: &str) -> Column {
        Column {
            name: name.to_string(),
            expr: self.expr.clone().alias(name),
            is_array_expr: self.is_array_expr,
            deferred: self.deferred,
            udf_call: self.udf_call.clone(),
            source_for_running: self.source_for_running.clone(),
            source_for_running_mean: self.source_for_running_mean.clone(),
            first_last_value: self.first_last_value.clone(),
            source_for_running_count: self.source_for_running_count.clone(),
        }
    }

    /// Ascending sort, nulls first (Spark default for ASC). PySpark asc.
    pub fn asc(&self) -> crate::functions::SortOrder {
        crate::functions::asc(self)
    }

    /// Ascending sort, nulls first. PySpark asc_nulls_first.
    pub fn asc_nulls_first(&self) -> crate::functions::SortOrder {
        crate::functions::asc_nulls_first(self)
    }

    /// Ascending sort, nulls last. PySpark asc_nulls_last.
    pub fn asc_nulls_last(&self) -> crate::functions::SortOrder {
        crate::functions::asc_nulls_last(self)
    }

    /// Descending sort, nulls last (Spark default for DESC). PySpark desc.
    pub fn desc(&self) -> crate::functions::SortOrder {
        crate::functions::desc(self)
    }

    /// Descending sort, nulls first. PySpark desc_nulls_first.
    pub fn desc_nulls_first(&self) -> crate::functions::SortOrder {
        crate::functions::desc_nulls_first(self)
    }

    /// Descending sort, nulls last. PySpark desc_nulls_last.
    pub fn desc_nulls_last(&self) -> crate::functions::SortOrder {
        crate::functions::desc_nulls_last(self)
    }

    /// Check if column is null
    pub fn is_null(&self) -> Column {
        Column {
            name: format!("({} IS NULL)", self.name),
            expr: self.expr.clone().is_null(),
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Check if column is not null
    pub fn is_not_null(&self) -> Column {
        Column {
            name: format!("({} IS NOT NULL)", self.name),
            expr: self.expr.clone().is_not_null(),
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: None,
            source_for_running_count: None,
        }
    }

    /// Alias for is_null. PySpark isnull.
    pub fn isnull(&self) -> Column {
        self.is_null()
    }

    /// Alias for is_not_null. PySpark isnotnull.
    pub fn isnotnull(&self) -> Column {
        self.is_not_null()
    }

    /// Create a null boolean expression
    fn null_boolean_expr() -> Expr {
        use polars::prelude::*;
        // Create an expression that is always a null boolean
        lit(NULL).cast(DataType::Boolean)
    }

    /// SQL LIKE pattern matching (% = any chars, _ = one char). PySpark like.
    /// When escape_char is Some(esc), esc + char treats that char as literal (e.g. \\% = literal %).
    pub fn like(&self, pattern: &str, escape_char: Option<char>) -> Column {
        let regex = like_pattern_to_regex(pattern, escape_char);
        self.regexp_like(&regex)
    }

    /// Case-insensitive LIKE. PySpark ilike.
    /// When escape_char is Some(esc), esc + char treats that char as literal.
    pub fn ilike(&self, pattern: &str, escape_char: Option<char>) -> Column {
        use polars::prelude::*;
        let regex = format!("(?i){}", like_pattern_to_regex(pattern, escape_char));
        Self::from_expr(self.expr().clone().str().contains(lit(regex), false), None)
    }

    /// PySpark-style equality comparison (NULL == NULL returns NULL, not True)
    /// Any comparison involving NULL returns NULL
    ///
    /// Explicitly wraps comparisons with null checks to ensure PySpark semantics.
    /// If either side is NULL, the result is NULL.
    pub fn eq_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard equality comparison
        let eq_result = self.expr().clone().eq(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(eq_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    /// PySpark-style inequality comparison (NULL != NULL returns NULL, not False)
    /// Any comparison involving NULL returns NULL
    pub fn ne_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard inequality comparison
        let ne_result = self.expr().clone().neq(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(ne_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    /// Null-safe equality (NULL <=> NULL returns True)
    /// PySpark's eqNullSafe() method. Applies type coercion (e.g. string vs int) for PySpark parity (#266).
    pub fn eq_null_safe(&self, other: &Column) -> Column {
        use crate::functions::{lit_bool, when};

        let (left_c, right_c) = crate::type_coercion::coerce_for_pyspark_eq_null_safe(
            self.expr().clone(),
            other.expr().clone(),
        )
        .unwrap_or_else(|_| (self.expr().clone(), other.expr().clone()));

        let left_null = left_c.clone().is_null();
        let right_null = right_c.clone().is_null();
        let both_null = left_null.clone().and(right_null.clone());
        let either_null = left_null.clone().or(right_null.clone());

        // Standard equality (on coerced exprs)
        let eq_result = left_c.eq(right_c);

        // If both are null, return True
        // If either is null (but not both), return False
        // Otherwise, return standard equality result
        when(&Self::from_expr(both_null, None))
            .then(&lit_bool(true))
            .otherwise(
                &when(&Self::from_expr(either_null, None))
                    .then(&lit_bool(false))
                    .otherwise(&Self::from_expr(eq_result, None)),
            )
    }

    /// Create a Column that is always a null boolean.
    /// This is useful for downstream bindings (e.g. PyO3) that need a null literal
    /// without depending directly on Polars types like `Expr` or `LiteralValue`.
    pub fn null_boolean() -> Column {
        Column::from_expr(Self::null_boolean_expr(), None)
    }

    /// Create a Column that is always a null value of the given type.
    /// `dtype` is a type name string (e.g. `"boolean"`, `"string"`, `"bigint"`, `"double"`).
    /// See [`crate::functions::parse_type_name`] for supported names.
    /// Returns `Err` on unknown type name so bindings get a clear error.
    pub fn lit_null(dtype: &str) -> Result<Column, String> {
        use polars::prelude::{NULL, lit};
        let dt = crate::functions::parse_type_name(dtype)?;
        Ok(Column::from_expr(lit(NULL).cast(dt), None))
    }

    /// Create a Column from a boolean literal. Convenience for bindings that prefer method form.
    pub fn from_bool(b: bool) -> Column {
        crate::functions::lit_bool(b)
    }

    /// Create a Column from an i64 literal. Convenience for bindings that prefer method form.
    pub fn from_i64(n: i64) -> Column {
        crate::functions::lit_i64(n)
    }

    /// Create a Column from a string literal. Convenience for bindings that prefer method form.
    pub fn from_string(s: &str) -> Column {
        crate::functions::lit_str(s)
    }

    /// PySpark-style greater-than comparison (NULL > value returns NULL)
    /// Any comparison involving NULL returns NULL
    pub fn gt_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard greater-than comparison
        let gt_result = self.expr().clone().gt(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(gt_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    /// PySpark-style greater-than-or-equal comparison
    /// Any comparison involving NULL returns NULL
    pub fn ge_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard greater-than-or-equal comparison
        let ge_result = self.expr().clone().gt_eq(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(ge_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    /// PySpark-style less-than comparison
    /// Any comparison involving NULL returns NULL
    pub fn lt_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard less-than comparison
        let lt_result = self.expr().clone().lt(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(lt_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    /// PySpark-style less-than-or-equal comparison
    /// Any comparison involving NULL returns NULL
    pub fn le_pyspark(&self, other: &Column) -> Column {
        // Check if either side is NULL
        let left_null = self.expr().clone().is_null();
        let right_null = other.expr().clone().is_null();
        let either_null = left_null.clone().or(right_null.clone());

        // Standard less-than-or-equal comparison
        let le_result = self.expr().clone().lt_eq(other.expr().clone());

        // Wrap: if either is null, return null boolean, else return comparison result
        let null_boolean = Self::null_boolean_expr();
        let null_aware_expr = crate::functions::when(&Self::from_expr(either_null, None))
            .then(&Self::from_expr(null_boolean, None))
            .otherwise(&Self::from_expr(le_result, None));

        Self::from_expr(null_aware_expr.into_expr(), None)
    }

    // Standard comparison methods that work with Expr (for literals and columns)
    // These delegate to Polars and may not match PySpark null semantics exactly.
    // Use _pyspark variants for explicit PySpark semantics.

    /// Greater than comparison
    pub fn gt(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().gt(other), None)
    }

    /// Greater than or equal comparison
    pub fn gt_eq(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().gt_eq(other), None)
    }

    /// Less than comparison
    pub fn lt(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().lt(other), None)
    }

    /// Less than or equal comparison
    pub fn lt_eq(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().lt_eq(other), None)
    }

    /// True if column value is between lower and upper (inclusive). PySpark between(low, high).
    /// Applies string–numeric coercion so col("val").between(1, 10) works when val is string (#628).
    pub fn between(&self, lower: &Column, upper: &Column) -> Column {
        use crate::type_coercion::{CompareOp, coerce_for_pyspark_comparison};
        use polars::prelude::*;

        let left = self.expr().clone();
        let lower_expr = lower.expr().clone();
        let upper_expr = upper.expr().clone();

        let infer_lit_type = |e: &Expr| -> Option<DataType> {
            if let Expr::Literal(lv) = e {
                let dt = lv.get_datatype();
                if matches!(dt, DataType::Unknown(_)) {
                    None
                } else {
                    Some(dt)
                }
            } else {
                None
            }
        };

        let lower_ty = infer_lit_type(&lower_expr).unwrap_or(DataType::String);
        let upper_ty = infer_lit_type(&upper_expr).unwrap_or(DataType::String);
        let lt = DataType::String;

        let (left_c, lower_c) = match coerce_for_pyspark_comparison(
            left.clone(),
            lower_expr.clone(),
            &lt,
            &lower_ty,
            &CompareOp::GtEq,
        ) {
            Ok((a, b)) => (a, b),
            Err(_) => (left.clone(), lower_expr),
        };

        let upper_clone = upper.expr().clone();
        let (left_cc, upper_c) = match coerce_for_pyspark_comparison(
            left_c.clone(),
            upper_expr,
            &lt,
            &upper_ty,
            &CompareOp::LtEq,
        ) {
            Ok((a, b)) => (a, b),
            Err(_) => (left_c.clone(), upper_clone),
        };

        let ge = left_cc.clone().gt_eq(lower_c);
        let le = left_cc.lt_eq(upper_c);
        Self::from_expr(ge.and(le), None)
    }

    /// Equality comparison
    pub fn eq(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().eq(other), None)
    }

    /// Inequality comparison
    pub fn neq(&self, other: Expr) -> Column {
        Self::from_expr(self.expr().clone().neq(other), None)
    }

    /// Logical AND of two boolean columns (PySpark and_).
    pub fn and_(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone().and(other.expr().clone()), None)
    }

    /// Logical OR of two boolean columns (PySpark or_).
    pub fn or_(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone().or(other.expr().clone()), None)
    }

    // Equality comparison with special handling for string-vs-numeric literals (issue #235).
    //
    // When comparing a column to a numeric literal (e.g. col("s") == lit(123)), Polars
    // normally raises `cannot compare string with numeric type` if the column is a
    // string column. PySpark, however, coerces types (string → numeric) and performs
    // the comparison, treating invalid strings as null (non-matching in filters).

    // --- String functions ---

    /// Convert string column to uppercase (PySpark upper)
    pub fn upper(&self) -> Column {
        Self::from_expr(self.expr().clone().str().to_uppercase(), None)
    }

    /// Convert string column to lowercase (PySpark lower)
    pub fn lower(&self) -> Column {
        Self::from_expr(self.expr().clone().str().to_lowercase(), None)
    }

    /// Alias for lower. PySpark lcase.
    pub fn lcase(&self) -> Column {
        self.lower()
    }

    /// Alias for upper. PySpark ucase.
    pub fn ucase(&self) -> Column {
        self.upper()
    }

    /// Substring with 1-based start (PySpark substring/substr semantics).
    /// - Start 0: 0-based (first char), for Sparkless/Python parity (#875).
    /// - Positive start: 1-based index (1 = first char).
    /// - Negative start: count from end (e.g. -3 = third char from end).
    /// - Length less than 1: empty string; null input yields null (Phase 7 / PySpark parity).
    pub fn substr(&self, start: i64, length: Option<i64>) -> Column {
        use polars::prelude::*;
        // PySpark: len < 1 -> empty string; null input must stay null (test_substr_pyspark_parity_comprehensive)
        if length.map(|l| l < 1).unwrap_or(false) {
            let expr = when(self.expr().clone().is_null())
                .then(lit(NULL))
                .otherwise(lit(""));
            return Self::from_expr(expr, None);
        }
        let len_chars = self.expr().clone().str().len_chars();
        // Start 0: 0-based (substr(0, 3) = "Hel"); start >= 1: 1-based; negative: from end.
        let offset_expr = if start == 0 {
            lit(0i64)
        } else if start >= 1 {
            lit((start - 1).max(0))
        } else {
            let from_end = len_chars + lit(start);
            when(from_end.clone().lt(lit(0i64)))
                .then(lit(0i64))
                .otherwise(from_end)
        };
        let length_expr = length.map(lit).unwrap_or_else(|| lit(i64::MAX));
        Self::from_expr(
            self.expr().clone().str().slice(offset_expr, length_expr),
            None,
        )
    }

    /// String length in characters (PySpark length)
    pub fn length(&self) -> Column {
        Self::from_expr(self.expr().clone().str().len_chars(), None)
    }

    /// Bit length of string in bytes * 8 (PySpark bit_length).
    pub fn bit_length(&self) -> Column {
        use polars::prelude::*;
        let len_bytes = self.expr().clone().str().len_bytes().cast(DataType::Int32);
        Self::from_expr((len_bytes * lit(8i32)).cast(DataType::Int32), None)
    }

    /// Length of string in bytes (PySpark octet_length).
    pub fn octet_length(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr().clone().str().len_bytes().cast(DataType::Int32),
            None,
        )
    }

    /// Length of string in characters (PySpark char_length). Alias of length().
    pub fn char_length(&self) -> Column {
        self.length()
    }

    /// Length of string in characters (PySpark character_length). Alias of length().
    pub fn character_length(&self) -> Column {
        self.length()
    }

    /// Encode string to binary (PySpark encode). Charset: UTF-8. Returns hex string.
    pub fn encode(&self, charset: &str) -> Column {
        let charset = charset.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_encode(s, &charset)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Decode binary (hex string) to string (PySpark decode). Charset: UTF-8.
    pub fn decode(&self, charset: &str) -> Column {
        let charset = charset.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_decode(s, &charset)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert to binary (PySpark to_binary). fmt: 'utf-8', 'hex'. Returns hex string.
    pub fn to_binary(&self, fmt: &str) -> Column {
        let fmt = fmt.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_to_binary(s, &fmt)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Try convert to binary; null on failure (PySpark try_to_binary).
    pub fn try_to_binary(&self, fmt: &str) -> Column {
        let fmt = fmt.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_try_to_binary(s, &fmt)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// AES encrypt (PySpark aes_encrypt). Key as string; AES-128-GCM. Output hex(nonce||ciphertext).
    pub fn aes_encrypt(&self, key: &str) -> Column {
        let key = key.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_aes_encrypt(s, &key)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// AES decrypt (PySpark aes_decrypt). Input hex(nonce||ciphertext). Null on failure.
    pub fn aes_decrypt(&self, key: &str) -> Column {
        let key = key.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_aes_decrypt(s, &key)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Try AES decrypt (PySpark try_aes_decrypt). Returns null on failure.
    pub fn try_aes_decrypt(&self, key: &str) -> Column {
        let key = key.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_try_aes_decrypt(s, &key)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Data type as string (PySpark typeof). Uses dtype from schema.
    pub fn typeof_(&self) -> Column {
        Self::from_expr(
            self.expr().clone().map(
                |s| expect_col(crate::udfs::apply_typeof(s)),
                |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
            ),
            None,
        )
    }

    /// Trim leading and trailing whitespace (PySpark trim). Default behavior trims ASCII space only
    /// so tabs are preserved when using nested LTRIM/RTRIM (issue #434 / #1078).
    pub fn trim(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().str().strip_chars(lit(" ")), None)
    }

    /// Trim leading whitespace (PySpark ltrim)
    pub fn ltrim(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().str().strip_chars_start(lit(" ")), None)
    }

    /// Trim trailing whitespace (PySpark rtrim)
    pub fn rtrim(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().str().strip_chars_end(lit(" ")), None)
    }

    /// Trim leading and trailing characters (PySpark btrim). trim_str defaults to whitespace.
    pub fn btrim(&self, trim_str: Option<&str>) -> Column {
        use polars::prelude::*;
        let chars = trim_str.unwrap_or(" ");
        Self::from_expr(self.expr().clone().str().strip_chars(lit(chars)), None)
    }

    /// Find substring position 1-based, starting at pos (PySpark locate). 0 if not found.
    pub fn locate(&self, substr: &str, pos: i64) -> Column {
        use polars::prelude::*;
        if substr.is_empty() {
            return Self::from_expr(lit(1i64), None);
        }
        let start = (pos - 1).max(0);
        let slice_expr = self.expr().clone().str().slice(lit(start), lit(i64::MAX));
        let found = slice_expr.str().find_literal(lit(substr.to_string()));
        let expr = (found.cast(DataType::Int64) + lit(start + 1))
            .fill_null(lit(0i64))
            .cast(DataType::Int64);
        Self::from_expr(expr, None)
    }

    /// Base conversion (PySpark conv). num_str from from_base to to_base.
    pub fn conv(&self, from_base: i32, to_base: i32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_conv(s, from_base, to_base)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert to hex string (PySpark hex). Int or string input.
    pub fn hex(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_hex(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert hex string to binary/string (PySpark unhex).
    pub fn unhex(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_unhex(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert integer to binary string (PySpark bin).
    pub fn bin(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_bin(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Get bit at 0-based position (PySpark getbit).
    pub fn getbit(&self, pos: i64) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_getbit(s, pos)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Bitwise AND of two integer/boolean columns (PySpark bit_and).
    pub fn bit_and(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().cast(DataType::Int64).map_many(
            |cols| expect_col(crate::udfs::apply_bit_and(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Bitwise OR of two integer/boolean columns (PySpark bit_or).
    pub fn bit_or(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().cast(DataType::Int64).map_many(
            |cols| expect_col(crate::udfs::apply_bit_or(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Bitwise XOR of two integer/boolean columns (PySpark bit_xor).
    pub fn bit_xor(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().cast(DataType::Int64).map_many(
            |cols| expect_col(crate::udfs::apply_bit_xor(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Count of set bits in the integer representation (PySpark bit_count).
    pub fn bit_count(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_bit_count(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Assert that all boolean values are true; errors otherwise (PySpark assert_true).
    /// When err_msg is Some, it is used in the error message when assertion fails.
    pub fn assert_true(&self, err_msg: Option<&str>) -> Column {
        let msg = err_msg.map(String::from);
        let expr = self.expr().clone().map(
            move |c| expect_col(crate::udfs::apply_assert_true(c, msg.as_deref())),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Bitwise NOT of an integer/boolean column (PySpark bitwise_not / bitwiseNOT).
    /// #859: Coerce via map so Unknown(Any) from when/otherwise is cast to Int64 at execution time.
    pub fn bitwise_not(&self) -> Column {
        use polars::prelude::Field;
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_coerce_to_int64_for_bitwise(col)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        let expr = (lit(-1i64) - expr).cast(DataType::Int64);
        Self::from_expr(expr, None)
    }

    /// Logical NOT of a boolean column (PySpark `~` on Column = boolean NOT).
    /// Fails at execution time if column is not Boolean (#405, #1236); use F.expr("~x") for bitwise NOT on integers.
    pub fn logical_not(&self) -> Column {
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_logical_not_boolean_only(col)),
            |_schema, field| {
                if field.dtype() == &DataType::Boolean {
                    Ok(field.clone())
                } else {
                    Err(PolarsError::ComputeError(
                        "logical NOT (~) requires boolean type".into(),
                    ))
                }
            },
        );
        Self::from_expr(expr, None)
    }

    /// Parse string to map (PySpark str_to_map). "k1:v1,k2:v2" -> map.
    pub fn str_to_map(&self, pair_delim: &str, key_value_delim: &str) -> Column {
        let pair_delim = pair_delim.to_string();
        let key_value_delim = key_value_delim.to_string();
        let expr = self.expr().clone().map(
            move |s| {
                expect_col(crate::udfs::apply_str_to_map(
                    s,
                    &pair_delim,
                    &key_value_delim,
                ))
            },
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// True if regex pattern contains lookahead/lookbehind (Polars regex does not support these).
    fn pattern_has_lookaround(pattern: &str) -> bool {
        let p = pattern.as_bytes();
        let n = p.len();
        let mut i = 0;
        while i + 2 < n {
            if p[i] == b'(' && p[i + 1] == b'?' {
                match p[i + 2] {
                    b'=' | b'!' => return true, // (?= (?! lookahead
                    b'<' if i + 4 <= n && (p[i + 3] == b'=' || p[i + 3] == b'!') => return true, // (?<= (?<! lookbehind
                    _ => {}
                }
            }
            i += 1;
        }
        false
    }

    /// Extract first match of regex pattern (PySpark regexp_extract). Group 0 = full match.
    /// When pattern contains lookahead/lookbehind, uses fancy-regex (Polars regex does not support them).
    pub fn regexp_extract(&self, pattern: &str, group_index: usize) -> Column {
        use polars::prelude::*;
        if Self::pattern_has_lookaround(pattern) {
            let pat = pattern.to_string();
            let group = group_index;
            Self::from_expr(
                self.expr().clone().map(
                    move |s| {
                        expect_col(crate::udfs::apply_regexp_extract_lookaround(s, &pat, group))
                    },
                    |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
                ),
                None,
            )
        } else {
            let pat = pattern.to_string();
            Self::from_expr(
                // PySpark implicitly casts non-string inputs to string for regex functions,
                // returns empty string for no-match, and null only for null input.
                self.expr().clone().map(
                    move |s| expect_col(crate::udfs::apply_regexp_extract(s, &pat, group_index)),
                    |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
                ),
                None,
            )
        }
    }

    /// Replace all matches of regex pattern (PySpark regexp_replace: global replace).
    pub fn regexp_replace(&self, pattern: &str, replacement: &str) -> Column {
        use polars::prelude::*;
        let pat = pattern.to_string();
        let rep = replacement.to_string();
        Self::from_expr(
            // PySpark implicitly casts non-string inputs to string for regex functions.
            self.expr()
                .clone()
                .cast(DataType::String)
                .str()
                .replace_all(lit(pat), lit(rep), false),
            None,
        )
    }

    /// Leftmost n characters (PySpark left).
    pub fn left(&self, n: i64) -> Column {
        use polars::prelude::*;
        let len = n.max(0) as u32;
        Self::from_expr(
            self.expr().clone().str().slice(lit(0i64), lit(len as i64)),
            None,
        )
    }

    /// Rightmost n characters (PySpark right).
    pub fn right(&self, n: i64) -> Column {
        use polars::prelude::*;
        let n_val = n.max(0);
        let n_expr = lit(n_val);
        let len_chars = self.expr().clone().str().len_chars().cast(DataType::Int64);
        let start = when((len_chars.clone() - n_expr.clone()).lt_eq(lit(0i64)))
            .then(lit(0i64))
            .otherwise(len_chars - n_expr.clone());
        Self::from_expr(self.expr().clone().str().slice(start, n_expr), None)
    }

    /// Replace all occurrences of literal search string with replacement (PySpark replace for literal).
    pub fn replace(&self, search: &str, replacement: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr().clone().str().replace_all(
                lit(search.to_string()),
                lit(replacement.to_string()),
                true,
            ),
            None,
        )
    }

    /// Replace multiple (search, replacement) pairs in order (PySpark replace with dict/list).
    pub fn replace_many(&self, pairs: &[(String, String)]) -> Column {
        let mut out = self.clone();
        for (search, replacement) in pairs {
            out = out.replace(search, replacement);
        }
        out
    }

    /// True if string starts with prefix (PySpark startswith).
    pub fn startswith(&self, prefix: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .starts_with(lit(prefix.to_string())),
            None,
        )
    }

    /// True if string ends with suffix (PySpark endswith).
    pub fn endswith(&self, suffix: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr().clone().str().ends_with(lit(suffix.to_string())),
            None,
        )
    }

    /// True if string contains substring (literal, not regex). PySpark contains.
    pub fn contains(&self, substring: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .contains(lit(substring.to_string()), true),
            None,
        )
    }

    /// Split string by delimiter (PySpark split). Returns list of strings.
    /// When limit is Some(n) with n > 0, at most n parts; remainder in last part. None or <= 0: no limit.
    /// Uses literal split so "|" is not interpreted as regex alternation.
    pub fn split(&self, delimiter: &str, limit: Option<i32>) -> Column {
        use polars::prelude::*;
        let use_limit = limit.is_some_and(|l| l > 0);
        if use_limit {
            let delim = delimiter.to_string();
            let lim = limit.unwrap_or(0);
            let expr = self.expr().clone().map(
                move |col| expect_col(crate::udfs::apply_split_with_limit(col, &delim, lim)),
                |_schema, field| {
                    Ok(Field::new(
                        field.name().clone(),
                        DataType::List(Box::new(DataType::String)),
                    ))
                },
            );
            Self::from_expr(expr, None)
        } else {
            Self::from_expr(
                self.expr().clone().str().split(lit(delimiter.to_string())),
                None,
            )
        }
    }

    /// Title case: first letter of each word uppercase (PySpark initcap). Uses UDF; Polars has no to_titlecase in 0.53.
    pub fn initcap(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_initcap(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Extract all matches of regex (PySpark regexp_extract_all). Returns list of strings.
    pub fn regexp_extract_all(&self, pattern: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .extract_all(lit(pattern.to_string())),
            None,
        )
    }

    /// Extract all matches of regex with capture group index (PySpark regexp_extract_all(col, pattern, idx)).
    /// idx=0 returns whole match; idx>0 returns capture group.
    pub fn regexp_extract_all_group(&self, pattern: &str, group_index: usize) -> Column {
        if group_index == 0 {
            return self.regexp_extract_all(pattern);
        }
        use polars::prelude::*;
        let pat = pattern.to_string();
        let idx = group_index;
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_regexp_extract_all_group(s, &pat, idx)),
            |_schema, field| {
                Ok(Field::new(
                    field.name().clone(),
                    DataType::List(Box::new(DataType::String)),
                ))
            },
        );
        Self::from_expr(expr, None)
    }

    /// Check if string matches regex (PySpark regexp_like / rlike).
    pub fn regexp_like(&self, pattern: &str) -> Column {
        use polars::prelude::*;
        // Polars regex engine (regex crate) does not support lookaround; use fancy-regex fallback.
        if pattern.contains("(?=")
            || pattern.contains("(?!")
            || pattern.contains("(?<=")
            || pattern.contains("(?<!")
        {
            let pat = pattern.to_string();
            let expr = self.expr().clone().map(
                move |s| expect_col(crate::udfs::apply_regexp_like_lookaround(s, &pat)),
                |_schema, field| Ok(Field::new(field.name().clone(), DataType::Boolean)),
            );
            return Self::from_expr(expr, None);
        }
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .contains(lit(pattern.to_string()), false),
            None,
        )
    }

    /// Count of non-overlapping regex matches (PySpark regexp_count).
    pub fn regexp_count(&self, pattern: &str) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .count_matches(lit(pattern.to_string()), false)
                .cast(DataType::Int64),
            None,
        )
    }

    /// First substring matching regex (PySpark regexp_substr). Null if no match.
    pub fn regexp_substr(&self, pattern: &str) -> Column {
        self.regexp_extract(pattern, 0)
    }

    /// 1-based position of first regex match (PySpark regexp_instr). group_idx 0 = full match; null if no match.
    pub fn regexp_instr(&self, pattern: &str, group_idx: Option<usize>) -> Column {
        let idx = group_idx.unwrap_or(0);
        let pattern = pattern.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_regexp_instr(s, pattern.clone(), idx)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// 1-based index of self in comma-delimited set column (PySpark find_in_set). 0 if not found or self contains comma.
    pub fn find_in_set(&self, set_column: &Column) -> Column {
        let args = [set_column.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_find_in_set(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Repeat string column n times (PySpark repeat). Each element repeated n times.
    pub fn repeat(&self, n: i32) -> Column {
        use polars::prelude::*;
        // repeat_by yields List[str]; join to get a single string per row.
        Self::from_expr(
            self.expr()
                .clone()
                .repeat_by(lit(n as u32))
                .list()
                .join(lit(""), false),
            None,
        )
    }

    /// Reverse string (PySpark reverse).
    pub fn reverse(&self) -> Column {
        Self::from_expr(self.expr().clone().str().reverse(), None)
    }

    /// Find substring position (1-based; 0 if not found). PySpark instr(col, substr).
    pub fn instr(&self, substr: &str) -> Column {
        use polars::prelude::*;
        let found = self
            .expr()
            .clone()
            .str()
            .find_literal(lit(substr.to_string()));
        // Polars find_literal returns 0-based index (null if not found); PySpark is 1-based, 0 when not found.
        Self::from_expr(
            (found.cast(DataType::Int64) + lit(1i64)).fill_null(lit(0i64)),
            None,
        )
    }

    /// Left-pad string to length with pad character (PySpark lpad).
    pub fn lpad(&self, length: i32, pad: &str) -> Column {
        let pad_str = if pad.is_empty() { " " } else { pad };
        let fill = pad_str.chars().next().unwrap_or(' ');
        Self::from_expr(
            self.expr()
                .clone()
                .str()
                .pad_start(lit(length as i64), fill),
            None,
        )
    }

    /// Right-pad string to length with pad character (PySpark rpad).
    pub fn rpad(&self, length: i32, pad: &str) -> Column {
        let pad_str = if pad.is_empty() { " " } else { pad };
        let fill = pad_str.chars().next().unwrap_or(' ');
        Self::from_expr(
            self.expr().clone().str().pad_end(lit(length as i64), fill),
            None,
        )
    }

    /// Character-by-character translation (PySpark translate). Replaces each char in from_str with corresponding in to_str; if to_str is shorter, extra from chars are removed.
    pub fn translate(&self, from_str: &str, to_str: &str) -> Column {
        use polars::prelude::*;
        let mut e = self.expr().clone();
        let from_chars: Vec<char> = from_str.chars().collect();
        let to_chars: Vec<char> = to_str.chars().collect();
        for (i, fc) in from_chars.iter().enumerate() {
            let f = fc.to_string();
            let t = to_chars
                .get(i)
                .map(|c| c.to_string())
                .unwrap_or_else(String::new); // PySpark: no replacement = drop char
            e = e.str().replace_all(lit(f), lit(t), true);
        }
        Self::from_expr(e, None)
    }

    /// Mask string: replace uppercase with upper_char, lowercase with lower_char, digits with digit_char (PySpark mask).
    /// Defaults: upper 'X', lower 'x', digit 'n'; other chars unchanged.
    pub fn mask(
        &self,
        upper_char: Option<char>,
        lower_char: Option<char>,
        digit_char: Option<char>,
        other_char: Option<char>,
    ) -> Column {
        use polars::prelude::*;
        let upper = upper_char.unwrap_or('X').to_string();
        let lower = lower_char.unwrap_or('x').to_string();
        let digit = digit_char.unwrap_or('n').to_string();
        let other = other_char.map(|c| c.to_string());
        let mut e = self
            .expr()
            .clone()
            .str()
            .replace_all(lit("[A-Z]".to_string()), lit(upper), false)
            .str()
            .replace_all(lit("[a-z]".to_string()), lit(lower), false)
            .str()
            .replace_all(lit(r"\d".to_string()), lit(digit), false);
        if let Some(o) = other {
            e = e
                .str()
                .replace_all(lit("[^A-Za-z0-9]".to_string()), lit(o), false);
        }
        Self::from_expr(e, None)
    }

    /// Split by delimiter and return 1-based part (PySpark split_part).
    /// part_num > 0: from left; part_num < 0: from right; part_num = 0: null; out-of-range: empty string.
    pub fn split_part(&self, delimiter: &str, part_num: i64) -> Column {
        use polars::prelude::*;
        if part_num == 0 {
            return Self::from_expr(lit(NULL), None);
        }
        let use_regex = delimiter == "|";
        if use_regex {
            let pattern = delimiter.to_string();
            let part = part_num;
            let get_expr = self.expr().clone().map(
                move |col| expect_col(crate::udfs::apply_split_part_regex(col, &pattern, part)),
                |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
            );
            let expr = when(self.expr().clone().is_null())
                .then(lit(NULL))
                .otherwise(get_expr.fill_null(lit("")));
            return Self::from_expr(expr, None);
        }
        let delim = delimiter.to_string();
        let split_expr = self.expr().clone().str().split(lit(delim));
        let index = if part_num > 0 {
            lit(part_num - 1)
        } else {
            lit(part_num)
        };
        let get_expr = split_expr.list().get(index, true).fill_null(lit(""));
        let expr = when(self.expr().clone().is_null())
            .then(lit(NULL))
            .otherwise(get_expr);
        Self::from_expr(expr, None)
    }

    /// Substring before/after nth delimiter (PySpark substring_index). count > 0: before nth from left; count < 0: after nth from right.
    pub fn substring_index(&self, delimiter: &str, count: i64) -> Column {
        use polars::prelude::*;
        // PySpark edge case: empty delimiter always yields empty string for non-null input,
        // and null when the input is null.
        if delimiter.is_empty() {
            let expr = when(self.expr().clone().is_null())
                .then(lit(NULL))
                .otherwise(lit("").cast(DataType::String));
            return Self::from_expr(expr, None);
        }
        let delim = delimiter.to_string();
        let split_expr = self.expr().clone().str().split(lit(delim.clone()));
        let n = count.unsigned_abs() as i64;
        let expr = if count > 0 {
            split_expr
                .clone()
                .list()
                .slice(lit(0i64), lit(n))
                .list()
                .join(lit(delim), false)
        } else {
            let len = split_expr.clone().list().len();
            let start = when(len.clone().gt(lit(n)))
                .then(len.clone() - lit(n))
                .otherwise(lit(0i64));
            let slice_len = when(len.clone().gt(lit(n))).then(lit(n)).otherwise(len);
            split_expr
                .list()
                .slice(start, slice_len)
                .list()
                .join(lit(delim), false)
        };
        Self::from_expr(expr, None)
    }

    /// Soundex code (PySpark soundex). Implemented via map UDF (strsim/soundex crates).
    pub fn soundex(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_soundex(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Levenshtein distance to another string (PySpark levenshtein). Implemented via map_many UDF (strsim).
    pub fn levenshtein(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_levenshtein(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// CRC32 checksum of string bytes (PySpark crc32). Implemented via map UDF (crc32fast).
    pub fn crc32(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_crc32(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// XXH64 hash of string (PySpark xxhash64). Implemented via map UDF (twox-hash).
    pub fn xxhash64(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_xxhash64(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// ASCII value of first character (PySpark ascii). Returns Int32.
    pub fn ascii(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_ascii(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Format numeric as string with fixed decimal places (PySpark format_number).
    pub fn format_number(&self, decimals: u32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_format_number(s, decimals)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Int to single-character string (PySpark char / chr). Valid codepoint only.
    pub fn char(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_char(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Alias for char (PySpark chr).
    pub fn chr(&self) -> Column {
        self.char()
    }

    /// Base64 encode string bytes (PySpark base64).
    pub fn base64(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_base64(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Base64 decode to string (PySpark unbase64). Invalid decode → null.
    pub fn unbase64(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_unbase64(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// SHA1 hash of string bytes, return hex string (PySpark sha1).
    pub fn sha1(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_sha1(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// SHA2 hash; bit_length 256, 384, or 512 (PySpark sha2). Default 256.
    pub fn sha2(&self, bit_length: i32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_sha2(s, bit_length)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// MD5 hash of string bytes, return hex string (PySpark md5).
    pub fn md5(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_md5(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Replace substring at 1-based position (PySpark overlay). replace is literal string.
    pub fn overlay(&self, replace: &str, pos: i64, length: i64) -> Column {
        use polars::prelude::*;
        let pos = pos.max(1);
        let replace_len = length.max(0);
        let start_left = 0i64;
        let len_left = (pos - 1).max(0);
        let start_right = (pos - 1 + replace_len).max(0);
        let len_right = 1_000_000i64; // "rest of string"
        let left = self
            .expr()
            .clone()
            .str()
            .slice(lit(start_left), lit(len_left));
        let mid = lit(replace.to_string());
        let right = self
            .expr()
            .clone()
            .str()
            .slice(lit(start_right), lit(len_right));
        let exprs = [left, mid, right];
        let concat_expr = polars::prelude::concat_str(&exprs, "", false);
        Self::from_expr(concat_expr, None)
    }

    // --- Math functions ---

    /// Absolute value (PySpark abs)
    pub fn abs(&self) -> Column {
        Self::from_expr(self.expr().clone().abs(), None)
    }

    /// Ceiling (PySpark ceil)
    pub fn ceil(&self) -> Column {
        use polars::prelude::*;
        // PySpark ceil returns an integral (bigint) type. Cast the ceil result
        // to Int64 so the schema reflects LongType/bigint instead of double.
        let expr = self.expr().clone().ceil().cast(DataType::Int64);
        Self::from_expr(expr, None)
    }

    /// Alias for ceil. PySpark ceiling.
    pub fn ceiling(&self) -> Column {
        self.ceil()
    }

    /// Floor (PySpark floor)
    pub fn floor(&self) -> Column {
        use polars::prelude::*;
        // PySpark floor returns an integral (bigint) type. Cast the floor result
        // to Int64 so the schema reflects LongType/bigint instead of double.
        let expr = self.expr().clone().floor().cast(DataType::Int64);
        Self::from_expr(expr, None)
    }

    /// Round to given decimal places (PySpark round). scale can be negative (e.g. -3 rounds to thousands).
    /// Supports string columns containing numeric values (implicit cast to double then round; parity with PySpark).
    pub fn round(&self, scale: i32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_round(s, scale)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Banker's rounding - round half to even (PySpark bround).
    pub fn bround(&self, scale: i32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_bround(s, scale)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Unary minus (PySpark negate, negative).
    pub fn negate(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone() * lit(-1), None)
    }

    /// Multiply with PySpark-style string/number coercion (used by Python Column operators).
    ///
    /// Both operands are coerced to Double when used from Python; string columns are parsed
    /// as doubles where possible, invalid strings become null.
    pub fn multiply_pyspark(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pyspark_multiply(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Add with PySpark-style string/number coercion (used by Python Column operators).
    /// For string columns, + attempts numeric addition so non-numeric strings become null (issue #1138).
    pub fn add_pyspark(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pyspark_add(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Subtract with PySpark-style string/number coercion (used by Python Column operators).
    pub fn subtract_pyspark(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pyspark_subtract(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Divide with PySpark-style string/number coercion (used by Python Column operators).
    pub fn divide_pyspark(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pyspark_divide(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Modulo with PySpark-style string/number coercion (used by Python Column operators).
    pub fn mod_pyspark(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pyspark_mod(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Multiply by another column or literal (PySpark multiply). Broadcasts scalars.
    pub fn multiply(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone() * other.expr().clone(), None)
    }

    /// Add another column or literal (PySpark +). Broadcasts scalars.
    pub fn add(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone() + other.expr().clone(), None)
    }

    /// Subtract another column or literal (PySpark -). Broadcasts scalars.
    pub fn subtract(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone() - other.expr().clone(), None)
    }

    /// Divide by another column or literal (PySpark /). Broadcasts scalars.
    pub fn divide(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone() / other.expr().clone(), None)
    }

    /// Modulo (PySpark %). Broadcasts scalars.
    pub fn mod_(&self, other: &Column) -> Column {
        Self::from_expr(self.expr().clone() % other.expr().clone(), None)
    }

    /// Square root (PySpark sqrt)
    pub fn sqrt(&self) -> Column {
        Self::from_expr(self.expr().clone().sqrt(), None)
    }

    /// Power (PySpark pow). Exponent can be literal or expression.
    pub fn pow(&self, exp: i64) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().pow(lit(exp)), None)
    }

    /// Power with column or scalar exponent (for __pow__ / col ** other).
    /// Uses float pow so fractional exponents work (#817); handles 0^positive=0 (#863).
    pub fn pow_with(&self, exponent: &Column) -> Column {
        let args = [exponent.expr().clone()];
        let expr = self.expr().clone().map_many(
            move |cols| expect_col(crate::udfs::apply_pow_pyspark(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Alias for pow. PySpark power.
    pub fn power(&self, exp: i64) -> Column {
        self.pow(exp)
    }

    /// Exponential (PySpark exp)
    pub fn exp(&self) -> Column {
        Self::from_expr(self.expr().clone().exp(), None)
    }

    /// Natural logarithm (PySpark log)
    pub fn log(&self) -> Column {
        Self::from_expr(self.expr().clone().log(lit(std::f64::consts::E)), None)
    }

    /// Alias for log. PySpark ln.
    pub fn ln(&self) -> Column {
        self.log()
    }

    /// Sine (radians). PySpark sin.
    pub fn sin(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_sin(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Cosine (radians). PySpark cos.
    pub fn cos(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_cos(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Tangent (radians). PySpark tan.
    pub fn tan(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_tan(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Cotangent: 1/tan (PySpark cot).
    pub fn cot(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_cot(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Cosecant: 1/sin (PySpark csc).
    pub fn csc(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_csc(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Secant: 1/cos (PySpark sec).
    pub fn sec(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_sec(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Arc sine. PySpark asin.
    pub fn asin(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_asin(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Arc cosine. PySpark acos.
    pub fn acos(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_acos(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Arc tangent. PySpark atan.
    pub fn atan(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_atan(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Two-argument arc tangent (y, x) -> angle in radians. PySpark atan2.
    pub fn atan2(&self, x: &Column) -> Column {
        let args = [x.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_atan2(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert radians to degrees. PySpark degrees.
    pub fn degrees(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_degrees(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Alias for degrees. PySpark toDegrees.
    pub fn to_degrees(&self) -> Column {
        self.degrees()
    }

    /// Convert degrees to radians. PySpark radians.
    pub fn radians(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_radians(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Alias for radians. PySpark toRadians.
    pub fn to_radians(&self) -> Column {
        self.radians()
    }

    /// Sign of the number (-1, 0, or 1). PySpark signum.
    pub fn signum(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_signum(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Hyperbolic cosine. PySpark cosh.
    pub fn cosh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_cosh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Hyperbolic sine. PySpark sinh.
    pub fn sinh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_sinh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Hyperbolic tangent. PySpark tanh.
    pub fn tanh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_tanh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Inverse hyperbolic cosine. PySpark acosh.
    pub fn acosh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_acosh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Inverse hyperbolic sine. PySpark asinh.
    pub fn asinh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_asinh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Inverse hyperbolic tangent. PySpark atanh.
    pub fn atanh(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_atanh(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Cube root. PySpark cbrt.
    pub fn cbrt(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_cbrt(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// exp(x) - 1. PySpark expm1.
    pub fn expm1(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_expm1(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// log(1 + x). PySpark log1p.
    pub fn log1p(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_log1p(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Base-10 logarithm. PySpark log10.
    pub fn log10(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_log10(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Base-2 logarithm. PySpark log2.
    pub fn log2(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_log2(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }
    /// Round to nearest integer. PySpark rint.
    pub fn rint(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_rint(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// sqrt(x^2 + y^2). PySpark hypot.
    pub fn hypot(&self, other: &Column) -> Column {
        let xx = self.expr().clone() * self.expr().clone();
        let yy = other.expr().clone() * other.expr().clone();
        Self::from_expr((xx + yy).sqrt(), None)
    }

    /// Cast to the given type (PySpark cast). Fails on invalid conversion.
    pub fn cast_to(&self, type_name: &str) -> Result<Column, String> {
        crate::functions::cast(self, type_name)
    }

    /// Cast to the given type, null on invalid conversion (PySpark try_cast).
    pub fn try_cast_to(&self, type_name: &str) -> Result<Column, String> {
        crate::functions::try_cast(self, type_name)
    }

    /// True where the float value is NaN (PySpark isnan). Non-float columns (e.g. string) return all False.
    pub fn is_nan(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_isnan_pyspark_parity(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Boolean)),
        );
        Self::from_expr(expr, None)
    }

    // --- Datetime functions ---

    /// Extract year from datetime column (PySpark year)
    pub fn year(&self) -> Column {
        let name = format!("year({})", self.name());
        use polars::prelude::*;
        let parsed = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        Self::from_expr(parsed.dt().year().alias(&name), Some(name))
    }

    /// Extract month from datetime column (PySpark month). Returns IntegerType (int) for schema parity (#1402).
    pub fn month(&self) -> Column {
        let name = format!("month({})", self.name());
        use polars::prelude::*;
        let parsed = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        let month_expr = parsed.dt().month().cast(DataType::Int32);
        Self::from_expr(month_expr.alias(&name), Some(name))
    }

    /// Extract day of month from datetime column (PySpark day)
    pub fn day(&self) -> Column {
        Self::from_expr(self.expr().clone().dt().day(), None)
    }

    /// Alias for day. PySpark dayofmonth. Returns IntegerType (int) for schema parity (#1403).
    pub fn dayofmonth(&self) -> Column {
        let name = format!("dayofmonth({})", self.name());
        use polars::prelude::*;
        let parsed = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        let day_expr = parsed.dt().day().cast(DataType::Int32);
        Self::from_expr(day_expr.alias(&name), Some(name))
    }

    /// Extract quarter (1-4) from date/datetime column (PySpark quarter).
    pub fn quarter(&self) -> Column {
        Self::from_expr(self.expr().clone().dt().quarter(), None)
    }

    /// Extract ISO week of year (1-53) (PySpark weekofyear / week).
    pub fn weekofyear(&self) -> Column {
        Self::from_expr(self.expr().clone().dt().week(), None)
    }

    /// Alias for weekofyear (PySpark week).
    pub fn week(&self) -> Column {
        self.weekofyear()
    }

    /// Day of week: 1 = Sunday, 2 = Monday, ..., 7 = Saturday (PySpark dayofweek).
    /// Polars weekday is Mon=1..Sun=7; we convert to Sun=1..Sat=7.
    pub fn dayofweek(&self) -> Column {
        use polars::prelude::*;
        let parsed = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        let w = parsed.dt().weekday().cast(DataType::Int32);
        let dayofweek = ((w % lit(7i32)) + lit(1i32)).cast(DataType::Int32);
        let name = format!("dayofweek({})", self.name());
        Self::from_expr(dayofweek.alias(&name), Some(name))
    }

    /// Day of year (1-366) (PySpark dayofyear).
    pub fn dayofyear(&self) -> Column {
        Self::from_expr(
            self.expr().clone().dt().ordinal_day().cast(DataType::Int32),
            None,
        )
    }

    /// Cast to date (PySpark to_date). Drops time component from datetime/timestamp.
    pub fn to_date(&self) -> Column {
        use polars::prelude::DataType;
        Self::from_expr(self.expr().clone().cast(DataType::Date), None)
    }

    /// Format date/datetime as string (PySpark date_format). Uses chrono strftime format.
    pub fn date_format(&self, format: &str) -> Column {
        Self::from_expr(self.expr().clone().dt().strftime(format), None)
    }

    /// Extract hour from datetime column (PySpark hour). Accepts string timestamp (#403).
    pub fn hour(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_hour(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Extract minute from datetime column (PySpark minute). Accepts string timestamp (#403).
    pub fn minute(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_minute(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Extract second from datetime column (PySpark second). Accepts string timestamp (#403).
    pub fn second(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_second(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Extract field from date/datetime (PySpark extract). field: "year","month","day","hour","minute","second","quarter","week","dayofweek","dayofyear".
    pub fn extract(&self, field: &str) -> Column {
        use polars::prelude::*;
        let e = self.expr().clone();
        let expr = match field.trim().to_lowercase().as_str() {
            "year" => e.dt().year(),
            "month" => e.dt().month(),
            "day" => e.dt().day(),
            "hour" => e.dt().hour(),
            "minute" => e.dt().minute(),
            "second" => e.dt().second(),
            "quarter" => e.dt().quarter(),
            "week" | "weekofyear" => e.dt().week(),
            "dayofweek" | "dow" => {
                let w = e.dt().weekday();
                (w % lit(7i32)) + lit(1i32)
            }
            "dayofyear" | "doy" => e.dt().ordinal_day().cast(DataType::Int32),
            _ => e.dt().year(), // fallback
        };
        Self::from_expr(expr, None)
    }

    /// Timestamp to microseconds since epoch (PySpark unix_micros).
    pub fn unix_micros(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().cast(DataType::Int64), None)
    }

    /// Timestamp to milliseconds since epoch (PySpark unix_millis).
    pub fn unix_millis(&self) -> Column {
        use polars::prelude::*;
        let micros = self.expr().clone().cast(DataType::Int64);
        Self::from_expr(micros / lit(1000i64), None)
    }

    /// Timestamp to seconds since epoch (PySpark unix_seconds).
    pub fn unix_seconds(&self) -> Column {
        use polars::prelude::*;
        let micros = self.expr().clone().cast(DataType::Int64);
        Self::from_expr(micros / lit(1_000_000i64), None)
    }

    /// Weekday name "Mon","Tue",... (PySpark dayname).
    pub fn dayname(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_dayname(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Weekday 0=Mon, 6=Sun (PySpark weekday).
    pub fn weekday(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_weekday(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Add n days to date/datetime column (PySpark date_add).
    pub fn date_add(&self, n: i32) -> Column {
        use polars::prelude::*;
        let date_expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        let dur = duration(DurationArgs::new().with_days(lit(n as i64)));
        let name = format!("date_add({}, {n})", self.name());
        Self::from_expr((date_expr + dur).alias(&name), Some(name))
    }

    /// Subtract n days from date/datetime column (PySpark date_sub).
    pub fn date_sub(&self, n: i32) -> Column {
        use polars::prelude::*;
        let date_expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_string_to_date_format(s, None, false)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        let dur = duration(DurationArgs::new().with_days(lit(n as i64)));
        let name = format!("date_sub({}, {n})", self.name());
        Self::from_expr((date_expr - dur).alias(&name), Some(name))
    }

    /// Number of days between two date/datetime columns (PySpark datediff). (end - start).
    pub fn datediff(&self, other: &Column) -> Column {
        use polars::prelude::*;
        let start = self.expr().clone().cast(DataType::Date);
        let end = other.expr().clone().cast(DataType::Date);
        // Cast to Int32 so schema maps to PySpark IntegerType instead of LongType.
        let expr = (end - start).dt().total_days(false).cast(DataType::Int32);
        Self::from_expr(expr, None)
    }

    /// Last day of the month for date/datetime column (PySpark last_day).
    pub fn last_day(&self) -> Column {
        Self::from_expr(self.expr().clone().dt().month_end(), None)
    }

    /// Add amount of unit to timestamp (PySpark timestampadd). unit: DAY, HOUR, MINUTE, SECOND, etc.
    pub fn timestampadd(&self, unit: &str, amount: &Column) -> Column {
        use polars::prelude::*;
        let ts = self.expr().clone();
        let amt = amount.expr().clone().cast(DataType::Int64);
        let dur = match unit.trim().to_uppercase().as_str() {
            "DAY" | "DAYS" => duration(DurationArgs::new().with_days(amt)),
            "HOUR" | "HOURS" => duration(DurationArgs::new().with_hours(amt)),
            "MINUTE" | "MINUTES" => duration(DurationArgs::new().with_minutes(amt)),
            "SECOND" | "SECONDS" => duration(DurationArgs::new().with_seconds(amt)),
            "WEEK" | "WEEKS" => duration(DurationArgs::new().with_weeks(amt)),
            _ => duration(DurationArgs::new().with_days(amt)),
        };
        Self::from_expr(ts + dur, None)
    }

    /// Difference between timestamps in given unit (PySpark timestampdiff). unit: DAY, HOUR, MINUTE, SECOND.
    pub fn timestampdiff(&self, unit: &str, other: &Column) -> Column {
        let start = self.expr().clone();
        let end = other.expr().clone();
        let diff = end - start;
        let expr = match unit.trim().to_uppercase().as_str() {
            "HOUR" | "HOURS" => diff.dt().total_hours(false),
            "MINUTE" | "MINUTES" => diff.dt().total_minutes(false),
            "SECOND" | "SECONDS" => diff.dt().total_seconds(false),
            "DAY" | "DAYS" => diff.dt().total_days(false),
            _ => diff.dt().total_days(false),
        };
        Self::from_expr(expr, None)
    }

    /// Interpret timestamp as UTC, convert to target timezone (PySpark from_utc_timestamp).
    pub fn from_utc_timestamp(&self, tz: &str) -> Column {
        let tz = tz.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_from_utc_timestamp(s, &tz)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Interpret timestamp as in tz, convert to UTC (PySpark to_utc_timestamp).
    pub fn to_utc_timestamp(&self, tz: &str) -> Column {
        let tz = tz.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_to_utc_timestamp(s, &tz)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Truncate date/datetime to unit. PySpark date_trunc/trunc use "year", "month", "day", "hour";
    /// Polars dt.truncate expects duration strings like "1y", "1mo", "1d", "1h". We map PySpark names.
    /// Uses a UDF path so string columns (e.g. from createDataFrame with Python datetime) are coerced to datetime then truncated.
    pub fn trunc(&self, format: &str) -> Column {
        use polars::prelude::*;
        let polars_duration = pyspark_trunc_format_to_polars_duration(format);
        let duration = polars_duration.clone();
        let expr = self.expr().clone().map(
            move |c| expect_col(crate::udfs::apply_date_trunc(c, &duration)),
            |_schema, field| {
                Ok(Field::new(
                    field.name().clone(),
                    DataType::Datetime(TimeUnit::Microseconds, None),
                ))
            },
        );
        Self::from_expr(expr, Some(self.name().to_string()))
    }

    /// Add n months to date/datetime column (PySpark add_months). Month-aware.
    pub fn add_months(&self, n: i32) -> Column {
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_add_months(col, n)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        Self::from_expr(expr, None)
    }

    /// Number of months between end and start dates, as fractional (PySpark months_between).
    /// When round_off is true, rounds to 8 decimal places (PySpark default).
    pub fn months_between(&self, start: &Column, round_off: bool) -> Column {
        let args = [start.expr().clone()];
        let expr = self.expr().clone().map_many(
            move |cols| expect_col(crate::udfs::apply_months_between(cols, round_off)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Next date that is the given day of week (e.g. "Mon", "Tue") (PySpark next_day).
    pub fn next_day(&self, day_of_week: &str) -> Column {
        let day = day_of_week.to_string();
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_next_day(col, &day)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        Self::from_expr(expr, None)
    }

    /// Parse string timestamp to seconds since epoch (PySpark unix_timestamp).
    pub fn unix_timestamp(&self, format: Option<&str>) -> Column {
        let fmt = format.map(String::from);
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_unix_timestamp(col, fmt.as_deref())),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert seconds since epoch to formatted string (PySpark from_unixtime).
    pub fn from_unixtime(&self, format: Option<&str>) -> Column {
        let fmt = format.map(String::from);
        let expr = self.expr().clone().map(
            move |col| expect_col(crate::udfs::apply_from_unixtime(col, fmt.as_deref())),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Convert seconds since epoch to timestamp (PySpark timestamp_seconds).
    pub fn timestamp_seconds(&self) -> Column {
        let expr = (self.expr().clone().cast(DataType::Int64) * lit(1_000_000i64))
            .cast(DataType::Datetime(TimeUnit::Microseconds, None));
        Self::from_expr(expr, None)
    }

    /// Convert milliseconds since epoch to timestamp (PySpark timestamp_millis).
    pub fn timestamp_millis(&self) -> Column {
        let expr = (self.expr().clone().cast(DataType::Int64) * lit(1000i64))
            .cast(DataType::Datetime(TimeUnit::Microseconds, None));
        Self::from_expr(expr, None)
    }

    /// Convert microseconds since epoch to timestamp (PySpark timestamp_micros).
    pub fn timestamp_micros(&self) -> Column {
        let expr = self
            .expr()
            .clone()
            .cast(DataType::Int64)
            .cast(DataType::Datetime(TimeUnit::Microseconds, None));
        Self::from_expr(expr, None)
    }

    /// Date to days since 1970-01-01 (PySpark unix_date).
    pub fn unix_date(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_unix_date(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int32)),
        );
        Self::from_expr(expr, None)
    }

    /// Days since epoch to date (PySpark date_from_unix_date).
    pub fn date_from_unix_date(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_date_from_unix_date(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Date)),
        );
        Self::from_expr(expr, None)
    }

    /// Positive modulus (PySpark pmod). Column method: pmod(self, other).
    pub fn pmod(&self, divisor: &Column) -> Column {
        let args = [divisor.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_pmod(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Float64)),
        );
        Self::from_expr(expr, None)
    }

    /// Factorial n! for n in 0..=20 (PySpark factorial).
    pub fn factorial(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_factorial(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    // --- Window functions ---

    /// Apply window partitioning. Returns a new Column with `.over(partition_by)`.
    /// Use after rank(), dense_rank(), row_number(), lag(), lead().
    pub fn over(&self, partition_by: &[&str]) -> Column {
        let partition_exprs: Vec<Expr> = if partition_by.is_empty() {
            vec![lit(1i32)]
        } else {
            partition_by.iter().map(|s| col(*s)).collect()
        };
        Self::from_expr(self.expr().clone().over(partition_exprs), None)
    }

    /// Apply window with optional order-by for running aggregates (e.g. sum, count).
    /// `order_by_encoded`: e.g. ["value"] for asc, ["-value"] for desc.
    /// When `use_running_aggregate` is true and we have `source_for_running`, use cum_sum for running semantics.
    /// When `first_last_value` is Some and order is present, use order-sensitive first/last (PySpark #1145).
    /// When `is_full_partition_frame` is false and we have last_value (first_last_value.is_last), use current-row semantics (PySpark default frame).
    pub fn over_window(
        &self,
        partition_by: &[&str],
        order_by_encoded: &[String],
        use_running_aggregate: bool,
        is_full_partition_frame: bool,
    ) -> Result<Column, PolarsError> {
        // PySpark does not support countDistinct().over(); approx_count_distinct().over() is allowed (#1218).
        if expr_is_or_contains_n_unique(self.expr()) && self.name.starts_with("count_distinct(") {
            return Err(PolarsError::InvalidOperation(
                "Distinct window functions are not supported".into(),
            ));
        }
        let partition_exprs: Vec<Expr> = if partition_by.is_empty() {
            vec![lit(1i32)]
        } else {
            partition_by.iter().map(|s| col(*s)).collect()
        };

        // last_value (or F.last().over()) with orderBy and default frame: "last" = current row (PySpark default frame).
        if let Some(ref fl) = self.first_last_value {
            if fl.is_last && !order_by_encoded.is_empty() && !is_full_partition_frame {
                let mut order_exprs: Vec<Expr> = Vec::with_capacity(order_by_encoded.len());
                let mut descending_multi: Vec<bool> = Vec::with_capacity(order_by_encoded.len());
                for s in order_by_encoded.iter() {
                    let s = s.trim();
                    let (name, descending) = if let Some(stripped) = s.strip_prefix('-') {
                        (stripped.trim(), true)
                    } else {
                        (s.trim(), false)
                    };
                    order_exprs.push(col(name));
                    descending_multi.push(descending);
                }
                let default_opts = SortOptions {
                    descending: descending_multi.first().copied().unwrap_or(false),
                    nulls_last: descending_multi.first().copied().unwrap_or(false),
                    ..Default::default()
                };
                let expr = fl.value_expr.clone().over_with_options(
                    Some(partition_exprs),
                    Some((order_exprs, default_opts)),
                    WindowMapping::default(),
                )?;
                return Ok(Self::from_expr(expr, None));
            }
        }

        let base_expr = if use_running_aggregate {
            if let Some(ref src) = self.source_for_running_mean {
                // Running mean in window order: cum_sum / cum_count on same column so order applies (#1241).
                let sum_expr = col(src).cast(DataType::Float64).cum_sum(false);
                let count_expr = col(src).cum_count(false).cast(DataType::Float64);
                sum_expr / count_expr
            } else if let Some(ref src) = self.source_for_running {
                // Running sum in window order. Cast to Float64 so string columns work
                // (PySpark parity, issue #393). Non-running sums still use native dtype.
                col(src).cast(DataType::Float64).cum_sum(false)
            } else if let Some(ref src) = self.source_for_running_count {
                // Running count in window order (PySpark count().over(order); #1218).
                col(src).cum_count(false).cast(DataType::Int64)
            } else {
                self.expr().clone()
            }
        } else {
            // Non-running aggregates over a window: use the aggregate expression as-is.
            // For sum() on string columns, this expression already includes a cast to Float64
            // (PySpark parity, issue #393).
            self.expr().clone()
        };
        let expr = if order_by_encoded.is_empty() {
            base_expr.over(partition_exprs)
        } else {
            // Build order exprs and sort options. Polars over_with_options uses (order_exprs, sort_options).
            // Use column as-is (no String cast) so numeric columns sort 80,90,100 (issue #1052).
            let mut order_exprs: Vec<Expr> = Vec::with_capacity(order_by_encoded.len());
            let mut descending_multi: Vec<bool> = Vec::with_capacity(order_by_encoded.len());
            for s in order_by_encoded.iter() {
                let s = s.trim();
                let (name, descending) = if let Some(stripped) = s.strip_prefix('-') {
                    (stripped.trim(), true)
                } else {
                    (s, false)
                };
                order_exprs.push(col(name));
                descending_multi.push(descending);
            }
            // Single sort_options for the window: use first column's direction. Polars may use this for the whole order.
            let default_opts = SortOptions {
                descending: descending_multi.first().copied().unwrap_or(false),
                nulls_last: descending_multi.first().copied().unwrap_or(false),
                ..Default::default()
            };
            base_expr.over_with_options(
                Some(partition_exprs),
                Some((order_exprs, default_opts)),
                WindowMapping::default(),
            )?
        };
        Ok(Self::from_expr(expr, None))
    }

    /// Rank (with ties, gaps). Use with `.over(partition_by)`.
    pub fn rank(&self, descending: bool) -> Column {
        let opts = RankOptions {
            method: RankMethod::Min,
            descending,
        };
        Self::from_expr(self.expr().clone().rank(opts, None), None)
    }

    /// Dense rank (no gaps). Use with `.over(partition_by)`.
    pub fn dense_rank(&self, descending: bool) -> Column {
        let opts = RankOptions {
            method: RankMethod::Dense,
            descending,
        };
        Self::from_expr(self.expr().clone().rank(opts, None), None)
    }

    /// Row number (1, 2, 3 by this column's order). Use with `.over(partition_by)`.
    /// Nulls in the order column get a rank (PySpark parity: nulls last for asc, nulls first for desc).
    pub fn row_number(&self, descending: bool) -> Column {
        use polars::prelude::*;
        let opts = RankOptions {
            method: RankMethod::Ordinal,
            descending,
        };
        // Fill nulls so rank returns a value: ascending -> nulls last (fill with inf); descending -> nulls first (fill with -inf)
        let rank_expr = self
            .expr()
            .clone()
            .cast(DataType::Float64)
            .fill_null(lit(if descending {
                f64::NEG_INFINITY
            } else {
                f64::INFINITY
            }))
            .rank(opts, None);
        Self::from_expr(rank_expr, None)
    }

    /// Row number with explicit multi-column order (PySpark Window.orderBy([...]) parity).
    /// Uses over_with_options; multi-column order is done via a struct sort so Type then Score then Name is respected.
    pub fn row_number_over(
        partition_by: &[&str],
        order_by_encoded: &[String],
    ) -> Result<Column, PolarsError> {
        use polars::prelude::*;
        if order_by_encoded.is_empty() {
            return Err(PolarsError::InvalidOperation(
                "row_number_over: order_by_encoded cannot be empty".into(),
            ));
        }
        let partition_exprs: Vec<Expr> = if partition_by.is_empty() {
            vec![lit(1i32)]
        } else {
            partition_by.iter().map(|s| col(*s)).collect()
        };
        // Parse encoded order: "-col" = descending, "col" = ascending. Trim whitespace (#1241).
        fn parse_order_key(s: &str) -> (&str, bool) {
            let s = s.trim();
            let descending = s.starts_with('-');
            let name = if descending {
                s.trim_start_matches('-').trim()
            } else {
                s
            };
            (name, descending)
        }
        let all_asc = order_by_encoded.iter().all(|s| !s.trim().starts_with('-'));
        // Row number = ordinal rank of the order key within partition. For multi-column mixed asc/desc, use struct with negated desc columns (#1241).
        let rank_expr = if order_by_encoded.len() == 1 {
            let (first_name, first_desc) = parse_order_key(order_by_encoded[0].trim());
            // For descending, rank by -col so ascending rank gives high values rank 1 (#1241).
            let order_col = col(first_name)
                .cast(DataType::Float64)
                .fill_null(lit(if first_desc {
                    f64::NEG_INFINITY
                } else {
                    f64::INFINITY
                }));
            let rank_input = if first_desc {
                order_col.neg()
            } else {
                order_col
            };
            let opts = RankOptions {
                method: RankMethod::Ordinal,
                descending: false,
            };
            rank_input.rank(opts, None)
        } else if all_asc {
            let struct_fields: Vec<Expr> = order_by_encoded
                .iter()
                .map(|s| col(parse_order_key(s).0))
                .collect();
            let opts = RankOptions {
                method: RankMethod::Ordinal,
                descending: false,
            };
            as_struct(struct_fields).rank(opts, None)
        } else {
            // Mixed asc/desc: rank by struct with desc columns negated so ascending struct sort gives correct order (#1241).
            let struct_fields: Vec<Expr> = order_by_encoded
                .iter()
                .map(|s| {
                    let (name, desc) = parse_order_key(s);
                    if desc {
                        (col(name)
                            .cast(DataType::Float64)
                            .fill_null(lit(f64::NEG_INFINITY)))
                        .neg()
                    } else {
                        col(name)
                            .cast(DataType::Float64)
                            .fill_null(lit(f64::INFINITY))
                    }
                })
                .collect();
            let opts = RankOptions {
                method: RankMethod::Ordinal,
                descending: false,
            };
            as_struct(struct_fields).rank(opts, None)
        };
        let expr = rank_expr.over(partition_exprs);
        Ok(Self::from_expr(expr, None))
    }

    /// Lag: value from n rows before. Use with `.over(partition_by)`.
    pub fn lag(&self, n: i64) -> Column {
        Self::from_expr(self.expr().clone().shift(polars::prelude::lit(n)), None)
    }

    /// Lead: value from n rows after. Use with `.over(partition_by)`.
    pub fn lead(&self, n: i64) -> Column {
        Self::from_expr(self.expr().clone().shift(polars::prelude::lit(-n)), None)
    }

    /// First value in partition (PySpark first_value). Use with `.over(partition_by)`.
    /// When the window has orderBy, over_window uses order-sensitive semantics (#1145).
    pub fn first_value(&self) -> Column {
        let value_expr = self.expr().clone();
        Column {
            name: "first_value".to_string(),
            expr: value_expr.clone().first(),
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: Some(FirstLastValue {
                value_expr,
                is_last: false,
            }),
            source_for_running_count: None,
        }
    }

    /// Last value in partition (PySpark last_value). Use with `.over(partition_by)`.
    /// When the window has orderBy, over_window uses order-sensitive semantics (#1145).
    pub fn last_value(&self) -> Column {
        let value_expr = self.expr().clone();
        Column {
            name: "last_value".to_string(),
            expr: value_expr.clone().last(),
            is_array_expr: false,
            deferred: None,
            udf_call: None,
            source_for_running: None,
            source_for_running_mean: None,
            first_last_value: Some(FirstLastValue {
                value_expr,
                is_last: true,
            }),
            source_for_running_count: None,
        }
    }

    /// Percent rank in partition: (rank - 1) / (count - 1). Window is applied; do not call .over() again.
    pub fn percent_rank(&self, partition_by: &[&str], descending: bool) -> Column {
        use polars::prelude::*;
        let partition_exprs: Vec<Expr> = partition_by.iter().map(|s| col(*s)).collect();
        let opts = RankOptions {
            method: RankMethod::Min,
            descending,
        };
        let rank_expr = self
            .expr()
            .clone()
            .rank(opts, None)
            .over(partition_exprs.clone());
        let count_expr = self.expr().clone().count().over(partition_exprs.clone());
        let rank_f = (rank_expr - lit(1i64)).cast(DataType::Float64);
        let count_f = (count_expr - lit(1i64)).cast(DataType::Float64);
        // Avoid division by zero: single-row partition -> 0.0 (PySpark parity)
        // When count=2, count_f=1.0; use gt(0) not gt(1) so we compute (rank-1)/1 correctly
        let pct = when(count_f.clone().gt(lit(0.0)))
            .then(rank_f / count_f)
            .otherwise(lit(0.0));
        Self::from_expr(pct, None)
    }

    /// Cumulative distribution in partition: row_number / count. Window is applied; do not call .over() again.
    pub fn cume_dist(&self, partition_by: &[&str], descending: bool) -> Column {
        use polars::prelude::*;
        let partition_exprs: Vec<Expr> = partition_by.iter().map(|s| col(*s)).collect();
        let opts = RankOptions {
            method: RankMethod::Ordinal,
            descending,
        };
        let row_num = self
            .expr()
            .clone()
            .rank(opts, None)
            .over(partition_exprs.clone());
        let count_expr = self.expr().clone().count().over(partition_exprs.clone());
        // Avoid division by zero when partition is empty
        let count_f = count_expr.clone().cast(DataType::Float64);
        let cume = when(count_f.clone().eq(lit(0.0)))
            .then(lit(0.0))
            .otherwise(row_num.cast(DataType::Float64) / count_f);
        Self::from_expr(cume, None)
    }

    /// Ntile: bucket 1..n by rank within partition (ceil(rank * n / count)). Window is applied; do not call .over() again.
    pub fn ntile(&self, n: u32, partition_by: &[&str], descending: bool) -> Column {
        use polars::prelude::*;
        let partition_exprs: Vec<Expr> = if partition_by.is_empty() {
            vec![lit(1i32)]
        } else {
            partition_by.iter().map(|s| col(*s)).collect()
        };
        let opts = RankOptions {
            method: RankMethod::Ordinal,
            descending,
        };
        let rank_expr = self
            .expr()
            .clone()
            .rank(opts, None)
            .over(partition_exprs.clone());
        let count_expr = self.expr().clone().count().over(partition_exprs.clone());
        let n_expr = lit(n as f64);
        let rank_f = rank_expr.cast(DataType::Float64);
        let count_f = count_expr.cast(DataType::Float64);
        // Avoid division by zero when partition is empty: use bucket 1.
        // PySpark parity: ntile(n) uses floor((rank - 1) * n / count) + 1 so that
        // the first buckets get the extra rows when count % n != 0.
        let bucket = when(count_f.clone().eq(lit(0.0))).then(lit(1.0)).otherwise(
            ((rank_f.clone() - lit(1.0)) * n_expr.clone() / count_f.clone()).floor() + lit(1.0),
        );
        let clamped = bucket.clip(lit(1.0), lit(n as f64));
        Self::from_expr(clamped.cast(DataType::Int32), None)
    }

    /// Nth value in partition by order (1-based n). Returns a Column with window already applied; do not call .over() again.
    pub fn nth_value(&self, n: i64, partition_by: &[&str], descending: bool) -> Column {
        use polars::prelude::*;
        let partition_exprs: Vec<Expr> = partition_by.iter().map(|s| col(*s)).collect();
        let opts = RankOptions {
            method: RankMethod::Ordinal,
            descending,
        };
        let rank_expr = self
            .expr()
            .clone()
            .rank(opts, None)
            .over(partition_exprs.clone());
        let cond_col = Self::from_expr(rank_expr.eq(lit(n)), None);
        let null_col = Self::from_expr(lit(NULL), None);
        let value_col = Self::from_expr(self.expr().clone(), None);
        let when_expr = crate::functions::when(&cond_col)
            .then(&value_col)
            .otherwise(&null_col)
            .into_expr();
        let windowed = when_expr.max().over(partition_exprs);
        Self::from_expr(windowed, None)
    }

    /// Number of elements in list (PySpark size / array_size). Returns Int32.
    pub fn array_size(&self) -> Column {
        use polars::prelude::*;
        Self::from_expr(
            self.expr().clone().list().len().cast(DataType::Int32),
            Some("size".to_string()),
        )
    }

    /// Cardinality: number of elements in array/list (PySpark cardinality). Alias for array_size.
    pub fn cardinality(&self) -> Column {
        self.array_size()
    }

    /// Check if list contains value (PySpark array_contains).
    pub fn array_contains(&self, value: Expr) -> Column {
        use polars::prelude::*;
        let args = [value];
        let base_expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_contains(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Boolean)),
        );
        // Ensure PySpark parity for null arrays: array_contains(null, x) -> null.
        let is_null = self.expr().clone().is_null();
        let expr = when(is_null)
            .then(lit(NULL))
            .otherwise(base_expr)
            .cast(DataType::Boolean);
        Self::from_expr(expr, None)
    }

    /// Join list of strings with separator (PySpark array_join).
    pub fn array_join(&self, separator: &str) -> Column {
        use polars::prelude::*;
        // PySpark array_join accepts arrays of any element type and stringifies elements.
        // Cast elements to String via list.eval before joining.
        let elem_to_str = col("").cast(DataType::String);
        let list_expr = self.expr().clone().list().eval(elem_to_str);
        let joined = list_expr.list().join(lit(separator.to_string()), false);
        Self::from_expr(joined, None)
    }

    /// Maximum element in list (PySpark array_max).
    pub fn array_max(&self) -> Column {
        Self::from_expr(self.expr().clone().list().max(), None)
    }

    /// Minimum element in list (PySpark array_min).
    pub fn array_min(&self) -> Column {
        Self::from_expr(self.expr().clone().list().min(), None)
    }

    /// Get element at 1-based index (PySpark element_at). Returns null if out of bounds.
    pub fn element_at(&self, index: i64) -> Column {
        use polars::prelude::*;
        // PySpark uses 1-based indexing; Polars uses 0-based. index 1 -> get(0).
        let idx = if index >= 1 { index - 1 } else { index };
        Self::from_expr(self.expr().clone().list().get(lit(idx), true), None)
    }

    /// Get element at 0-based index (PySpark Column.getItem). Returns null if out of bounds.
    pub fn get_item(&self, index: i64) -> Column {
        use polars::prelude::*;
        Self::from_expr(self.expr().clone().list().get(lit(index), true), None)
    }

    /// Get struct field by name (PySpark Column.getField).
    pub fn get_field(&self, name: &str) -> Column {
        Self::from_expr(
            self.expr().clone().struct_().field_by_name(name),
            Some(name.to_string()),
        )
    }

    /// Add or replace a struct field (PySpark Column.withField).
    ///
    /// Panics if the column is not a struct type. If you need error handling, use
    /// [`Column::try_with_field`].
    pub fn with_field(&self, name: &str, value: &Column) -> Column {
        self.try_with_field(name, value)
            .expect("with_field: column must be struct type")
    }

    /// Add or replace a struct field (PySpark Column.withField), returning an error if the
    /// column is not a struct type. Uses a map_many UDF so we don't rely on Polars "*" wildcard
    /// (removed in 0.53). Schema callback returns the extended struct so collected rows include
    /// the new field (issue #1066).
    pub fn try_with_field(
        &self,
        name: &str,
        value: &Column,
    ) -> Result<Column, polars::error::PolarsError> {
        use polars::prelude::PlSmallStr;
        let field_name = name.to_string();
        let field_name_schema = field_name.clone();
        let args = [value.expr().clone()];
        let expr = self.expr().clone().map_many(
            move |cols| {
                // map_many passes [self, ...args]: self is struct, args[0] is value
                expect_col(crate::udfs::apply_struct_with_field(
                    cols[0].clone(),
                    cols[1].clone(),
                    &field_name,
                ))
            },
            &args,
            move |_schema, fields| {
                let struct_field = &fields[0];
                let struct_dtype = struct_field.dtype();
                let inner: &[Field] = match struct_dtype {
                    DataType::Struct(f) => f.as_ref(),
                    _ => return Ok(struct_field.clone()),
                };
                let value_dtype = fields[1].dtype().clone();
                let known_value_dtype = if value_dtype.is_known() {
                    value_dtype
                } else if let DataType::Unknown(uk) = &value_dtype {
                    uk.materialize().unwrap_or(DataType::String)
                } else {
                    DataType::String
                };
                let mut new_fields: Vec<Field> = inner.to_vec();
                let mut replaced = false;
                for f in &mut new_fields {
                    if f.name.as_str() == field_name_schema {
                        // When replacing, use value's dtype so type changes (e.g. int -> string) and
                        // new complex types (e.g. array) are reflected in schema (Issue #1263).
                        let dtype = if known_value_dtype.is_known() {
                            known_value_dtype.clone()
                        } else if f.dtype.is_known() {
                            f.dtype.clone()
                        } else {
                            DataType::String
                        };
                        *f = Field::new(PlSmallStr::from(f.name.as_str()), dtype);
                        replaced = true;
                        break;
                    }
                }
                if !replaced {
                    new_fields.push(Field::new(
                        PlSmallStr::from(field_name_schema.as_str()),
                        known_value_dtype,
                    ));
                }
                let out_dtype = DataType::Struct(new_fields);
                Ok(Field::new(struct_field.name().clone(), out_dtype))
            },
        );
        Ok(Self::from_expr(expr, None))
    }

    /// Sort list elements (PySpark array_sort). Ascending, nulls last.
    pub fn array_sort(&self) -> Column {
        use polars::prelude::SortOptions;
        let opts = SortOptions {
            descending: false,
            nulls_last: true,
            ..Default::default()
        };
        Self::from_expr(self.expr().clone().list().sort(opts), None)
    }

    /// Distinct elements in list (PySpark array_distinct). Preserves first-occurrence order.
    pub fn array_distinct(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_array_distinct_first_order(s)),
            |_schema, field| {
                let new_name = format!("array_distinct({})", field.name());
                Ok(Field::new(new_name.into(), field.dtype().clone()))
            },
        );
        Self::from_expr(expr, None)
    }

    /// Mode aggregation - most frequent value (PySpark mode).
    /// Uses value_counts sorted by count descending, then first.
    pub fn mode(&self) -> Column {
        // value_counts(sort=true, parallel=false, name="count", normalize=false)
        // puts highest count first; first() gives the mode
        // Struct has "count" and value field; field 0 is typically the value
        let vc = self
            .expr()
            .clone()
            .value_counts(true, false, "count", false);
        let first_struct = vc.first();
        let val_expr = first_struct.struct_().field_by_index(0);
        Self::from_expr(val_expr, Some("mode".to_string()))
    }

    /// Slice list from start with optional length (PySpark slice). 1-based start.
    pub fn array_slice(&self, start: i64, length: Option<i64>) -> Column {
        use polars::prelude::*;
        let start_expr = lit((start - 1).max(0)); // 1-based to 0-based
        let length_expr = length.map(lit).unwrap_or_else(|| lit(i64::MAX));
        Self::from_expr(
            self.expr().clone().list().slice(start_expr, length_expr),
            None,
        )
    }

    /// Explode list into one row per element (PySpark explode).
    pub fn explode(&self) -> Column {
        use polars::prelude::ExplodeOptions;
        Self::from_expr(
            self.expr().clone().explode(ExplodeOptions {
                empty_as_null: false,
                keep_nulls: false,
            }),
            None,
        )
    }

    /// Explode list; null/empty produces one row with null (PySpark explode_outer).
    pub fn explode_outer(&self) -> Column {
        use polars::prelude::ExplodeOptions;
        Self::from_expr(
            self.expr().clone().explode(ExplodeOptions {
                empty_as_null: true,
                keep_nulls: true,
            }),
            None,
        )
    }

    /// Posexplode with null preservation (PySpark posexplode_outer).
    ///
    /// Implementation detail:
    /// - Build a list of structs per row: [{pos, col}, {pos, col}, ...].
    /// - Explode that single list column once, then project struct fields "pos" and "col".
    ///   This avoids applying two separate explode() expressions on the same list column,
    ///   which can lead to length mismatches when both position and value are selected
    ///   in the same DataFrame.select call.
    pub fn posexplode_outer(&self) -> (Column, Column) {
        use polars::prelude::{ExplodeOptions, as_struct};

        let opts = ExplodeOptions {
            empty_as_null: true,
            keep_nulls: true,
        };

        // In list.eval context, col("") is the current list element. cum_count(false)
        // yields 1-based positions, so subtract 1 for 0-based PySpark posexplode parity.
        let pos_inner = (col("").cum_count(false) - lit(1i64)).alias("pos");
        let val_inner = col("").alias("col");
        let struct_expr = as_struct(vec![pos_inner, val_inner]);

        // list of structs [{pos, col}, ...] per input row
        let list_struct_expr = self.expr().clone().list().eval(struct_expr);
        // explode once to get a struct column with one row per element (or null/empty handling via opts)
        let struct_exploded = list_struct_expr.explode(opts);

        let pos_expr = struct_exploded.clone().struct_().field_by_name("pos");
        let val_expr = struct_exploded.struct_().field_by_name("col");

        (
            Self::from_expr(pos_expr, Some("pos".to_string())),
            Self::from_expr(val_expr, Some("col".to_string())),
        )
    }

    /// Zip two arrays element-wise into array of structs (PySpark arrays_zip).
    pub fn arrays_zip(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_arrays_zip(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// True if two arrays have any element in common (PySpark arrays_overlap).
    pub fn arrays_overlap(&self, other: &Column) -> Column {
        use polars::prelude::*;

        let args = [other.expr().clone()];
        let base_expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_arrays_overlap(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Boolean)),
        );

        // PySpark parity: arrays_overlap(null_array, other) and arrays_overlap(array, null_array)
        // both yield null, not false. This also drives Column-arg array_contains.
        let is_null = self
            .expr()
            .clone()
            .is_null()
            .or(other.expr().clone().is_null());
        let expr = polars::prelude::when(is_null)
            .then(lit(NULL))
            .otherwise(base_expr)
            .cast(DataType::Boolean);

        Self::from_expr(expr, None)
    }

    /// Collect to array (PySpark array_agg). Alias for implode in group context.
    pub fn array_agg(&self) -> Column {
        Self::from_expr(self.expr().clone().implode(), None)
    }

    /// 1-based index of first occurrence of value in list, or 0 if not found (PySpark array_position).
    /// Uses Polars list.eval with col("") as element (requires polars list_eval feature).
    pub fn array_position(&self, value: Expr) -> Column {
        use polars::prelude::{DataType, NULL};
        // In list.eval context, col("") refers to the current list element.
        let cond = Self::from_expr(col("").eq(value), None);
        let then_val = Self::from_expr(col("").cum_count(false), None);
        let else_val = Self::from_expr(lit(NULL), None);
        let idx_expr = crate::functions::when(&cond)
            .then(&then_val)
            .otherwise(&else_val)
            .into_expr();
        let list_expr = self
            .expr()
            .clone()
            .list()
            .eval(idx_expr)
            .list()
            .min()
            .fill_null(lit(0i64))
            .cast(DataType::Int64);
        Self::from_expr(list_expr, Some("array_position".to_string()))
    }

    /// Remove null elements from list (PySpark array_compact). Preserves order.
    pub fn array_compact(&self) -> Column {
        let list_expr = self.expr().clone().list().drop_nulls();
        Self::from_expr(list_expr, None)
    }

    /// New list with all elements equal to value removed (PySpark array_remove).
    /// Uses list.eval + drop_nulls (requires polars list_eval and list_drop_nulls).
    pub fn array_remove(&self, value: Expr) -> Column {
        use polars::prelude::NULL;
        // when(element != value) then element else null; then drop_nulls.
        let cond = Self::from_expr(col("").neq(value), None);
        let then_val = Self::from_expr(col(""), None);
        let else_val = Self::from_expr(lit(NULL), None);
        let elem_neq = crate::functions::when(&cond)
            .then(&then_val)
            .otherwise(&else_val)
            .into_expr();
        let list_expr = self
            .expr()
            .clone()
            .list()
            .eval(elem_neq)
            .list()
            .drop_nulls();
        Self::from_expr(list_expr, None)
    }

    /// Repeat each element n times (PySpark array_repeat). Implemented via map UDF.
    pub fn array_repeat(&self, n: i64) -> Column {
        let expr = self.expr().clone().map(
            move |c| expect_col(crate::udfs::apply_array_repeat(c, n)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Flatten list of lists to one list (PySpark flatten). Implemented via map UDF.
    pub fn array_flatten(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_array_flatten(s)),
            |_schema, field| Ok(field.clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Append element to end of list (PySpark array_append).
    pub fn array_append(&self, elem: &Column) -> Column {
        let args = [elem.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_append(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Prepend element to start of list (PySpark array_prepend).
    pub fn array_prepend(&self, elem: &Column) -> Column {
        let args = [elem.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_prepend(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Insert element at 1-based position (PySpark array_insert).
    pub fn array_insert(&self, pos: &Column, elem: &Column) -> Column {
        let args = [pos.expr().clone(), elem.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_insert(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Elements in first array not in second (PySpark array_except).
    pub fn array_except(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_except(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Elements in both arrays (PySpark array_intersect).
    pub fn array_intersect(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_intersect(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Distinct elements from both arrays (PySpark array_union).
    pub fn array_union(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_array_union(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Zip two arrays element-wise with merge function (PySpark zip_with). Shorter array padded with null.
    /// Merge Expr uses col("").struct_().field_by_name("left") and field_by_name("right").
    pub fn zip_with(&self, other: &Column, merge: Expr) -> Column {
        let args = [other.expr().clone()];
        let zip_expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_zip_arrays_to_struct(cols)),
            &args,
            |_schema, fields| {
                let left_inner = match &fields[0].dtype {
                    DataType::List(inner) => *inner.clone(),
                    _ => DataType::Unknown(Default::default()),
                };
                let right_inner = match fields.get(1).map(|f| &f.dtype) {
                    Some(DataType::List(inner)) => *inner.clone(),
                    _ => DataType::Unknown(Default::default()),
                };
                let struct_dtype = DataType::Struct(vec![
                    Field::new("left".into(), left_inner),
                    Field::new("right".into(), right_inner),
                ]);
                Ok(Field::new(
                    fields[0].name().clone(),
                    DataType::List(Box::new(struct_dtype)),
                ))
            },
        );
        let list_expr = zip_expr.list().eval(merge);
        Self::from_expr(list_expr, None)
    }

    /// True if any list element satisfies the predicate (PySpark exists). Uses list.eval(pred).list().any().
    pub fn array_exists(&self, predicate: Expr) -> Column {
        let pred_expr = self.expr().clone().list().eval(predicate).list().any();
        Self::from_expr(pred_expr, Some("exists".to_string()))
    }

    /// True if all list elements satisfy the predicate (PySpark forall). Uses list.eval(pred).list().all().
    pub fn array_forall(&self, predicate: Expr) -> Column {
        let pred_expr = self.expr().clone().list().eval(predicate).list().all();
        Self::from_expr(pred_expr, Some("forall".to_string()))
    }

    /// Filter list elements by predicate (PySpark filter). Keeps elements where predicate is true.
    pub fn array_filter(&self, predicate: Expr) -> Column {
        use polars::prelude::NULL;
        let then_val = Self::from_expr(col(""), None);
        let else_val = Self::from_expr(lit(NULL), None);
        let elem_expr = crate::functions::when(&Self::from_expr(predicate, None))
            .then(&then_val)
            .otherwise(&else_val)
            .into_expr();
        let list_expr = self
            .expr()
            .clone()
            .list()
            .eval(elem_expr)
            .list()
            .drop_nulls();
        Self::from_expr(list_expr, None)
    }

    /// Transform list elements by expression (PySpark transform). list.eval(expr).
    pub fn array_transform(&self, f: Expr) -> Column {
        let list_expr = self.expr().clone().list().eval(f);
        Self::from_expr(list_expr, None)
    }

    /// Sum of list elements (PySpark aggregate with sum). Uses list.sum().
    pub fn array_sum(&self) -> Column {
        Self::from_expr(self.expr().clone().list().sum(), None)
    }

    /// Array fold/aggregate (PySpark aggregate). Simplified: zero + sum(list). Full (zero, merge, finish) deferred.
    pub fn array_aggregate(&self, zero: &Column) -> Column {
        let sum_expr = self.expr().clone().list().sum();
        Self::from_expr(sum_expr + zero.expr().clone(), None)
    }

    /// Mean of list elements (PySpark aggregate with avg). Uses list.mean().
    pub fn array_mean(&self) -> Column {
        Self::from_expr(self.expr().clone().list().mean(), None)
    }

    /// Explode list with position (PySpark posexplode). Returns (pos_col, value_col).
    ///
    /// Implementation detail:
    /// - Build a list of structs per row: [{pos, col}, {pos, col}, ...].
    /// - Explode that single list column once, then project struct fields "pos" and "col".
    ///   This ensures that selecting both position and value in the same DataFrame.select
    ///   call yields a single exploded DataFrame (matching PySpark posexplode semantics)
    ///   instead of attempting two independent explode() calls on the same list column.
    pub fn posexplode(&self) -> (Column, Column) {
        use polars::prelude::{ExplodeOptions, as_struct};

        let opts = ExplodeOptions {
            empty_as_null: false,
            keep_nulls: false,
        };

        // In list.eval context, col("") is the current list element. cum_count(false)
        // yields 1-based positions, so subtract 1 for 0-based PySpark posexplode parity.
        let pos_inner = (col("").cum_count(false) - lit(1i64)).alias("pos");
        let val_inner = col("").alias("col");
        let struct_expr = as_struct(vec![pos_inner, val_inner]);

        // list of structs [{pos, col}, ...] per input row
        let list_struct_expr = self.expr().clone().list().eval(struct_expr);
        // explode once to get a struct column with one row per element
        let struct_exploded = list_struct_expr.explode(opts);

        let pos_expr = struct_exploded.clone().struct_().field_by_name("pos");
        let val_expr = struct_exploded.struct_().field_by_name("col");

        (
            Self::from_expr(pos_expr, Some("pos".to_string())),
            Self::from_expr(val_expr, Some("col".to_string())),
        )
    }

    /// Extract keys from a map column (PySpark map_keys). Map column is List(Struct{key, value}).
    pub fn map_keys(&self) -> Column {
        let elem_key = col("").struct_().field_by_name("key");
        let list_expr = self.expr().clone().list().eval(elem_key);
        Self::from_expr(list_expr, None)
    }

    /// Extract values from a map column (PySpark map_values). Map column is List(Struct{key, value}).
    pub fn map_values(&self) -> Column {
        let elem_val = col("").struct_().field_by_name("value");
        let list_expr = self.expr().clone().list().eval(elem_val);
        Self::from_expr(list_expr, None)
    }

    /// Return map as list of structs {key, value} (PySpark map_entries). Identity for List(Struct) column.
    pub fn map_entries(&self) -> Column {
        Self::from_expr(self.expr().clone(), None)
    }

    /// Build map from two array columns (keys, values) (PySpark map_from_arrays). Implemented via map_many UDF.
    pub fn map_from_arrays(&self, values: &Column) -> Column {
        let args = [values.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_map_from_arrays(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Merge two map columns (PySpark map_concat). Last value wins for duplicate keys.
    pub fn map_concat(&self, other: &Column) -> Column {
        let args = [other.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_map_concat(cols)),
            &args,
            |_schema, fields| Ok(fields[0].clone()),
        );
        Self::from_expr(expr, None)
    }

    /// Transform each map key by expr (PySpark transform_keys). key_expr should use col("").struct_().field_by_name("key").
    pub fn transform_keys(&self, key_expr: Expr) -> Column {
        use polars::prelude::as_struct;
        let value = col("").struct_().field_by_name("value");
        let new_struct = as_struct(vec![key_expr.alias("key"), value.alias("value")]);
        let list_expr = self.expr().clone().list().eval(new_struct);
        Self::from_expr(list_expr, None)
    }

    /// Transform each map value by expr (PySpark transform_values). value_expr should use col("").struct_().field_by_name("value").
    pub fn transform_values(&self, value_expr: Expr) -> Column {
        use polars::prelude::as_struct;
        let key = col("").struct_().field_by_name("key");
        let new_struct = as_struct(vec![key.alias("key"), value_expr.alias("value")]);
        let list_expr = self.expr().clone().list().eval(new_struct);
        Self::from_expr(list_expr, None)
    }

    /// Merge two maps by key with merge function (PySpark map_zip_with).
    /// Merge Expr uses col("").struct_().field_by_name("value1") and field_by_name("value2").
    pub fn map_zip_with(&self, other: &Column, merge: Expr) -> Column {
        use polars::prelude::as_struct;
        let args = [other.expr().clone()];
        let zip_expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_map_zip_to_struct(cols)),
            &args,
            |_schema, fields| {
                let list_inner = match &fields[0].dtype {
                    DataType::List(inner) => *inner.clone(),
                    _ => return Ok(fields[0].clone()),
                };
                let (key_dtype, value_dtype) = match &list_inner {
                    DataType::Struct(struct_fields) => {
                        let k = struct_fields
                            .iter()
                            .find(|f| f.name.as_str() == "key")
                            .map(|f| f.dtype.clone())
                            .unwrap_or(DataType::String);
                        let v = struct_fields
                            .iter()
                            .find(|f| f.name.as_str() == "value")
                            .map(|f| f.dtype.clone())
                            .unwrap_or(DataType::String);
                        (k, v)
                    }
                    _ => (DataType::String, DataType::String),
                };
                let out_struct = DataType::Struct(vec![
                    Field::new("key".into(), key_dtype),
                    Field::new("value1".into(), value_dtype.clone()),
                    Field::new("value2".into(), value_dtype),
                ]);
                Ok(Field::new(
                    fields[0].name().clone(),
                    DataType::List(Box::new(out_struct)),
                ))
            },
        );
        let key_field = col("").struct_().field_by_name("key").alias("key");
        let value_field = merge.alias("value");
        let merge_expr = as_struct(vec![key_field, value_field]);
        let list_expr = zip_expr.list().eval(merge_expr);
        Self::from_expr(list_expr, None)
    }

    /// Filter map entries by predicate (PySpark map_filter). Keeps key-value pairs where predicate is true.
    /// Predicate uses col("").struct_().field_by_name("key") and field_by_name("value") to reference key/value.
    pub fn map_filter(&self, predicate: Expr) -> Column {
        use polars::prelude::NULL;
        let then_val = Self::from_expr(col(""), None);
        let else_val = Self::from_expr(lit(NULL), None);
        let elem_expr = crate::functions::when(&Self::from_expr(predicate, None))
            .then(&then_val)
            .otherwise(&else_val)
            .into_expr();
        let list_expr = self
            .expr()
            .clone()
            .list()
            .eval(elem_expr)
            .list()
            .drop_nulls();
        Self::from_expr(list_expr, None)
    }

    /// Array of structs {key, value} to map (PySpark map_from_entries). Identity for List(Struct) format.
    pub fn map_from_entries(&self) -> Column {
        Self::from_expr(self.expr().clone(), None)
    }

    /// True if map contains key (PySpark map_contains_key).
    pub fn map_contains_key(&self, key: &Column) -> Column {
        let args = [key.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_map_contains_key(cols)),
            &args,
            |_schema, fields| Ok(Field::new(fields[0].name().clone(), DataType::Boolean)),
        );
        Self::from_expr(expr, None)
    }

    /// Get value for key from map, or null (PySpark get).
    pub fn get(&self, key: &Column) -> Column {
        let args = [key.expr().clone()];
        let expr = self.expr().clone().map_many(
            |cols| expect_col(crate::udfs::apply_get(cols)),
            &args,
            |_schema, fields| {
                let dtype = &fields[0].dtype;
                let value_dtype = match dtype {
                    DataType::List(inner) => match inner.as_ref() {
                        DataType::Struct(struct_fields) => struct_fields
                            .iter()
                            .find(|f| f.name == "value")
                            .map(|f| f.dtype.clone())
                            .unwrap_or(DataType::String),
                        _ => DataType::String,
                    },
                    DataType::Struct(struct_fields) => struct_fields
                        .first()
                        .map(|f| f.dtype.clone())
                        .unwrap_or(DataType::String),
                    _ => DataType::String,
                };
                Ok(Field::new(fields[0].name().clone(), value_dtype))
            },
        );
        Self::from_expr(expr, None)
    }

    /// Extract JSON path from string column (PySpark get_json_object). UDF returns string always (#1146).
    pub fn get_json_object(&self, path: &str) -> Column {
        let path = path.to_string();
        let expr = self
            .expr()
            .clone()
            .map(
                move |s| expect_col(crate::udfs::apply_get_json_object(s, &path)),
                |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
            )
            .cast(DataType::String);
        Self::from_expr(expr, None)
    }

    /// Parse string column as JSON into struct (PySpark from_json). Uses Polars str().json_decode.
    pub fn from_json(&self, schema: Option<polars::datatypes::DataType>) -> Column {
        use polars::prelude::DataType;
        let dtype = schema.unwrap_or(DataType::String);
        let out = self.expr().clone().str().json_decode(dtype);
        Self::from_expr(out, None)
    }

    /// Serialize struct column to JSON string (PySpark to_json). Uses Polars struct().json_encode.
    pub fn to_json(&self) -> Column {
        let out = self.expr().clone().struct_().json_encode();
        Self::from_expr(out, None)
    }

    /// Length of JSON array at path (PySpark json_array_length). UDF.
    pub fn json_array_length(&self, path: &str) -> Column {
        let path = path.to_string();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_json_array_length(s, &path)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Keys of JSON object (PySpark json_object_keys). Returns list of strings. UDF.
    pub fn json_object_keys(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_json_object_keys(s)),
            |_schema, field| {
                Ok(Field::new(
                    field.name().clone(),
                    DataType::List(Box::new(DataType::String)),
                ))
            },
        );
        Self::from_expr(expr, None)
    }

    /// Extract keys from JSON as struct (PySpark json_tuple). UDF. Returns struct with one string field per key.
    pub fn json_tuple(&self, keys: &[&str]) -> Column {
        let keys_vec: Vec<String> = keys.iter().map(|s| (*s).to_string()).collect();
        let struct_fields: Vec<polars::datatypes::Field> = keys_vec
            .iter()
            .map(|k| polars::datatypes::Field::new(k.as_str().into(), DataType::String))
            .collect();
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_json_tuple(s, &keys_vec)),
            move |_schema, field| {
                Ok(Field::new(
                    field.name().clone(),
                    DataType::Struct(struct_fields.clone()),
                ))
            },
        );
        Self::from_expr(expr, None)
    }

    /// Parse CSV string to struct (PySpark from_csv). Minimal: split by comma, up to 32 columns. UDF.
    pub fn from_csv(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_from_csv(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Struct(vec![]))),
        );
        Self::from_expr(expr, None)
    }

    /// Format struct as CSV string (PySpark to_csv). Minimal. UDF.
    pub fn to_csv(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_to_csv(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Parse URL and extract part (PySpark parse_url). UDF.
    /// When part is QUERY/QUERYSTRING and key is Some(k), returns the value for that query parameter only.
    pub fn parse_url(&self, part: &str, key: Option<&str>) -> Column {
        let part = part.to_string();
        let key_owned = key.map(String::from);
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_parse_url(s, &part, key_owned.as_deref())),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Hash of column value (PySpark hash). Single-column version.
    pub fn hash(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_hash_one(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }

    /// Check if column values are in the other column's list/series (PySpark isin).
    pub fn isin(&self, other: &Column) -> Column {
        let out = self.expr().clone().is_in(other.expr().clone(), false);
        Self::from_expr(out, None)
    }

    /// Percent-decode URL-encoded string (PySpark url_decode). Uses UDF.
    pub fn url_decode(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_url_decode(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Percent-encode string for URL (PySpark url_encode). Uses UDF.
    pub fn url_encode(&self) -> Column {
        let expr = self.expr().clone().map(
            |s| expect_col(crate::udfs::apply_url_encode(s)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::String)),
        );
        Self::from_expr(expr, None)
    }

    /// Bitwise left shift (PySpark shiftLeft). col << n = col * 2^n.
    pub fn shift_left(&self, n: i32) -> Column {
        use polars::prelude::*;
        let pow = lit(2i64).pow(lit(n as i64));
        Self::from_expr(
            (self.expr().clone().cast(DataType::Int64) * pow).cast(DataType::Int64),
            None,
        )
    }

    /// Bitwise signed right shift (PySpark shiftRight). col >> n = col / 2^n.
    pub fn shift_right(&self, n: i32) -> Column {
        use polars::prelude::*;
        let pow = lit(2i64).pow(lit(n as i64));
        Self::from_expr(
            (self.expr().clone().cast(DataType::Int64) / pow).cast(DataType::Int64),
            None,
        )
    }

    /// Bitwise unsigned right shift (PySpark shiftRightUnsigned). Logical shift.
    pub fn shift_right_unsigned(&self, n: i32) -> Column {
        let expr = self.expr().clone().map(
            move |s| expect_col(crate::udfs::apply_shift_right_unsigned(s, n)),
            |_schema, field| Ok(Field::new(field.name().clone(), DataType::Int64)),
        );
        Self::from_expr(expr, None)
    }
}

#[cfg(test)]
mod tests {
    use super::Column;
    use polars::prelude::{IntoLazy, col, df, lit};

    /// Helper to create a simple DataFrame for testing
    fn test_df() -> polars::prelude::DataFrame {
        df!(
            "a" => &[1, 2, 3, 4, 5],
            "b" => &[10, 20, 30, 40, 50]
        )
        .unwrap()
    }

    /// Helper to create a DataFrame with nulls for testing
    fn test_df_with_nulls() -> polars::prelude::DataFrame {
        df!(
            "a" => &[Some(1), Some(2), None, Some(4), None],
            "b" => &[Some(10), None, Some(30), None, None]
        )
        .unwrap()
    }

    #[test]
    fn test_column_new() {
        let column = Column::new("age".to_string());
        assert_eq!(column.name(), "age");
    }

    #[test]
    fn test_column_from_expr() {
        let expr = col("test");
        let column = Column::from_expr(expr, Some("test".to_string()));
        assert_eq!(column.name(), "test");
    }

    #[test]
    fn test_column_from_expr_default_name() {
        let expr = col("test").gt(lit(5));
        let column = Column::from_expr(expr, None);
        assert_eq!(column.name(), "<expr>");
    }

    #[test]
    fn test_column_alias() {
        let column = Column::new("original".to_string());
        let aliased = column.alias("new_name");
        assert_eq!(aliased.name(), "new_name");
    }

    #[test]
    fn test_column_gt() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.gt(lit(3));

        // Apply the expression to filter the DataFrame
        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 2); // rows with a > 3: 4, 5
    }

    #[test]
    fn test_column_lt() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.lt(lit(3));

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 2); // rows with a < 3: 1, 2
    }

    #[test]
    fn test_column_eq() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.eq(lit(3));

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 1); // only row with a == 3
    }

    #[test]
    fn test_column_neq() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.neq(lit(3));

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 4); // rows with a != 3
    }

    #[test]
    fn test_column_gt_eq() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.gt_eq(lit(3));

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 3); // rows with a >= 3: 3, 4, 5
    }

    #[test]
    fn test_column_lt_eq() {
        let df = test_df();
        let column = Column::new("a".to_string());
        let result = column.lt_eq(lit(3));

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 3); // rows with a <= 3: 1, 2, 3
    }

    #[test]
    fn test_column_is_null() {
        let df = test_df_with_nulls();
        let column = Column::new("a".to_string());
        let result = column.is_null();

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 2); // 2 null values in column 'a'
    }

    #[test]
    fn test_column_is_not_null() {
        let df = test_df_with_nulls();
        let column = Column::new("a".to_string());
        let result = column.is_not_null();

        let filtered = df.lazy().filter(result.into_expr()).collect().unwrap();
        assert_eq!(filtered.height(), 3); // 3 non-null values in column 'a'
    }

    #[test]
    fn test_null_boolean_column_produces_null_bool_series() {
        let df = test_df();
        let expr = Column::null_boolean().into_expr();
        let out = df
            .lazy()
            .select([expr.alias("null_bool")])
            .collect()
            .unwrap();
        let s = out.column("null_bool").unwrap();
        assert_eq!(s.dtype(), &polars::prelude::DataType::Boolean);
        assert_eq!(s.null_count(), s.len());
    }

    #[test]
    fn test_eq_null_safe_both_null() {
        // Create a DataFrame where both columns have NULL at the same row
        let df = df!(
            "a" => &[Some(1), None, Some(3)],
            "b" => &[Some(1), None, Some(4)]
        )
        .unwrap();

        let col_a = Column::new("a".to_string());
        let col_b = Column::new("b".to_string());
        let result = col_a.eq_null_safe(&col_b);

        // Apply the expression and collect
        let result_df = df
            .lazy()
            .with_column(result.into_expr().alias("eq_null_safe"))
            .collect()
            .unwrap();

        // Get the result column
        let eq_col = result_df.column("eq_null_safe").unwrap();
        let values: Vec<Option<bool>> = eq_col.bool().unwrap().into_iter().collect();

        // Row 0: 1 == 1 -> true
        // Row 1: NULL <=> NULL -> true
        // Row 2: 3 == 4 -> false
        assert_eq!(values[0], Some(true));
        assert_eq!(values[1], Some(true)); // NULL-safe: both NULL = true
        assert_eq!(values[2], Some(false));
    }

    #[test]
    fn test_eq_null_safe_one_null() {
        // Create a DataFrame where only one column has NULL
        let df = df!(
            "a" => &[Some(1), None, Some(3)],
            "b" => &[Some(1), Some(2), None]
        )
        .unwrap();

        let col_a = Column::new("a".to_string());
        let col_b = Column::new("b".to_string());
        let result = col_a.eq_null_safe(&col_b);

        let result_df = df
            .lazy()
            .with_column(result.into_expr().alias("eq_null_safe"))
            .collect()
            .unwrap();

        let eq_col = result_df.column("eq_null_safe").unwrap();
        let values: Vec<Option<bool>> = eq_col.bool().unwrap().into_iter().collect();

        // Row 0: 1 == 1 -> true
        // Row 1: NULL <=> 2 -> false (one is null, not both)
        // Row 2: 3 <=> NULL -> false (one is null, not both)
        assert_eq!(values[0], Some(true));
        assert_eq!(values[1], Some(false));
        assert_eq!(values[2], Some(false));
    }
}