rayforce-sys 1.3.0

/*
 *   Copyright (c) 2025-2026 Anton Kundenko <singaraiona@gmail.com>
 *   All rights reserved.

 *   Permission is hereby granted, free of charge, to any person obtaining a copy
 *   of this software and associated documentation files (the "Software"), to deal
 *   in the Software without restriction, including without limitation the rights
 *   to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 *   copies of the Software, and to permit persons to whom the Software is
 *   furnished to do so, subject to the following conditions:

 *   The above copyright notice and this permission notice shall be included in all
 *   copies or substantial portions of the Software.

 *   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 *   IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 *   FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 *   AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 *   LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 *   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 *   SOFTWARE.
 */

#include <inttypes.h>
#include <stdatomic.h>
#include "ops/internal.h"
#include "ops/hash.h"
#include "ops/idxop.h"
#include "ops/rowsel.h"
#include "table/domain.h"   /* raw vocabulary snapshot for if_sym_side_t */
#include "core/pool.h"
#include "lang/format.h"    /* ray_type_name (error context) */

/* Resolved string atom (borrowed) of a SYM-scalar broadcast input
 * (atom -RAY_SYM, or a 1-elem RAY_SYM_W{8,16,32,64} vec used as
 * scalar).  Atoms are runtime-domain by design; a vec scalar is
 * CELL-DATA and resolves through its own domain.  Hides the atom/vec
 * dispatch: ray_t->i64 aliases ray_t->len, so reading `v->i64` on a
 * vec silently yields `len` (= 1) instead of the element value —
 * always go through these helpers for a then_v/else_v that may be
 * either atom or 1-elem vec. */
static inline ray_t* sym_scalar_str(ray_t* v) {
    return ray_is_atom(v) ? ray_sym_str(v->i64) : ray_sym_vec_cell(v, 0);
}

/* A SYM id taken from `col`'s cells, re-expressed in the RUNTIME domain
 * (fresh SYM output vecs built here are runtime-domain).  Fast path: a
 * runtime-domain column's ids ARE runtime ids — raw copy.  Otherwise a
 * lock-free read of the per-domain runtime-id LUT (pivot is sequential,
 * so the LUT's first-request vocabulary intern is safe here; mirrors
 * lang/internal.h sym_id_runtime — exact no-op while every domain is
 * the runtime singleton). */
static inline int64_t sym_id_runtime(ray_t* col, int64_t id) {
    struct ray_sym_domain_s* dom = ray_sym_vec_domain(col);
    if (dom == ray_sym_runtime_domain()) return id;
    const int64_t* lut = ray_sym_domain_runtime_lut(dom);
    if (!lut || id < 0 || id >= ray_sym_domain_count(dom)) return -1;
    return lut[id];
}

static inline int64_t sym_cell_runtime_id(ray_t* v, int64_t i) {
    return sym_id_runtime(v, ray_read_sym(ray_data(v), i, v->type, v->attrs));
}

static inline int64_t sym_scalar_runtime_id(ray_t* v) {
    return ray_is_atom(v) ? v->i64 : sym_cell_runtime_id(v, 0);
}

static inline ray_t* str_vec_pool_obj(ray_t* v) {
    if (!v || v->type != RAY_STR) return NULL;
    ray_t* owner = (v->attrs & RAY_ATTR_SLICE) ? v->slice_parent : v;
    return owner ? owner->str_pool : NULL;
}

enum { IF_BRANCH_MAX_SCAN_SYMS = 64 };

typedef struct {
    int64_t syms[IF_BRANCH_MAX_SCAN_SYMS];
    int     n_syms;
} if_branch_plan_t;

static bool if_plan_add_sym(if_branch_plan_t* p, int64_t sym) {
    for (int i = 0; i < p->n_syms; i++)
        if (p->syms[i] == sym) return true;
    if (p->n_syms >= IF_BRANCH_MAX_SCAN_SYMS) return false;
    p->syms[p->n_syms++] = sym;
    return true;
}

static bool if_op_rowwise(uint16_t opc) {
    switch (opc) {
    case OP_SCAN: case OP_CONST:
    case OP_NEG: case OP_ABS: case OP_NOT: case OP_SQRT:
    case OP_LOG: case OP_EXP: case OP_CEIL: case OP_FLOOR: case OP_ROUND:
    case OP_SIN: case OP_ASIN: case OP_COS: case OP_ACOS:
    case OP_TAN: case OP_ATAN: case OP_RECIPROCAL: case OP_SIGNUM:
    case OP_ISNULL: case OP_CAST:
    case OP_ADD: case OP_SUB: case OP_MUL: case OP_DIV: case OP_IDIV: case OP_MOD:
    case OP_POW:
    case OP_EQ: case OP_NE: case OP_LT: case OP_LE:
    case OP_GT: case OP_GE: case OP_AND: case OP_OR:
    case OP_MIN2: case OP_MAX2: case OP_IF:
    case OP_LIKE: case OP_ILIKE: case OP_UPPER: case OP_LOWER:
    case OP_STRLEN: case OP_SUBSTR: case OP_REPLACE: case OP_TRIM:
    case OP_CONCAT: case OP_STR_FIND:
    case OP_EXTRACT: case OP_DATE_TRUNC:
        return true;
    default:
        return false;
    }
}

static bool if_collect_branch(ray_graph_t* g, ray_op_t* op,
                              if_branch_plan_t* plan, uint64_t* visited) {
    if (!g || !op) return false;
    uint32_t nid = op->id;
    if (nid < g->node_count) {
        uint64_t bit = (uint64_t)1 << (nid & 63);
        uint64_t* word = &visited[nid >> 6];
        if (*word & bit) return true;
        *word |= bit;
    }

    uint16_t opc = op->opcode;
    if (opc == OP_CONST) {
        ray_op_ext_t* ext = find_ext(g, op->id);
        return ext && ext->literal && ray_is_atom(ext->literal);
    }
    if (opc == OP_SCAN) {
        ray_op_ext_t* ext = find_ext(g, op->id);
        if (!ext) return false;
        uint16_t stored_table_id = 0;
        memcpy(&stored_table_id, ext->base.pad, sizeof(uint16_t));
        if (stored_table_id != 0) return false;
        return if_plan_add_sym(plan, ext->sym);
    }
    if (!if_op_rowwise(opc)) return false;

    for (uint8_t i = 0; i < op->arity && i < 2; i++) {
        if (op->in_id[i] == RAY_OP_NONE) continue;
        if (!if_collect_branch(g, op_child(g, op, i), plan, visited))
            return false;
    }

    ray_op_ext_t* ext = find_ext(g, op->id);
    if ((opc == OP_IF || opc == OP_SUBSTR || opc == OP_REPLACE) && ext) {
        if (ext->third_in != RAY_OP_NONE &&
            !if_collect_branch(g, op_node(g, ext->third_in), plan, visited))
            return false;
    } else if (opc == OP_CONCAT && ext) {
        int n_args = (int)ext->sym;
        uint32_t* trail = (uint32_t*)((char*)(ext + 1));
        for (int i = 2; i < n_args; i++) {
            if (!if_collect_branch(g, op_node(g, trail[i - 2]), plan, visited))
                return false;
        }
    }
    return true;
}

static bool if_make_branch_plan(ray_graph_t* g, ray_op_t* root,
                                if_branch_plan_t* plan) {
    memset(plan, 0, sizeof(*plan));
    if (!g || !root || g->node_count == 0) return false;
    ray_t* hdr = NULL;
    size_t words = ((size_t)g->node_count + 63u) >> 6;
    uint64_t* visited = (uint64_t*)scratch_calloc(&hdr, words * sizeof(uint64_t));
    if (!visited) return false;
    bool ok = if_collect_branch(g, root, plan, visited);
    scratch_free(hdr);
    return ok;
}

static int64_t if_selected_true_count(const uint8_t* cond, int64_t nrows,
                                      ray_t* outer_sel) {
    int64_t total = 0;
    if (!outer_sel) {
        for (int64_t i = 0; i < nrows; i++) total += cond[i] ? 1 : 0;
        return total;
    }
    ray_rowsel_t* m = ray_rowsel_meta(outer_sel);
    const uint8_t* flags = ray_rowsel_flags(outer_sel);
    const uint32_t* offsets = ray_rowsel_offsets(outer_sel);
    const uint16_t* idx = ray_rowsel_idx(outer_sel);
    for (uint32_t seg = 0; seg < m->n_segs; seg++) {
        uint8_t f = flags[seg];
        if (f == RAY_SEL_NONE) continue;
        int64_t base = (int64_t)seg * RAY_MORSEL_ELEMS;
        int64_t end = base + RAY_MORSEL_ELEMS;
        if (end > nrows) end = nrows;
        if (f == RAY_SEL_ALL) {
            for (int64_t r = base; r < end; r++) total += cond[r] ? 1 : 0;
        } else {
            uint32_t off = offsets[seg];
            uint32_t cnt = offsets[seg + 1] - off;
            for (uint32_t i = 0; i < cnt; i++) {
                int64_t r = base + idx[off + i];
                total += cond[r] ? 1 : 0;
            }
        }
    }
    return total;
}

static void if_fill_branch_ids(const uint8_t* cond, int64_t nrows,
                               ray_t* outer_sel, int64_t* true_ids,
                               int64_t* false_ids) {
    int64_t ti = 0, fi = 0;
#define IF_ADD_ROW(R) do { int64_t _r = (R); if (cond[_r]) true_ids[ti++] = _r; else false_ids[fi++] = _r; } while (0)
    if (!outer_sel) {
        for (int64_t r = 0; r < nrows; r++) IF_ADD_ROW(r);
    } else {
        ray_rowsel_t* m = ray_rowsel_meta(outer_sel);
        const uint8_t* flags = ray_rowsel_flags(outer_sel);
        const uint32_t* offsets = ray_rowsel_offsets(outer_sel);
        const uint16_t* idx = ray_rowsel_idx(outer_sel);
        for (uint32_t seg = 0; seg < m->n_segs; seg++) {
            uint8_t f = flags[seg];
            if (f == RAY_SEL_NONE) continue;
            int64_t base = (int64_t)seg * RAY_MORSEL_ELEMS;
            int64_t end = base + RAY_MORSEL_ELEMS;
            if (end > nrows) end = nrows;
            if (f == RAY_SEL_ALL) {
                for (int64_t r = base; r < end; r++) IF_ADD_ROW(r);
            } else {
                uint32_t off = offsets[seg];
                uint32_t cnt = offsets[seg + 1] - off;
                for (uint32_t i = 0; i < cnt; i++) IF_ADD_ROW(base + idx[off + i]);
            }
        }
    }
#undef IF_ADD_ROW
}

static ray_t* if_eval_branch(ray_graph_t* g, ray_op_t* branch,
                             const if_branch_plan_t* plan,
                             int64_t* ids, int64_t count, int64_t nrows) {
    if (count <= 0) return NULL;
    ray_t* saved_table = g->table;
    ray_t* saved_sel = g->selection;
    ray_t* sub = NULL;
    ray_t* sel = NULL;

    if (plan->n_syms > 0 && count != nrows) {
        sel = ray_index_rowsel_from_ids(nrows, ids, count);
        if (!sel) return ray_error("oom", NULL);
        sub = sel_compact(g, saved_table, sel, plan->syms, plan->n_syms);
        ray_release(sel);
        if (!sub || RAY_IS_ERR(sub)) return sub ? sub : ray_error("oom", NULL);
    } else {
        sub = saved_table;
        ray_retain(sub);
    }

    g->table = sub;
    g->selection = NULL;
    /* The branch's shared nodes get a memo over the branch's rows; the
     * outer memo (values of the full table) is set aside meanwhile. */
    ray_exec_memo_save_t memo_save;
    ray_exec_memo_push(g, branch, &memo_save);
    ray_t* value = exec_node(g, branch);
    ray_exec_memo_pop(g, &memo_save);
    if (g->selection) {
        ray_release(g->selection);
        g->selection = NULL;
    }
    g->table = saved_table;
    g->selection = saved_sel;
    ray_release(sub);
    return value;
}

static inline bool if_value_scalar(ray_t* v) {
    return ray_is_atom(v) || (v->type > 0 && v->len == 1);
}

static int64_t if_value_index(ray_t* v, int64_t* ids, int64_t j,
                              int64_t count, int64_t nrows) {
    if (if_value_scalar(v)) return 0;
    if (v->len == count) return j;
    if (v->len == nrows) return ids[j];
    return -1;
}

/* A double with no integer image — a NaN (the F64 null) or a magnitude past
 * the int64 range — must not be cast.  The cast is undefined, and where it
 * does not trap the result differs per target, so a null would narrow into a
 * different arbitrary number on each platform.  The integer null is the one
 * in-band value that means "no representable answer", so use it for both. */
static inline int64_t if_f64_to_i64(double x) {
    if (!(x >= -9223372036854775808.0 && x < 9223372036854775808.0))
        return NULL_I64;
    return (int64_t)x;
}

static int64_t if_atom_i64(ray_t* v) {
    switch (-v->type) {
    case RAY_I64: case RAY_TIMESTAMP: return v->i64;
    case RAY_I32: case RAY_DATE: case RAY_TIME: return v->i32;
    case RAY_I16: return v->i16;
    case RAY_BOOL: case RAY_U8: return v->b8;
    case RAY_F64: case RAY_F32: return if_f64_to_i64(v->f64);
    default: return 0;
    }
}

/* A null atom stays null across the widths.  Converting the value instead —
 * an I64 sentinel read as a double — produces an ordinary huge number, so
 * `(if c f 0N)` over an F64 column wrote -9.22e18 where a null belongs. */
static double if_atom_f64(ray_t* v) {
    if (RAY_ATOM_IS_NULL(v)) return (double)NAN;
    return (v->type == -RAY_F64 || v->type == -RAY_F32)
         ? v->f64 : (double)if_atom_i64(v);
}

/* Does this branch value contribute a null to the output?  A vector says so
 * with its attr; an atom says so by BEING the sentinel, and nothing sets the
 * attr for it — so `(if x d 0Nd)` wrote the sentinel and then rendered it as
 * an ordinary date, since the formatter reads the attr, not the value. */
static bool if_val_has_null(ray_t* v) {
    if (!v) return false;
    if (!ray_is_atom(v)) return (v->attrs & RAY_ATTR_HAS_NULLS) != 0;
    /* RAY_ATOM_IS_NULL answers for the atom's OWN type, which is the right
     * question: keying off out_type missed a null whose sentinel is a
     * different width from the promoted result, so `(if c f 0N)` over an F64
     * column wrote the I64 sentinel as a number with no null marked. */
    return RAY_ATOM_IS_NULL(v);
}

static int64_t if_vec_i64(ray_t* v, int64_t idx) {
    if (v->type == RAY_F64) return if_f64_to_i64(((double*)ray_data(v))[idx]);
    if (v->type == RAY_F32) return if_f64_to_i64((double)((float*)ray_data(v))[idx]);
    return read_col_i64(ray_data(v), idx, v->type, v->attrs);
}

static double if_vec_f64(ray_t* v, int64_t idx) {
    if (v->type == RAY_F64) return ((double*)ray_data(v))[idx];
    if (v->type == RAY_F32) return (double)((float*)ray_data(v))[idx];
    return (double)if_vec_i64(v, idx);
}

static int64_t if_scalar_i64(ray_t* v) {
    return ray_is_atom(v) ? if_atom_i64(v) : if_vec_i64(v, 0);
}

static double if_scalar_f64(ray_t* v) {
    return ray_is_atom(v) ? if_atom_f64(v) : if_vec_f64(v, 0);
}

static int8_t if_value_type(ray_t* v) {
    if (!v) return 0;
    if (ray_is_atom(v) && v->type == -RAY_STR) return RAY_SYM;
    int8_t t = ray_is_atom(v) ? (int8_t)(-(int)v->type) : v->type;
    return (t == RAY_F32) ? RAY_F64 : t;
}



/* The graph already promoted the two branch types into op->out_type, and that
 * is the answer — with one exception it cannot know statically.  A branch that
 * ends up supplying no rows never materialises its type, so an F64 demanded by
 * a float branch that never ran should not survive: the integer side keeps its
 * own width.  That is the whole of the narrowing this arm does, and it is why
 * (if (> v 0) v (sqrt s)) sums to 15 rather than 15.0 when every row is
 * positive.
 *
 * And only where the eager arm could not have handled the node at all — a STR
 * output, or a branch it cannot fill elementwise.  Where it could, it fills
 * every row from op->out_type and never narrows, so narrowing here would make
 * the same expression answer differently by worker count, which is the defect
 * being fixed.
 *
 * Anything beyond that was a second promotion ladder living beside the graph's,
 * and the two had drifted: it folded two matching temporal types down to their
 * integer storage, so the same expression answered TIMESTAMP or I64 depending
 * on which arm of exec_if ran. */
/* How wide a value of this type is as a hash/store key, for the widen-only
 * merge below.  Only the numeric ladder is ordered; anything else answers 0
 * and is never used to widen. */
static int if_type_rank(int8_t t) {
    switch (t) {
        case RAY_BOOL: return 1;
        case RAY_U8:   return 2;
        case RAY_I16:  return 3;
        case RAY_I32:  return 4;
        case RAY_I64:  return 5;
        case RAY_F64:  return 6;
        default:       return 0;
    }
}

/* The graph promoted the branch types into op->out_type, but it promoted the
 * declared types of the branch EXPRESSIONS, and arithmetic widens beyond
 * them: (* a a) over an I32 column is computed and stored as I64, and calling
 * the result I32 truncates every value.  So the declared type is a floor, not
 * an answer — whichever of it and the materialised values is wider wins.
 *
 * Narrowing is only ever right for a branch that supplied no rows, where the
 * type it would have contributed never materialised; even then it is limited
 * to nodes the eager arm could not have taken, since that arm fills from
 * op->out_type and never narrows, and the two answering differently by worker
 * count is the defect this is all about. */
static int8_t if_selected_type(ray_t* then_v, ray_t* else_v, int8_t declared,
                               bool eager_possible) {
    int8_t tt = if_value_type(then_v);
    int8_t et = if_value_type(else_v);

    /* Narrowing is for a branch that supplied no rows, and only where the
     * declared type came FROM that branch — otherwise it is not narrowing but
     * disagreeing with the eager arm.  Two shapes qualify:
     *
     *   a float branch that never ran, leaving integers that keep their own
     *   width — this is what makes (if (> v 0) v (sqrt s)) sum to 15 and not
     *   15.0;
     *
     *   a string branch that never ran, leaving a number that would otherwise
     *   be written into a string column and read back as empty.
     *
     * Two temporal branches do not qualify: their I64 comes from the pair, not
     * from either one, so narrowing to the survivor made a nested-if branch
     * answer DATE where the same expression with a bare column answered I64. */
    if (!eager_possible && (!then_v || !else_v)) {
        int8_t live = then_v ? tt : et;
        int lr = if_type_rank(live);
        bool float_branch_absent = declared == RAY_F64 &&
                                   lr && lr < if_type_rank(RAY_F64);
        bool string_branch_absent = (declared == RAY_STR || declared == RAY_SYM) &&
                                    lr > 0;
        if (float_branch_absent || string_branch_absent) return live;
    }

    int dr = if_type_rank(declared);
    if (dr) {
        int8_t widest = declared;
        int wr = dr;
        if (if_type_rank(tt) > wr) { widest = tt; wr = if_type_rank(tt); }
        if (if_type_rank(et) > wr) { widest = et; }
        return widest;
    }
    return declared;
}

static bool if_scatter_numeric(ray_t* result, ray_t* value, int64_t* ids,
                               int64_t count, int64_t nrows, int8_t out_type) {
    if (!value) return true;
    for (int64_t j = 0; j < count; j++) {
        int64_t src = if_value_index(value, ids, j, count, nrows);
        if (src < 0) return false;
        int64_t dst = ids[j];
        if (out_type == RAY_F64) {
            double x = ray_is_atom(value) ? if_atom_f64(value)
                     : if_vec_f64(value, src);
            ((double*)ray_data(result))[dst] = x;
        } else if (out_type == RAY_I64 || out_type == RAY_TIMESTAMP) {
            int64_t x = ray_is_atom(value) ? if_atom_i64(value) : if_vec_i64(value, src);
            ((int64_t*)ray_data(result))[dst] = x;
        } else if (out_type == RAY_I32 || out_type == RAY_DATE || out_type == RAY_TIME) {
            int64_t x = ray_is_atom(value) ? if_atom_i64(value) : if_vec_i64(value, src);
            ((int32_t*)ray_data(result))[dst] = (int32_t)x;
        } else if (out_type == RAY_I16) {
            int64_t x = ray_is_atom(value) ? if_atom_i64(value) : if_vec_i64(value, src);
            ((int16_t*)ray_data(result))[dst] = (int16_t)x;
        } else {
            int64_t x = ray_is_atom(value) ? if_atom_i64(value) : if_vec_i64(value, src);
            ((uint8_t*)ray_data(result))[dst] = (uint8_t)x;
        }
    }
    return true;
}

static bool if_string_cell(ray_t* value, int64_t src,
                           const char** sp, size_t* sl) {
    if (value->type == -RAY_STR) {
        *sp = ray_str_ptr(value); *sl = ray_str_len(value); return true;
    }
    if (value->type == RAY_STR) {
        *sp = ray_str_vec_get(value, src, sl);
        if (!*sp) { *sp = ""; *sl = 0; }
        return true;
    }
    if (value->type == -RAY_SYM || RAY_IS_SYM(value->type)) {
        ray_t* s = ray_is_atom(value) ? sym_scalar_str(value)
                                      : ray_sym_vec_cell(value, src);
        *sp = s ? ray_str_ptr(s) : "";
        *sl = s ? ray_str_len(s) : 0;
        return true;
    }
    *sp = ""; *sl = 0;
    return true;
}

static ray_t* if_scatter_str(ray_t* result, ray_t* value, int64_t* ids,
                             int64_t count, int64_t nrows) {
    if (!value) return result;
    for (int64_t j = 0; j < count; j++) {
        int64_t src = if_value_index(value, ids, j, count, nrows);
        if (src < 0) { ray_release(result); return ray_error("length", "if: branch length mismatch"); }
        const char* sp = NULL;
        size_t sl = 0;
        if_string_cell(value, src, &sp, &sl);
        ray_t* next = ray_str_vec_set(result, ids[j], sp ? sp : "", sp ? sl : 0);
        if (!next || RAY_IS_ERR(next)) {
            ray_release(result);
            return next ? next : ray_error("oom", NULL);
        }
        result = next;
    }
    return result;
}

/* Descriptor scatter for the STR arm of the selected path.  A side is a
 * STR vector with one row per id (or as many rows as the table, indexed by
 * the id), or a broadcast scalar.  The result takes the side's 16-byte
 * descriptor at ids[j] instead of appending its bytes row by row: pooled
 * strings keep pointing into their pool when the rows keep most of it;
 * otherwise (two different pools, a pooled scalar, or a small share of a
 * big pool) the result gets a pool of exactly its own bytes.  Rows not in
 * either id list stay the null descriptor the caller
 * zeroed.  Returns NULL, the result untouched, for a side this cannot take
 * (a SYM branch, a length that is neither) — the caller then falls back to
 * the per-row scatter, which also reports the length error. */
typedef struct {
    ray_t*           v;
    int64_t*         ids;
    int64_t          n;
    bool             scalar;
    bool             full;     /* vector of nrows: row ids[j] */
    const ray_str_t* desc;
    const char*      bytes;    /* its pool's bytes (NULL when inline-only) */
    ray_t*           pool;
    const char*      sp;       /* scalar bytes */
    size_t           sl;
} if_str_side_t;

static bool if_str_side_init(ray_t* v, int64_t* ids, int64_t n, int64_t nrows,
                             if_str_side_t* s) {
    memset(s, 0, sizeof(*s));
    s->ids = ids; s->n = n;
    if (!v || n <= 0) return true;
    s->v = v;
    if (v->type == -RAY_STR) {
        s->scalar = true; s->sp = ray_str_ptr(v); s->sl = ray_str_len(v);
        return true;
    }
    if (v->type != RAY_STR) return false;
    if (v->len == 1) {
        s->scalar = true;
        s->sp = ray_str_vec_get(v, 0, &s->sl);
        if (!s->sp) { s->sp = ""; s->sl = 0; }
        return true;
    }
    if (v->len == n) s->full = false;
    else if (v->len == nrows) s->full = true;
    else return false;
    str_resolve(v, &s->desc, &s->bytes);
    s->pool = str_vec_pool_obj(v);
    if (s->pool && RAY_IS_ERR(s->pool)) return false;
    return true;
}

/* The side's descriptor for its j-th id. */
static inline ray_str_t if_str_side_desc(const if_str_side_t* s, int64_t j) {
    return s->full ? s->desc[s->ids[j]] : s->desc[j];
}

/* Bytes the side's pooled descriptors point at. */
static uint64_t if_str_side_pooled_bytes(const if_str_side_t* s) {
    if (!s->v || s->scalar) return 0;
    uint64_t sum = 0;
    for (int64_t j = 0; j < s->n; j++) {
        ray_str_t d = if_str_side_desc(s, j);
        if (!ray_str_is_inline(&d)) sum += d.len;
    }
    return sum;
}

/* Assign compact offsets to the side's pooled descriptors from *run. */
static void if_str_side_offsets(const if_str_side_t* s, uint32_t* newoff, uint64_t* run) {
    for (int64_t j = 0; j < s->n; j++) {
        ray_str_t d = if_str_side_desc(s, j);
        if (ray_str_is_inline(&d)) continue;
        newoff[j] = (uint32_t)*run;
        *run += d.len;
    }
}

typedef struct {
    const int64_t*   ids;
    const ray_str_t* src;
    bool             scalar;
    bool             full;
    ray_str_t        sd;       /* the scalar's descriptor */
    ray_str_t*       dst;
    /* compaction: pooled bytes move to dst_bytes at newoff[j] */
    const uint32_t*  newoff;
    const char*      src_bytes;
    char*            dst_bytes;
} if_str_scatter_ctx_t;

static void if_str_scatter_fn(void* vctx, uint32_t wid, int64_t start, int64_t end) {
    (void)wid;
    const if_str_scatter_ctx_t* c = (const if_str_scatter_ctx_t*)vctx;
    if (c->scalar) {
        for (int64_t j = start; j < end; j++) c->dst[c->ids[j]] = c->sd;
        return;
    }
    for (int64_t j = start; j < end; j++) {
        ray_str_t d = c->full ? c->src[c->ids[j]] : c->src[j];
        if (c->newoff && !ray_str_is_inline(&d)) {
            memcpy(c->dst_bytes + c->newoff[j], c->src_bytes + d.pool_off, d.len);
            d.pool_off = c->newoff[j];
        }
        c->dst[c->ids[j]] = d;
    }
}

static void if_str_scatter_side(const if_str_side_t* s, const uint32_t* newoff, char* dst_bytes,
                                uint32_t scalar_off, ray_str_t* dst) {
    if (!s->v) return;
    if_str_scatter_ctx_t c = {
        .ids = s->ids, .src = s->desc, .scalar = s->scalar, .full = s->full,
        .dst = dst, .newoff = newoff, .src_bytes = s->bytes, .dst_bytes = dst_bytes,
    };
    if (s->scalar) {
        memset(&c.sd, 0, sizeof(c.sd));
        c.sd.len = (uint32_t)s->sl;
        if (s->sl <= RAY_STR_INLINE_MAX) {
            if (s->sl) memcpy(c.sd.data, s->sp, s->sl);
        } else {
            memcpy(c.sd.prefix, s->sp, 4);
            c.sd.pool_off = scalar_off;
        }
    }
    ray_pool_t* pool = ray_pool_get();
    if (ray_pool_par_dispatch_ok(pool, s->n, RAY_PARALLEL_THRESHOLD))
        ray_pool_dispatch(pool, if_str_scatter_fn, &c, s->n);
    else
        if_str_scatter_fn(&c, 0, 0, s->n);
}

static ray_t* if_scatter_str_desc(ray_t* result,
                                  ray_t* then_v, int64_t* t_ids, int64_t t_n,
                                  ray_t* else_v, int64_t* e_ids, int64_t e_n,
                                  int64_t nrows) {
    if_str_side_t t, e;
    if (!if_str_side_init(then_v, t_ids, t_n, nrows, &t)) return NULL;
    if (!if_str_side_init(else_v, e_ids, e_n, nrows, &e)) return NULL;

    ray_t* pt = (t.v && !t.scalar) ? t.pool : NULL;
    ray_t* pe = (e.v && !e.scalar) ? e.pool : NULL;
    bool t_big = t.v && t.scalar && t.sl > RAY_STR_INLINE_MAX;
    bool e_big = e.v && e.scalar && e.sl > RAY_STR_INLINE_MAX;
    uint64_t t_ref = if_str_side_pooled_bytes(&t);
    uint64_t e_ref = if_str_side_pooled_bytes(&e);
    ray_str_t* dst = (ray_str_t*)ray_data(result);

    /* One pool (or none) behind both sides, and the rows keep most of it:
     * point into it as is. */
    bool one_pool = (pt == pe) || !pt || !pe;
    ray_t* shared = pt ? pt : pe;
    if (!t_big && !e_big && one_pool &&
        (!shared || !ray_str_view_should_compact(t_ref + e_ref, shared->len))) {
        if (shared) { ray_retain(shared); result->str_pool = shared; }
        if_str_scatter_side(&t, NULL, NULL, 0, dst);
        if_str_scatter_side(&e, NULL, NULL, 0, dst);
        return result;
    }

    /* Otherwise a pool of exactly the bytes the result points at: the two
     * sides' pooled rows one after the other, then the pooled scalars. */
    uint64_t total = t_ref + e_ref + (t_big ? (uint64_t)t.sl : 0) + (e_big ? (uint64_t)e.sl : 0);
    if (total > UINT32_MAX) return NULL;
    int64_t n_off = (t.v && !t.scalar ? t.n : 0) + (e.v && !e.scalar ? e.n : 0);
    ray_t* off_hdr = NULL;
    uint32_t* newoff = (uint32_t*)scratch_alloc(&off_hdr, (size_t)(n_off > 0 ? n_off : 1) * sizeof(uint32_t));
    if (!newoff) return NULL;
    uint32_t* t_off = (t.v && !t.scalar) ? newoff : NULL;
    uint32_t* e_off = (e.v && !e.scalar) ? newoff + (t.v && !t.scalar ? t.n : 0) : NULL;
    uint64_t run = 0;
    if (t_off) if_str_side_offsets(&t, t_off, &run);
    if (e_off) if_str_side_offsets(&e, e_off, &run);
    uint32_t ts_off = 0, es_off = 0;
    ray_t* np = ray_alloc(total > 0 ? (size_t)total : 1);
    if (!np || RAY_IS_ERR(np)) { scratch_free(off_hdr); return NULL; }
    np->type = RAY_U8;
    np->len  = (int64_t)total;
    char* dst_bytes = (char*)ray_data(np);
    if (t_big) { memcpy(dst_bytes + run, t.sp, t.sl); ts_off = (uint32_t)run; run += t.sl; }
    if (e_big) { memcpy(dst_bytes + run, e.sp, e.sl); es_off = (uint32_t)run; run += e.sl; }
    result->str_pool = np;
    if_str_scatter_side(&t, t_off, dst_bytes, ts_off, dst);
    if_str_scatter_side(&e, e_off, dst_bytes, es_off, dst);
    scratch_free(off_hdr);
    return result;
}

/* Runtime-id translation for one SYM branch of an `if`.
 *
 * A branch that scans a FILE-domain column used to go through the
 * domain's whole-vocabulary LUT (ray_sym_domain_runtime_lut): the first
 * request interns EVERY entry of the vocabulary into the global table —
 * tens of GB, permanently, for a 20M-entry column — even when the branch
 * touches a few of them.  This translator interns only the positions the
 * branch actually reads: `lut` starts at -1 and a slot is resolved (raw
 * file bytes when available, the domain atom otherwise) the first time
 * that position is met.  Resolution interns, so it runs on the calling
 * thread only: the serial scatter path resolves on the fly; the parallel
 * fill first walks the mask serially and resolves exactly the positions
 * the fill will read — a row's then-cell where the mask is set, its
 * else-cell where it is clear (if_sym_side_prepare) — and the fill then
 * reads only that side of each row through if_sym_side_lookup, which
 * never interns.  Runtime-domain, STR and scalar branches need no table. */
typedef struct {
    ray_t*                   v;
    struct ray_sym_domain_s* dom;      /* non-runtime domain, else NULL */
    int64_t                  dn;
    int32_t*                 lut;      /* dn slots, -1 = unresolved (ids fit: the table counts them in 32 bits) */
    ray_t*                   lut_hdr;
    ray_sym_domain_raw_t     raw;
    bool                     raw_ok;
} if_sym_side_t;

static bool if_sym_side_init(if_sym_side_t* side, ray_t* v) {
    memset(side, 0, sizeof(*side));
    side->v = v;
    if (!v || ray_is_atom(v) || !RAY_IS_SYM(v->type)) return true;
    struct ray_sym_domain_s* dom = ray_sym_vec_domain(v);
    if (!dom || dom == ray_sym_runtime_domain()) return true;
    int64_t dn = ray_sym_domain_count(dom);
    if (dn <= 0) return true;
    side->lut = (int32_t*)scratch_alloc(&side->lut_hdr, (size_t)dn * sizeof(int32_t));
    if (!side->lut) return false;
    memset(side->lut, 0xff, (size_t)dn * sizeof(int32_t));
    side->dom = dom;
    side->dn = dn;
    side->raw_ok = ray_sym_domain_raw_pin(dom, &side->raw);
    return true;
}

static void if_sym_side_free(if_sym_side_t* side) {
    if (side->raw_ok) ray_sym_domain_raw_unpin(side->dom);
    scratch_free(side->lut_hdr);
    side->lut = NULL;
    side->lut_hdr = NULL;
}

/* Runtime id of the vocabulary position `pos`; interns on first use.
 * Calling thread only. */
static int64_t if_sym_side_resolve(if_sym_side_t* side, int64_t pos) {
    if (pos < 0 || pos >= side->dn) return -1;
    int64_t id = side->lut[pos];
    if (id >= 0) return id;
    const char* sp = NULL;
    size_t sl = 0;
    if (side->raw_ok && pos < side->raw.count) {
        sp = ray_sym_domain_raw_str(&side->raw, pos, &sl);
    } else {
        ray_t* s = ray_sym_domain_str(side->dom, pos);
        if (s) { sp = ray_str_ptr(s); sl = ray_str_len(s); }
    }
    id = ray_sym_intern(sp ? sp : "", sp ? sl : 0);
    if (id >= 0 && id <= INT32_MAX) side->lut[pos] = (int32_t)id;
    return id;
}

/* Read-only lookup for pool workers: the position must have been
 * prepared; nothing is interned here.  An unprepared position (a bug in
 * the caller's prepare pass) maps to the empty symbol rather than
 * touching the table from a worker. */
static inline int64_t if_sym_side_lookup(const if_sym_side_t* side, int64_t pos) {
    if (pos < 0 || pos >= side->dn) return 0;
    int32_t id = side->lut[pos];
    return id >= 0 ? (int64_t)id : 0;
}

/* Pre-resolve the positions rows [0, len) read on this side under the
 * mask (`want` selects which mask value picks this side).  Serial. */
static void if_sym_side_prepare(if_sym_side_t* side, const uint8_t* cond,
                                int64_t len, uint8_t want) {
    if (!side->dom) return;
    const void* base = ray_data(side->v);
    for (int64_t i = 0; i < len; i++) {
        if ((cond[i] != 0) != (want != 0)) continue;
        int64_t pos = ray_read_sym(base, i, side->v->type, side->v->attrs);
        if (pos >= 0 && pos < side->dn && side->lut[pos] < 0)
            (void)if_sym_side_resolve(side, pos);
    }
}

/* Cell value as a runtime id.  A FILE-domain cell reads the side's table
 * (resolving on the calling thread when not prepared). */
static int64_t if_sym_cell_value(if_sym_side_t* side, int64_t src) {
    ray_t* value = side->v;
    if (value->type == -RAY_STR) return ray_sym_intern(ray_str_ptr(value), ray_str_len(value));
    if (value->type == RAY_STR) {
        size_t sl = 0;
        const char* sp = ray_str_vec_get(value, src, &sl);
        return ray_sym_intern(sp ? sp : "", sp ? sl : 0);
    }
    if (ray_is_atom(value)) return sym_scalar_runtime_id(value);
    if (!side->dom) return sym_cell_runtime_id(value, src);
    int64_t pos = ray_read_sym(ray_data(value), src, value->type, value->attrs);
    if (pos >= 0 && pos < side->dn && side->lut[pos] >= 0) return side->lut[pos];
    return if_sym_side_resolve(side, pos);
}

static bool if_scatter_sym(ray_t* result, ray_t* value, int64_t* ids,
                           int64_t count, int64_t nrows) {
    if (!value) return true;
    if_sym_side_t side;
    if (!if_sym_side_init(&side, value)) return false;
    int64_t* dst = (int64_t*)ray_data(result);
    bool ok = true;
    for (int64_t j = 0; j < count && ok; j++) {
        int64_t src = if_value_index(value, ids, j, count, nrows);
        if (src < 0) { ok = false; break; }
        dst[ids[j]] = if_sym_cell_value(&side, src);
    }
    if_sym_side_free(&side);
    return ok;
}

static bool if_lazy_supported_type(int8_t out_type) {
    return out_type == RAY_F64 || out_type == RAY_I64 ||
           out_type == RAY_I32 || out_type == RAY_I16 ||
           out_type == RAY_BOOL || out_type == RAY_U8 ||
           out_type == RAY_TIMESTAMP || out_type == RAY_TIME ||
           out_type == RAY_DATE || out_type == RAY_STR ||
           out_type == RAY_SYM;
}

/* A branch that is just a column scan or a scalar constant costs nothing
 * to evaluate over ALL rows — for such branches the lazy selected path's
 * scaffolding (serial true-count + id-list build + per-branch gather +
 * scatter over nrows) strictly loses to the eager elementwise fill, which
 * is pool-parallel.  The selected path pays off only when a branch is an
 * EXPRESSION whose evaluation is worth restricting to its passing rows. */
static bool if_branch_trivial(ray_graph_t* g, ray_op_t* op) {
    if (!op) return false;
    if (op->opcode == OP_ALIAS || op->opcode == OP_MATERIALIZE)
        return if_branch_trivial(g, op_child(g, op, 0));
    if (op->opcode == OP_SCAN) return true;
    if (op->opcode == OP_CONST) {
        ray_op_ext_t* ext = find_ext(g, op->id);
        return ext && ext->literal && ray_is_atom(ext->literal);
    }
    return false;
}

/* A branch that is cheap to evaluate over ALL rows and total (no row can
 * fail): scans, constants, substrings, string case/trim/length, add/sub/mul,
 * comparisons, and/or/not, and `if` over such — a descriptor-only STR
 * result then costs less than the selected path's compaction of the rows
 * for each branch (a serial gather of the branch's columns) plus its
 * scatter.  Searches (str-find, like, replace, concat) and everything else
 * are worth restricting to the rows that need them. */
static bool if_branch_cheap(ray_graph_t* g, ray_op_t* op, int depth) {
    if (!op || depth > 32) return false;
    switch (op->opcode) {
    case OP_ALIAS: case OP_MATERIALIZE:
        return if_branch_cheap(g, op_child(g, op, 0), depth + 1);
    case OP_SCAN: return true;
    case OP_CONST: {
        ray_op_ext_t* ext = find_ext(g, op->id);
        return ext && ext->literal && ray_is_atom(ext->literal);
    }
    case OP_SUBSTR: case OP_IF: {
        ray_op_ext_t* ext = find_ext(g, op->id);
        if (!ext || ext->third_in >= g->node_count) return false;
        for (int k = 0; k < op->arity && k < 2; k++)
            if (!if_branch_cheap(g, op_child(g, op, k), depth + 1)) return false;
        return if_branch_cheap(g, op_node(g, ext->third_in), depth + 1);
    }
    case OP_STRLEN: case OP_UPPER: case OP_LOWER: case OP_TRIM:
    case OP_ADD: case OP_SUB: case OP_MUL: case OP_NEG: case OP_ABS:
    case OP_EQ: case OP_NE: case OP_LT: case OP_LE: case OP_GT: case OP_GE:
    case OP_AND: case OP_OR: case OP_NOT: case OP_ISNULL:
        for (int k = 0; k < op->arity && k < 2; k++)
            if (!if_branch_cheap(g, op_child(g, op, k), depth + 1)) return false;
        return true;
    default:
        return false;
    }
}

/* Is a trivial branch of static type `bt` filled CORRECTLY by the eager
 * elementwise path for result type `out`?  Mixed numeric/string branch
 * combinations (e.g. `(if c n1 s2)` with I64 + STR) rely on the selected
 * path's per-value string conversion — keep those there. */
static bool if_type_eager_ok(int8_t bt, int8_t out) {
    int8_t b = bt < 0 ? (int8_t)-bt : bt;
    if (out == RAY_SYM)
        return b == RAY_SYM || bt == -RAY_STR;
    switch (b) {
        /* No RAY_F32: if_fill_range has no F32 fill case, so an F32 branch
         * must stay on the selected path (unreachable today — kept off the
         * whitelist so a future if_lazy_supported_type change cannot route
         * it to an unwritten result buffer). */
        case RAY_F64: case RAY_I64: case RAY_I32: case RAY_I16:
        case RAY_U8: case RAY_BOOL:
        case RAY_TIMESTAMP: case RAY_TIME: case RAY_DATE: case RAY_GUID:
            return true;
        default: return false;
    }
}

static ray_t* exec_if_selected(ray_graph_t* g, ray_op_t* op, ray_t* cond_v) {
    if (!g || !g->table || !cond_v || cond_v->type != RAY_BOOL)
        return NULL;
    int64_t nrows = ray_table_nrows(g->table);
    if (cond_v->len != nrows) return NULL;

    if (!if_lazy_supported_type(op->out_type)) return NULL;

    ray_op_t* then_op = op_child(g, op, 1);
    ray_op_ext_t* ext = find_ext(g, op->id);
    ray_op_t* else_op = ext ? op_node(g, ext->third_in) : NULL;
    if (!then_op || !else_op) return NULL;

    /* Both branches trivial AND type-safe for the eager fill → eager
     * (POOL-PARALLEL elementwise) wins over the serial id-scaffolding
     * here.  Only when workers exist: with a 0-worker pool (-c 1) the
     * eager fill runs serially over ALL rows for BOTH branches, which
     * loses to the selected path's touch-only-passing-rows scheme. */
    /* Whether the eager arm is even a candidate for this node — not whether
     * it runs, which also depends on the pool.  Where it is, this arm must
     * answer exactly as it would, or the same expression gets two types
     * depending on the worker count.  Where it is not, this arm is the only
     * one there is and may report the type the rows actually have. */
    ray_t* outer_sel = g->selection;
    if (outer_sel) {
        ray_rowsel_t* sm = ray_rowsel_meta(outer_sel);
        if (!sm || sm->nrows != nrows) return NULL;
    }

    int64_t selected = outer_sel ? ray_rowsel_meta(outer_sel)->total_pass : nrows;
    if (selected < 0 || selected > nrows) return NULL;
    /* Under a sparse outer selection (a `where:` keeping a quarter of the
     * rows or fewer) the STR eager arm is a bad trade: it builds a
     * descriptor for every row of the table, and the filtered result then
     * keeps that whole intermediate — both parents' bytes — alive for its
     * few rows.  Touching only the selected rows costs less and holds only
     * their bytes.  Same threshold as the projection's pre-compaction. */
    bool sparse_sel = outer_sel && selected * 4 <= nrows;
    /* The SYM text build below writes a fresh pool for every row of the
     * table, which the filtered result then keeps: only worth it when the
     * selection keeps nearly every row. */
    bool near_full_sel = !outer_sel || selected * 10 >= nrows * 9;

    bool eager_possible = (op->out_type != RAY_STR &&
                           if_branch_trivial(g, then_op) &&
                           if_branch_trivial(g, else_op) &&
                           if_type_eager_ok(then_op->out_type, op->out_type) &&
                           if_type_eager_ok(else_op->out_type, op->out_type)) ||
                          /* two STR vector branches that are cheap and total:
                           * the eager arm picks descriptors over one pass */
                          (op->out_type == RAY_STR && !sparse_sel &&
                           then_op->out_type == RAY_STR && else_op->out_type == RAY_STR &&
                           if_branch_cheap(g, then_op, 0) && if_branch_cheap(g, else_op, 0)) ||
                          /* a SYM column or literal against a SYM or STR
                           * one, over enough rows to split: the eager arm
                           * resolves each row's text on the pool
                           * (if_str_text_build).  Below that the selected
                           * arm's per-row scatter costs less. */
                          (op->out_type == RAY_STR && near_full_sel && nrows >= RAY_PARALLEL_THRESHOLD &&
                           (then_op->out_type == RAY_SYM || else_op->out_type == RAY_SYM) &&
                           (then_op->out_type == RAY_SYM || then_op->out_type == RAY_STR) &&
                           (else_op->out_type == RAY_SYM || else_op->out_type == RAY_STR) &&
                           if_branch_trivial(g, then_op) && if_branch_trivial(g, else_op));
    {
        ray_pool_t* rp = ray_pool_get();
        if (rp && rp->n_workers > 0 && eager_possible)
            return NULL;
    }

    if_branch_plan_t then_plan, else_plan;
    if (!if_make_branch_plan(g, then_op, &then_plan) ||
        !if_make_branch_plan(g, else_op, &else_plan))
        return NULL;

    uint8_t* cond = (uint8_t*)ray_data(cond_v);
    int64_t true_count = if_selected_true_count(cond, nrows, outer_sel);
    int64_t false_count = selected - true_count;
    if (true_count < 0 || false_count < 0) return NULL;

    ray_t* ids_hdr = NULL;
    int64_t* ids = NULL;
    if (selected > 0) {
        ids = (int64_t*)scratch_alloc(&ids_hdr, (size_t)selected * sizeof(int64_t));
        if (!ids) return ray_error("oom", NULL);
        if_fill_branch_ids(cond, nrows, outer_sel, ids, ids + true_count);
    }
    int64_t* true_ids = ids;
    int64_t* false_ids = ids ? ids + true_count : NULL;

    ray_t* then_v = if_eval_branch(g, then_op, &then_plan, true_ids, true_count, nrows);
    if (then_v && RAY_IS_ERR(then_v)) { scratch_free(ids_hdr); return then_v; }
    ray_t* else_v = if_eval_branch(g, else_op, &else_plan, false_ids, false_count, nrows);
    if (else_v && RAY_IS_ERR(else_v)) {
        if (then_v) ray_release(then_v);
        scratch_free(ids_hdr);
        return else_v;
    }

    int8_t out_type = if_selected_type(then_v, else_v, op->out_type, eager_possible);
    if (!if_lazy_supported_type(out_type)) {
        if (then_v) ray_release(then_v);
        if (else_v) ray_release(else_v);
        scratch_free(ids_hdr);
        return NULL;
    }

    ray_t* result = ray_vec_new(out_type, nrows);
    if (!result || RAY_IS_ERR(result)) {
        if (then_v) ray_release(then_v);
        if (else_v) ray_release(else_v);
        scratch_free(ids_hdr);
        return result;
    }
    result->len = nrows;
    if (if_val_has_null(then_v) || if_val_has_null(else_v))
        result->attrs |= RAY_ATTR_HAS_NULLS;
    if (out_type == RAY_STR)
        memset(ray_data(result), 0, (size_t)nrows * sizeof(ray_str_t));
    else
        memset(ray_data(result), 0, (size_t)nrows * ray_sym_elem_size(out_type, result->attrs));

    bool ok = true;
    if (out_type == RAY_STR) {
        ray_t* fast = if_scatter_str_desc(result, then_v, true_ids, true_count,
                                          else_v, false_ids, false_count, nrows);
        if (fast) {
            result = fast;
        } else {
            result = if_scatter_str(result, then_v, true_ids, true_count, nrows);
            if (result && !RAY_IS_ERR(result))
                result = if_scatter_str(result, else_v, false_ids, false_count, nrows);
        }
    } else if (out_type == RAY_SYM) {
        ok = if_scatter_sym(result, then_v, true_ids, true_count, nrows) &&
             if_scatter_sym(result, else_v, false_ids, false_count, nrows);
    } else {
        ok = if_scatter_numeric(result, then_v, true_ids, true_count, nrows, out_type) &&
             if_scatter_numeric(result, else_v, false_ids, false_count, nrows, out_type);
    }

    if (then_v) ray_release(then_v);
    if (else_v) ray_release(else_v);
    scratch_free(ids_hdr);

    if (!result || RAY_IS_ERR(result)) return result;
    if (!ok) {
        ray_release(result);
        return ray_error("length", "if: branch length mismatch");
    }
    return result;
}

/* ============================================================================
 * OP_IF: ternary select  result[i] = cond[i] ? then[i] : else[i]
 * ============================================================================ */

/* Shared elementwise fill for the fixed-width OP_IF branches — one body
 * driven serially (0..len) or as a pool worker over disjoint ranges.
 * STR output stays in the serial append path (variable-width build).
 * SYM ranges are dispatch-safe only after the caller warms each side's
 * runtime-id LUT (sym.c frozen-table rule; mirrors window.c setup). */
typedef struct {
    const uint8_t* cond;
    ray_t*  then_v;
    ray_t*  else_v;
    bool    then_scalar;
    bool    else_scalar;
    double  t_f64, e_f64;     /* pre-resolved scalar broadcasts */
    int64_t t_i64, e_i64;
    void*   dst;
    int8_t  out_type;
    if_sym_side_t* t_side;    /* SYM output: per-side id translation */
    if_sym_side_t* e_side;
    bool    sides_prepared;   /* positions resolved up front: read-only fill (workers) */
} if_fill_ctx_t;

/* One SYM cell of a fill side as a runtime id: the selected side only. */
static inline int64_t if_fill_sym_cell(const if_fill_ctx_t* c, if_sym_side_t* side,
                                       int64_t i) {
    ray_t* v = side->v;
    if (c->sides_prepared && side->dom) {
        int64_t pos = ray_read_sym(ray_data(v), i, v->type, v->attrs);
        return if_sym_side_lookup(side, pos);
    }
    return if_sym_cell_value(side, i);
}

static void if_fill_range(const if_fill_ctx_t* c, int64_t i0, int64_t i1) {
    const uint8_t* cond_p = c->cond;
    switch (c->out_type) {
    case RAY_F64: {
        double* dst = (double*)c->dst;
        for (int64_t i = i0; i < i1; i++) {
            double tv = c->then_scalar ? c->t_f64 : if_vec_f64(c->then_v, i);
            double ev = c->else_scalar ? c->e_f64 : if_vec_f64(c->else_v, i);
            dst[i] = cond_p[i] ? tv : ev;
        }
        break; }
    case RAY_I64: case RAY_TIMESTAMP: {
        int64_t* dst = (int64_t*)c->dst;
        for (int64_t i = i0; i < i1; i++) {
            int64_t tv = c->then_scalar ? c->t_i64 : if_vec_i64(c->then_v, i);
            int64_t ev = c->else_scalar ? c->e_i64 : if_vec_i64(c->else_v, i);
            dst[i] = cond_p[i] ? tv : ev;
        }
        break; }
    case RAY_SYM: {
        int64_t* dst = (int64_t*)c->dst;
        /* Only the chosen side is read: the prepare pass resolved exactly
         * these cells, and the other side's cell may be unresolved. */
        for (int64_t i = i0; i < i1; i++) {
            if (cond_p[i])
                dst[i] = c->then_scalar ? c->t_i64 : if_fill_sym_cell(c, c->t_side, i);
            else
                dst[i] = c->else_scalar ? c->e_i64 : if_fill_sym_cell(c, c->e_side, i);
        }
        break; }
    case RAY_I32: case RAY_TIME: case RAY_DATE: {
        int32_t* dst = (int32_t*)c->dst;
        for (int64_t i = i0; i < i1; i++) {
            int32_t tv = c->then_scalar ? (int32_t)c->t_i64 : (int32_t)if_vec_i64(c->then_v, i);
            int32_t ev = c->else_scalar ? (int32_t)c->e_i64 : (int32_t)if_vec_i64(c->else_v, i);
            dst[i] = cond_p[i] ? tv : ev;
        }
        break; }
    case RAY_I16: {
        int16_t* dst = (int16_t*)c->dst;
        for (int64_t i = i0; i < i1; i++) {
            int16_t tv = c->then_scalar ? (int16_t)c->t_i64 : (int16_t)if_vec_i64(c->then_v, i);
            int16_t ev = c->else_scalar ? (int16_t)c->e_i64 : (int16_t)if_vec_i64(c->else_v, i);
            dst[i] = cond_p[i] ? tv : ev;
        }
        break; }
    case RAY_BOOL: case RAY_U8: {
        uint8_t* dst = (uint8_t*)c->dst;
        for (int64_t i = i0; i < i1; i++) {
            uint8_t tv = c->then_scalar ? (uint8_t)c->t_i64 : (uint8_t)if_vec_i64(c->then_v, i);
            uint8_t ev = c->else_scalar ? (uint8_t)c->e_i64 : (uint8_t)if_vec_i64(c->else_v, i);
            dst[i] = cond_p[i] ? tv : ev;
        }
        break; }
    case RAY_GUID: {
        /* 16-byte cells: copy the chosen side's cell.  A GUID atom keeps
         * its bytes in obj; a one-element vector is row 0. */
        uint8_t* dst = (uint8_t*)c->dst;
        const uint8_t* tb = ray_is_atom(c->then_v) ? (const uint8_t*)ray_data(c->then_v->obj)
                                                   : (const uint8_t*)ray_data(c->then_v);
        const uint8_t* eb = ray_is_atom(c->else_v) ? (const uint8_t*)ray_data(c->else_v->obj)
                                                   : (const uint8_t*)ray_data(c->else_v);
        for (int64_t i = i0; i < i1; i++) {
            const uint8_t* src = cond_p[i]
                ? tb + (c->then_scalar ? 0 : i) * 16
                : eb + (c->else_scalar ? 0 : i) * 16;
            memcpy(dst + i * 16, src, 16);
        }
        break; }
    default: break;
    }
}

static void if_fill_par_fn(void* ctx, uint32_t wid, int64_t start, int64_t end) {
    (void)wid;
    if_fill_range((const if_fill_ctx_t*)ctx, start, end);
}

/* Descriptor fill for the STR arm of exec_if_eager: dst[r] is the chosen
 * side's descriptor; a pooled else-descriptor moves by e_shift when the two
 * pools were laid end to end.  Rows are independent; workers write disjoint
 * ranges. */
typedef struct {
    const uint8_t*   cond;
    const ray_str_t* t;
    const ray_str_t* e;
    ray_str_t*       dst;
    /* two pools: the chosen row's pooled bytes move to dst_bytes at newoff[r] */
    const uint32_t*  newoff;
    const char*      t_bytes;
    const char*      e_bytes;
    char*            dst_bytes;
} if_str_desc_ctx_t;

static void if_str_desc_fn(void* vctx, uint32_t worker_id, int64_t start, int64_t end) {
    (void)worker_id;
    const if_str_desc_ctx_t* c = (const if_str_desc_ctx_t*)vctx;
    for (int64_t r = start; r < end; r++) {
        ray_str_t d = c->cond[r] ? c->t[r] : c->e[r];
        if (c->newoff && !ray_str_is_inline(&d)) {
            const char* src = c->cond[r] ? c->t_bytes : c->e_bytes;
            memcpy(c->dst_bytes + c->newoff[r], src + d.pool_off, d.len);
            d.pool_off = c->newoff[r];
        }
        c->dst[r] = d;
    }
}

/* Truthiness of one cell of an `if` condition vector, the rule is_truthy
 * applies to an atom: null and zero are false, anything else is true. */
static uint8_t if_cond_cell(ray_t* c, int64_t i) {
    const void* d = ray_data(c);
    switch (c->type) {
    case RAY_BOOL: case RAY_U8:
        return ((const uint8_t*)d)[i] != 0;
    default: break;
    }
    if (ray_vec_is_null(c, i)) return 0;
    switch (c->type) {
    case RAY_I16:   return ((const int16_t*)d)[i] != 0;
    case RAY_I32: case RAY_DATE: case RAY_TIME:
                    return ((const int32_t*)d)[i] != 0;
    case RAY_I64: case RAY_TIMESTAMP:
                    return ((const int64_t*)d)[i] != 0;
    case RAY_F32: { float f = ((const float*)d)[i]; return f != 0.0f && f == f; }
    case RAY_F64: { double f = ((const double*)d)[i]; return f != 0.0 && f == f; }
    default:        return 0;   /* refused by if_cond_type_ok first */
    }
}

/* Truthiness of an atom condition: is_truthy's rule (lang/internal.h, which
 * this file cannot include), kept in step with it. */
static uint8_t if_cond_atom(ray_t* x) {
    if (RAY_ATOM_IS_NULL(x)) return 0;
    switch (x->type) {
    case -RAY_BOOL: return x->b8 != 0;
    case -RAY_U8:   return x->u8 != 0;
    case -RAY_I16:  return x->i16 != 0;
    case -RAY_I32:  return x->i32 != 0;
    case -RAY_I64: case -RAY_DATE: case -RAY_TIME: case -RAY_TIMESTAMP:
        return x->i64 != 0;
    case -RAY_F32: case -RAY_F64:
        return x->f64 != 0.0 && x->f64 == x->f64;
    default:
        return 1;
    }
}

/* A condition vector needs a truth value per cell: booleans, numbers and
 * temporals have one.  Text, GUID and boxed lists do not. */
static bool if_cond_type_ok(int8_t t) {
    switch (t) {
    case RAY_BOOL: case RAY_U8: case RAY_I16: case RAY_I32: case RAY_I64:
    case RAY_F32: case RAY_F64: case RAY_DATE: case RAY_TIME: case RAY_TIMESTAMP:
        return true;
    default:
        return false;
    }
}

/* STR output with a SYM side (the other a SYM or STR vector, or a
 * broadcast scalar): each row takes its chosen side's text into a fresh
 * descriptor; texts past the inline width go to one pool laid out in row
 * order.  A SYM cell is read through a snapshot taken here, serially,
 * before any worker runs — the runtime table's string array
 * (ray_sym_strings_borrow) or a FILE domain's raw vocabulary
 * (ray_sym_domain_raw_pin) — so nothing is interned or materialised and
 * both passes run on the pool.  A cell neither snapshot can read (a
 * FILE-domain position appended after the file was mapped, an id past the
 * table) makes the build decline; the caller then appends serially. */
enum { IF_TXT_SCALAR = 1, IF_TXT_STR, IF_TXT_SYM_RT, IF_TXT_SYM_FILE };

typedef struct {
    uint8_t                  kind;
    const char*              sp;          /* scalar text */
    uint32_t                 sl;
    const ray_str_t*         desc;        /* STR vector */
    const char*              bytes;
    const void*              sym;         /* SYM vector cells */
    int8_t                   sym_type;
    uint8_t                  sym_attrs;
    ray_t**                  rt_strings;  /* runtime-domain snapshot */
    uint32_t                 rt_count;
    struct ray_sym_domain_s* dom;         /* FILE domain, pinned */
    ray_sym_domain_raw_t     raw;
} if_txt_side_t;

static bool if_txt_side_init(if_txt_side_t* s, ray_t* v, bool scalar, int64_t len) {
    memset(s, 0, sizeof(*s));
    if (scalar) {
        const char* sp = "";
        size_t sl = 0;
        if (v->type == -RAY_STR) {
            sp = ray_str_ptr(v); sl = ray_str_len(v);
        } else if (v->type == RAY_STR) {
            sp = ray_str_vec_get(v, 0, &sl);
            if (!sp) { sp = ""; sl = 0; }
        } else if (v->type == -RAY_SYM || RAY_IS_SYM(v->type)) {
            ray_t* a = sym_scalar_str(v);
            sp = a ? ray_str_ptr(a) : "";
            sl = a ? ray_str_len(a) : 0;
        } else {
            return false;
        }
        if (sl > UINT32_MAX) return false;
        s->kind = IF_TXT_SCALAR; s->sp = sp; s->sl = (uint32_t)sl;
        return true;
    }
    if (v->len < len) return false;
    if (v->type == RAY_STR) {
        s->kind = IF_TXT_STR;
        str_resolve(v, &s->desc, &s->bytes);
        return true;
    }
    if (!RAY_IS_SYM(v->type)) return false;
    s->sym = ray_data(v);
    s->sym_type = v->type;
    s->sym_attrs = v->attrs;
    struct ray_sym_domain_s* dom = ray_sym_vec_domain(v);
    if (dom == ray_sym_runtime_domain()) {
        ray_sym_strings_borrow(&s->rt_strings, &s->rt_count);
        if (!s->rt_strings) return false;
        s->kind = IF_TXT_SYM_RT;
        return true;
    }
    if (!ray_sym_domain_raw_pin(dom, &s->raw)) return false;
    s->dom = dom;
    s->kind = IF_TXT_SYM_FILE;
    return true;
}

static void if_txt_side_free(if_txt_side_t* s) {
    if (s->kind == IF_TXT_SYM_FILE) ray_sym_domain_raw_unpin(s->dom);
    s->kind = 0;
}

/* Text of row r on this side; false for a cell the snapshot cannot read. */
static inline bool if_txt_get(const if_txt_side_t* s, int64_t r,
                              const char** sp, uint32_t* sl) {
    switch (s->kind) {
    case IF_TXT_SCALAR:
        *sp = s->sp; *sl = s->sl;
        return true;
    case IF_TXT_STR: {
        const ray_str_t* d = &s->desc[r];
        *sl = d->len;
        *sp = d->len == 0 ? "" : ray_str_is_inline(d) ? d->data : s->bytes + d->pool_off;
        return true;
    }
    case IF_TXT_SYM_RT: {
        int64_t id = ray_read_sym(s->sym, r, s->sym_type, s->sym_attrs);
        if (id < 0 || id >= (int64_t)s->rt_count || !s->rt_strings[id]) return false;
        ray_t* a = s->rt_strings[id];
        *sp = ray_str_ptr(a); *sl = (uint32_t)ray_str_len(a);
        return true;
    }
    case IF_TXT_SYM_FILE: {
        int64_t pos = ray_read_sym(s->sym, r, s->sym_type, s->sym_attrs);
        if (pos < 0 || pos >= s->raw.count) return false;
        size_t l = 0;
        *sp = ray_sym_domain_raw_str(&s->raw, pos, &l);
        *sl = (uint32_t)l;
        return true;
    }
    default:
        return false;
    }
}

typedef struct {
    const uint8_t*       cond;
    const if_txt_side_t* t;
    const if_txt_side_t* e;
    uint32_t*            off;    /* pooled length per row, then its pool offset */
    ray_str_t*           dst;
    char*                pool;
    _Atomic(int)         miss;
} if_txt_ctx_t;

static void if_txt_len_fn(void* vctx, uint32_t wid, int64_t start, int64_t end) {
    (void)wid;
    if_txt_ctx_t* c = (if_txt_ctx_t*)vctx;
    for (int64_t r = start; r < end; r++) {
        const char* sp; uint32_t sl;
        if (!if_txt_get(c->cond[r] ? c->t : c->e, r, &sp, &sl)) {
            atomic_store_explicit(&c->miss, 1, memory_order_relaxed);
            return;
        }
        c->off[r] = sl > RAY_STR_INLINE_MAX ? sl : 0;
    }
}

static void if_txt_fill_fn(void* vctx, uint32_t wid, int64_t start, int64_t end) {
    (void)wid;
    if_txt_ctx_t* c = (if_txt_ctx_t*)vctx;
    for (int64_t r = start; r < end; r++) {
        const char* sp = ""; uint32_t sl = 0;
        if (!if_txt_get(c->cond[r] ? c->t : c->e, r, &sp, &sl)) { sp = ""; sl = 0; }
        ray_str_t* d = &c->dst[r];
        memset(d, 0, sizeof(*d));
        d->len = sl;
        if (sl <= RAY_STR_INLINE_MAX) {
            if (sl) memcpy(d->data, sp, sl);
        } else {
            memcpy(d->prefix, sp, 4);
            d->pool_off = c->off[r];
            memcpy(c->pool + c->off[r], sp, sl);
        }
    }
}

/* Fills `result` (STR, len rows) and returns it, or NULL with `result`
 * untouched when the build does not apply. */
static ray_t* if_str_text_build(ray_t* result, const uint8_t* cond, int64_t len,
                                ray_t* then_v, bool then_scalar,
                                ray_t* else_v, bool else_scalar) {
    bool t_sym = then_v->type == -RAY_SYM || RAY_IS_SYM(then_v->type);
    bool e_sym = else_v->type == -RAY_SYM || RAY_IS_SYM(else_v->type);
    if (!t_sym && !e_sym) return NULL;
    if (len <= 0) return NULL;

    if_txt_side_t t, e;
    if (!if_txt_side_init(&t, then_v, then_scalar, len)) {
        if_txt_side_free(&t);
        return NULL;
    }
    if (!if_txt_side_init(&e, else_v, else_scalar, len)) {
        if_txt_side_free(&t); if_txt_side_free(&e);
        return NULL;
    }
    ray_t* off_hdr = NULL;
    uint32_t* off = (uint32_t*)scratch_alloc(&off_hdr, (size_t)len * sizeof(uint32_t));
    if (!off) { if_txt_side_free(&t); if_txt_side_free(&e); return NULL; }

    if_txt_ctx_t c = { .cond = cond, .t = &t, .e = &e, .off = off,
                       .dst = (ray_str_t*)ray_data(result), .pool = NULL };
    atomic_init(&c.miss, 0);
    ray_pool_t* pool = ray_pool_get();
    bool par = ray_pool_par_dispatch_ok(pool, len, RAY_PARALLEL_THRESHOLD);
    if (par) ray_pool_dispatch(pool, if_txt_len_fn, &c, len);
    else     if_txt_len_fn(&c, 0, 0, len);

    uint64_t run = 0;
    bool ok = atomic_load_explicit(&c.miss, memory_order_relaxed) == 0;
    if (ok) {
        for (int64_t r = 0; r < len; r++) {
            uint32_t l = off[r];
            off[r] = (uint32_t)run;
            run += l;
            if (run > UINT32_MAX) { ok = false; break; }
        }
    }
    ray_t* np = NULL;
    if (ok && run > 0) {
        np = ray_alloc((size_t)run);
        if (!np || RAY_IS_ERR(np)) { np = NULL; ok = false; }
        else { np->type = RAY_U8; np->len = (int64_t)run; c.pool = (char*)ray_data(np); }
    }
    if (ok) {
        if (par) ray_pool_dispatch(pool, if_txt_fill_fn, &c, len);
        else     if_txt_fill_fn(&c, 0, 0, len);
        result->str_pool = np;
    }
    scratch_free(off_hdr);
    if_txt_side_free(&t);
    if_txt_side_free(&e);
    return ok ? result : NULL;
}

/* `cond_in`: the condition exec_if already evaluated (ownership passes
 * here), or NULL to evaluate it. */
static ray_t* exec_if_eager(ray_graph_t* g, ray_op_t* op, ray_t* cond_in) {
    /* cond = inputs[0], then = inputs[1], else_id stored in ext->third_in */
    ray_t* cond_v = cond_in ? cond_in : exec_node(g, op_child(g, op, 0));
    ray_t* then_v = exec_node(g, op_child(g, op, 1));

    ray_op_ext_t* ext = find_ext(g, op->id);
    uint32_t else_id = ext->third_in;
    ray_t* else_v = exec_node(g, op_node(g, else_id));

    if (!cond_v || RAY_IS_ERR(cond_v)) {
        if (then_v && !RAY_IS_ERR(then_v)) ray_release(then_v);
        if (else_v && !RAY_IS_ERR(else_v)) ray_release(else_v);
        return cond_v;
    }
    if (!then_v || RAY_IS_ERR(then_v)) {
        ray_release(cond_v);
        if (else_v && !RAY_IS_ERR(else_v)) ray_release(else_v);
        return then_v;
    }
    if (!else_v || RAY_IS_ERR(else_v)) {
        ray_release(cond_v); ray_release(then_v);
        return else_v;
    }

    /* The fills below read the condition as one BOOL byte per row.  A scalar
     * condition has one cell, and indexing it per row read past it, so
     * `(if false f 1.0)` returned a mixture of both branches.  A reduction
     * in the condition, such as `(> (count x) 3)`, reaches here as a
     * one-element vector rather than an atom and is the same scalar against
     * vector branches.  A condition of another type, such as `(count x)` or
     * an I64 column, was read byte by byte.  Every such condition becomes a
     * full-length BOOL mask, with is_truthy's rule per cell; the
     * selected arm declines all of them, which is why only this arm
     * mattered. */
    int64_t branch_len = 1;
    if (!ray_is_atom(then_v) && then_v->len > branch_len) branch_len = then_v->len;
    if (!ray_is_atom(else_v) && else_v->len > branch_len) branch_len = else_v->len;
    if (!ray_is_atom(cond_v) && !if_cond_type_ok(cond_v->type)) {
        int8_t ct = cond_v->type;
        g->if_refused = true;
        ray_release(cond_v); ray_release(then_v); ray_release(else_v);
        return ray_error("type", "if: condition must be boolean, numeric or temporal, got %s",
                         ray_type_name(ct));
    }
    bool cond_bcast = ray_is_atom(cond_v) || (cond_v->len == 1 && branch_len > 1);
    if (cond_bcast || cond_v->type != RAY_BOOL) {
        int64_t n = cond_bcast ? branch_len : cond_v->len;
        ray_t* mask = ray_vec_new(RAY_BOOL, n);
        if (!mask || RAY_IS_ERR(mask)) {
            ray_release(cond_v); ray_release(then_v); ray_release(else_v);
            return mask ? mask : ray_error("oom", NULL);
        }
        uint8_t* m = (uint8_t*)ray_data(mask);
        if (cond_bcast)
            memset(m, ray_is_atom(cond_v) ? if_cond_atom(cond_v)
                                          : if_cond_cell(cond_v, 0), (size_t)n);
        else
            for (int64_t i = 0; i < n; i++) m[i] = if_cond_cell(cond_v, i);
        mask->len = n;
        ray_release(cond_v);
        cond_v = mask;
    }

    int64_t len = cond_v->len;
    bool then_scalar = ray_is_atom(then_v) || (then_v->type > 0 && then_v->len == 1);
    bool else_scalar = ray_is_atom(else_v) || (else_v->type > 0 && else_v->len == 1);
    if (then_scalar && !else_scalar) len = else_v->len;
    if (!then_scalar) len = then_v->len;

    int8_t out_type = op->out_type;
    ray_t* result = ray_vec_new(out_type, len);
    if (!result || RAY_IS_ERR(result)) {
        ray_release(cond_v); ray_release(then_v); ray_release(else_v);
        return result;
    }
    result->len = len;
    /* Same reason as the selected arm: an atom null is the sentinel value and
     * carries no attr of its own, so without this the sentinel is rendered as
     * an ordinary date or timestamp. */
    if (if_val_has_null(then_v) || if_val_has_null(else_v))
        result->attrs |= RAY_ATTR_HAS_NULLS;

    uint8_t* cond_p = (uint8_t*)ray_data(cond_v);

    if (out_type == RAY_STR) {
        if (if_str_text_build(result, cond_p, len, then_v, then_scalar,
                              else_v, else_scalar)) {
            ray_release(cond_v); ray_release(then_v); ray_release(else_v);
            return result;
        }
        /* A SYM side the text build could not read (a symbol appended to
         * its domain after the snapshot) is resolved row by row below, from
         * the branches already materialised. */
        /* Two STR vectors: the result is descriptors only.  Each row takes
         * its side's 16-byte descriptor; pooled strings keep pointing into
         * their pool.  One shared pool (or one side inline-only) is reused
         * as is; two different pools give a pool of exactly the chosen
         * rows' bytes.  Nulls
         * are empty descriptors and travel unchanged.  No per-row append,
         * no rehash; the fill runs on the worker pool. */
        if (!then_scalar && !else_scalar &&
            then_v->type == RAY_STR && else_v->type == RAY_STR &&
            len <= then_v->len && len <= else_v->len) {
            ray_t* then_pool = str_vec_pool_obj(then_v);
            ray_t* else_pool = str_vec_pool_obj(else_v);
            const ray_str_t* t_desc = NULL;
            const ray_str_t* e_desc = NULL;
            const char* t_bytes = NULL;
            const char* e_bytes = NULL;
            str_resolve(then_v, &t_desc, &t_bytes);
            str_resolve(else_v, &e_desc, &e_bytes);
            bool ok = true;
            ray_t* off_hdr = NULL;
            uint32_t* newoff = NULL;
            if (then_pool == else_pool || !then_pool || !else_pool) {
                ray_t* out_pool = then_pool ? then_pool : else_pool;
                if (out_pool && !RAY_IS_ERR(out_pool)) {
                    ray_retain(out_pool);
                    result->str_pool = out_pool;
                }
            } else if (RAY_IS_ERR(then_pool) || RAY_IS_ERR(else_pool)) {
                ok = false;
            } else {
                /* Two pools: a pool of exactly the chosen rows' bytes (a
                 * serial pass assigns the offsets, the fill copies). */
                newoff = (uint32_t*)scratch_alloc(&off_hdr, (size_t)(len > 0 ? len : 1) * sizeof(uint32_t));
                if (!newoff) {
                    ok = false;
                } else {
                    uint64_t run = 0;
                    for (int64_t r = 0; r < len; r++) {
                        const ray_str_t* d = cond_p[r] ? &t_desc[r] : &e_desc[r];
                        if (ray_str_is_inline(d)) continue;
                        newoff[r] = (uint32_t)run;
                        run += d->len;
                    }
                    ray_t* np = (run <= UINT32_MAX) ? ray_alloc(run > 0 ? (size_t)run : 1) : NULL;
                    if (!np || RAY_IS_ERR(np)) {
                        ok = false;
                    } else {
                        np->type = RAY_U8;
                        np->len  = (int64_t)run;
                        result->str_pool = np;
                    }
                }
                if (!ok && off_hdr) { scratch_free(off_hdr); off_hdr = NULL; newoff = NULL; }
            }
            if (ok) {
                if_str_desc_ctx_t dctx = {
                    .cond = cond_p, .t = t_desc, .e = e_desc,
                    .dst = (ray_str_t*)ray_data(result),
                    .newoff = newoff, .t_bytes = t_bytes, .e_bytes = e_bytes,
                    .dst_bytes = result->str_pool ? (char*)ray_data(result->str_pool) : NULL,
                };
                ray_pool_t* pool = ray_pool_get();
                if (ray_pool_par_dispatch_ok(pool, len, RAY_PARALLEL_THRESHOLD))
                    ray_pool_dispatch(pool, if_str_desc_fn, &dctx, len);
                else
                    if_str_desc_fn(&dctx, 0, 0, len);
                if (off_hdr) scratch_free(off_hdr);
                if (ray_vec_may_have_nulls(then_v) || ray_vec_may_have_nulls(else_v))
                    result->attrs |= RAY_ATTR_HAS_NULLS;
                ray_release(cond_v); ray_release(then_v); ray_release(else_v);
                return result;
            }
        }

        /* RAY_STR: resolve each side to string data and ray_str_vec_append.
         * Scalars may be -RAY_STR or RAY_SYM atoms. */
        result->len = 0; /* ray_str_vec_append manages len */
        for (int64_t i = 0; i < len; i++) {
            const char* sp;
            size_t sl;
            if (cond_p[i]) {
                if (then_scalar) {
                    if (then_v->type == -RAY_STR) {
                        sp = ray_str_ptr(then_v);
                        sl = ray_str_len(then_v);
                    } else if (then_v->type == RAY_STR) {
                        sp = ray_str_vec_get(then_v, 0, &sl);
                        if (!sp) { sp = ""; sl = 0; }
                    } else if (RAY_IS_SYM(then_v->type) || then_v->type == -RAY_SYM) {
                        ray_t* s = sym_scalar_str(then_v);
                        sp = s ? ray_str_ptr(s) : "";
                        sl = s ? ray_str_len(s) : 0;
                    } else { sp = ""; sl = 0; }
                } else if (then_v->type == RAY_STR) {
                    sp = ray_str_vec_get(then_v, i, &sl);
                    if (!sp) { sp = ""; sl = 0; }
                } else {
                    /* RAY_SYM column: cell-data, resolve through its domain */
                    ray_t* sa = ray_sym_vec_cell(then_v, i);
                    sp = sa ? ray_str_ptr(sa) : "";
                    sl = sa ? ray_str_len(sa) : 0;
                }
            } else {
                if (else_scalar) {
                    if (else_v->type == -RAY_STR) {
                        sp = ray_str_ptr(else_v);
                        sl = ray_str_len(else_v);
                    } else if (else_v->type == RAY_STR) {
                        sp = ray_str_vec_get(else_v, 0, &sl);
                        if (!sp) { sp = ""; sl = 0; }
                    } else if (RAY_IS_SYM(else_v->type) || else_v->type == -RAY_SYM) {
                        ray_t* s = sym_scalar_str(else_v);
                        sp = s ? ray_str_ptr(s) : "";
                        sl = s ? ray_str_len(s) : 0;
                    } else { sp = ""; sl = 0; }
                } else if (else_v->type == RAY_STR) {
                    sp = ray_str_vec_get(else_v, i, &sl);
                    if (!sp) { sp = ""; sl = 0; }
                } else {
                    /* RAY_SYM column: cell-data, resolve through its domain */
                    ray_t* sa = ray_sym_vec_cell(else_v, i);
                    sp = sa ? ray_str_ptr(sa) : "";
                    sl = sa ? ray_str_len(sa) : 0;
                }
            }
            result = ray_str_vec_append(result, sp, sl);
            if (RAY_IS_ERR(result)) break;
        }
    } else {
        /* Fixed-width branches (F64/I64/I32/I16/BOOL/U8/SYM/temporal): one
         * shared elementwise fill, dispatched across the worker pool for
         * large inputs — this op previously ran serial-only, capping every
         * `if`-projection at single-core speed regardless of -c. */
        if_fill_ctx_t fc = {
            .cond = cond_p, .then_v = then_v, .else_v = else_v,
            .then_scalar = then_scalar, .else_scalar = else_scalar,
            .t_f64 = 0.0, .e_f64 = 0.0, .t_i64 = 0, .e_i64 = 0,
            .dst = ray_data(result), .out_type = out_type,
        };
        /* Pre-resolve scalar broadcasts once (serial — SYM scalars may
         * intern, which must never happen inside a pool worker). */
        if (out_type == RAY_SYM) {
            if (then_scalar)
                fc.t_i64 = (then_v->type == -RAY_STR)
                    ? ray_sym_intern(ray_str_ptr(then_v), ray_str_len(then_v))
                    : sym_scalar_runtime_id(then_v);
            if (else_scalar)
                fc.e_i64 = (else_v->type == -RAY_STR)
                    ? ray_sym_intern(ray_str_ptr(else_v), ray_str_len(else_v))
                    : sym_scalar_runtime_id(else_v);
        } else {
            if (then_scalar) { fc.t_f64 = if_scalar_f64(then_v); fc.t_i64 = if_scalar_i64(then_v); }
            if (else_scalar) { fc.e_f64 = if_scalar_f64(else_v); fc.e_i64 = if_scalar_i64(else_v); }
        }

        ray_pool_t* pool = ray_pool_get();
        bool par = ray_pool_par_dispatch_ok(pool, len, RAY_PARALLEL_THRESHOLD);
        if_sym_side_t t_side, e_side;
        bool sides_ok = true;
        if (out_type == RAY_SYM) {
            /* Both inits run (each zeroes its side) so both frees are safe. */
            bool t_ok = if_sym_side_init(&t_side, then_scalar ? NULL : then_v);
            bool e_ok = if_sym_side_init(&e_side, else_scalar ? NULL : else_v);
            sides_ok = t_ok && e_ok;
            fc.t_side = &t_side;
            fc.e_side = &e_side;
        }
        if (!sides_ok) {
            if (out_type == RAY_SYM) { if_sym_side_free(&t_side); if_sym_side_free(&e_side); }
            ray_release(cond_v); ray_release(then_v); ray_release(else_v);
            ray_release(result);
            return ray_error("oom", NULL);
        }
        if (par && out_type == RAY_SYM) {
            /* Vector STR sides intern per element — serial only.  A SYM
             * column side is dispatch-safe once every position the fill
             * will read has a runtime id (sym.c frozen-table rule: no
             * interning inside a worker) — resolve those here, serially,
             * and only those; the column's whole vocabulary is never
             * interned. */
            ray_t* sides[2] = { then_scalar ? NULL : then_v,
                                else_scalar ? NULL : else_v };
            for (int s = 0; s < 2 && par; s++) {
                if (!sides[s]) continue;
                if (sides[s]->type != RAY_SYM) { par = false; break; }
            }
            if (par) {
                if_sym_side_prepare(&t_side, cond_p, len, 1);
                if_sym_side_prepare(&e_side, cond_p, len, 0);
                fc.sides_prepared = true;
            }
        }
        if (par)
            ray_pool_dispatch(pool, if_fill_par_fn, &fc, len);
        else
            if_fill_range(&fc, 0, len);
        if (out_type == RAY_SYM) { if_sym_side_free(&t_side); if_sym_side_free(&e_side); }
    }

    ray_release(cond_v); ray_release(then_v); ray_release(else_v);
    return result;
}

static bool if_type_is_text(int8_t t) { return t == RAY_STR || t == RAY_SYM; }

/* A branch that can stand in a text result: a text column or literal, a
 * boxed list, an untyped node, or a null literal such as `0N`, which the
 * fills write as the text null. */
static bool if_branch_text_ok(ray_graph_t* g, ray_op_t* b) {
    if (!b) return true;
    int8_t t = b->out_type;
    if (t <= 0 || t == RAY_LIST || if_type_is_text(t)) return true;
    if (b->opcode == OP_CONST) {
        ray_op_ext_t* ext = find_ext(g, b->id);
        if (ext && ext->literal && ray_is_atom(ext->literal) &&
            RAY_ATOM_IS_NULL(ext->literal))
            return true;
    }
    return false;
}

ray_t* exec_if(ray_graph_t* g, ray_op_t* op) {
    /* A text result with a non-text branch has no column type: the fills
     * read the numbers as string descriptors or symbol ids and returned
     * garbage, e.g. `(if (< x 3) s x)` with s SYM and x I64. */
    if (if_type_is_text(op->out_type)) {
        ray_op_t* then_op = op_child(g, op, 1);
        ray_op_ext_t* ext = find_ext(g, op->id);
        ray_op_t* else_op = ext ? op_node(g, ext->third_in) : NULL;
        int8_t tt = then_op ? then_op->out_type : 0;
        int8_t et = else_op ? else_op->out_type : 0;
        if (!if_branch_text_ok(g, then_op) || !if_branch_text_ok(g, else_op)) {
            g->if_refused = true;
            return ray_error("type", "if: branches mix %s and %s",
                             ray_type_name(tt), ray_type_name(et));
        }
    }
    ray_t* cond_v = exec_node(g, op_child(g, op, 0));
    if (!cond_v || RAY_IS_ERR(cond_v)) return cond_v;

    ray_t* selected = exec_if_selected(g, op, cond_v);
    if (selected) {
        ray_release(cond_v);
        return selected;
    }

    /* the condition is handed on, not evaluated twice */
    return exec_if_eager(g, op, cond_v);
}

/* Fold the null-mask words [nullw, nullw+null_words) into a running hash.
 * Mirrors group.c's ght_hash_null_words (file-local there, so pivot.c
 * carries its own copy): an all-zero word contributes nothing, so the
 * single-word (<=64-key) case is byte-identical to the historical
 * `if (idx_nmask) combine` this replaces. */
static inline uint64_t pivot_hash_null_words(uint64_t h, const int64_t* nullw,
                                             uint32_t null_words) {
    for (uint32_t w = 0; w < null_words; w++)
        if (nullw[w]) h = ray_hash_combine(h, ray_hash_i64(nullw[w]));
    return h;
}

/* ============================================================================
 * exec_pivot — single-pass hash-aggregated pivot table
 *
 * Groups by (index_cols, pivot_col), aggregates value_col, then unstacks
 * pivot values into separate output columns.
 * ============================================================================ */

/* Segment count of a parted table = length of any parted column's segs array
 * (mirrors build_segment_table exec.c). */
static int64_t pivot_parted_seg_count(ray_t* tbl) {
    int64_t nc = ray_table_ncols(tbl);
    for (int64_t c = 0; c < nc; c++) {
        ray_t* col = ray_table_get_col_idx(tbl, c);
        if (col && RAY_IS_PARTED(col->type)) return col->len;
    }
    return 0;
}

static bool pivot_table_has_parted_col(ray_t* tbl) {
    int64_t nc = ray_table_ncols(tbl);
    for (int64_t c = 0; c < nc; c++) {
        ray_t* col = ray_table_get_col_idx(tbl, c);
        if (col && (RAY_IS_PARTED(col->type) || col->type == RAY_MAPCOMMON))
            return true;
    }
    return false;
}

ray_t* exec_pivot(ray_graph_t* g, ray_op_t* op, ray_t* tbl) {
    if (!tbl || RAY_IS_ERR(tbl)) return tbl;

    /* Parted-input guard: exec_pivot fetches columns via ray_table_get_col and
     * indexes them by row up to ray_table_nrows.  A raw RAY_PARTED/MAPCOMMON
     * column (whose data is a tiny segs array, not nrows elements) would be
     * indexed out of bounds → OOB read / SIGSEGV.  Flatten (concat all
     * segments) and recompute on the flat table, mirroring exec_window /
     * join_with_parted_guard.  A flat table has no parted column, so the
     * recursive call falls straight through this guard (one level). */
    if (pivot_table_has_parted_col(tbl)) {
        int64_t nseg = pivot_parted_seg_count(tbl);
        ray_t* flat = NULL;
        for (int64_t s = 0; s < nseg; s++) {
            ray_t* seg = build_segment_table(tbl, (int32_t)s);
            if (!seg || RAY_IS_ERR(seg)) { if (flat) ray_release(flat); return seg ? seg : ray_error("oom", NULL); }
            if (!flat) flat = seg;
            else {
                ray_t* m = ray_result_merge(flat, seg);
                ray_release(flat); ray_release(seg);
                if (!m || RAY_IS_ERR(m)) return m ? m : ray_error("oom", NULL);
                flat = m;
            }
        }
        ray_t* r = exec_pivot(g, op, flat);
        ray_release(flat);
        return r;
    }

    ray_op_ext_t* ext = find_ext(g, op->id);
    if (!ext) return ray_error("nyi", NULL);

    uint32_t n_idx  = ext->pivot.n_index;
    uint16_t agg_op = ext->pivot.agg_op;
    int64_t nrows   = ray_table_nrows(tbl);

    /* n_idx is USER-DATA-sized (the index-arg vector's length, tblop.c's
     * pivot_fn_impl), not AST-bounded — task 6 lifts both this function's
     * own former n_keys > 8 gate and tblop.c's caller-side check (the ght
     * layout below spills to a heap block for n_keys > 8, same mechanism
     * group.c's exec_group_run already relies on), so a caller can drive
     * n_idx into the millions.  A stack VLA at that size is a stack-clash
     * SIGSEGV waiting to happen; heap-carve instead.  One consolidated
     * block holds all six per-index/per-key arrays below (idx_vecs,
     * key_data, key_vecs — pointer-sized, placed first so every array
     * after the first stays 8-byte aligned — then idx_wide, key_types,
     * key_attrs, the byte-sized ones); freed on every exit from here to
     * pivot_cleanup below.  n_idx_cap floors n_idx at 1 (an index-less
     * pivot is legal — grouping by the pivot column alone), n_keys = n_idx
     * + 1 is already >= 1. */
    uint32_t n_idx_cap = n_idx ? n_idx : 1;
    uint32_t n_keys = n_idx + 1;
    size_t key_ptr_bytes  = (size_t)n_idx_cap * sizeof(ray_t*)   /* idx_vecs */
                          + (size_t)n_keys   * sizeof(void*)     /* key_data */
                          + (size_t)n_keys   * sizeof(ray_t*);   /* key_vecs */
    size_t key_byte_bytes = (size_t)n_idx_cap * sizeof(bool)     /* idx_wide */
                          + (size_t)n_keys   * sizeof(int8_t)    /* key_types */
                          + (size_t)n_keys   * sizeof(uint8_t);  /* key_attrs */
    ray_t* key_hdr = NULL;
    uint8_t* key_base = (uint8_t*)scratch_alloc(&key_hdr, key_ptr_bytes + key_byte_bytes);
    if (!key_base) return ray_error("oom", NULL);
    ray_t**  idx_vecs  = (ray_t**)key_base;
    void**   key_data  = (void**)(idx_vecs + n_idx_cap);
    ray_t**  key_vecs  = (ray_t**)(key_data + n_keys);
    bool*    idx_wide  = (bool*)(key_vecs + n_keys);
    int8_t*  key_types = (int8_t*)(idx_wide + n_idx_cap);
    uint8_t* key_attrs = (uint8_t*)(key_types + n_keys);
    memset(idx_wide, 0, (size_t)n_idx_cap * sizeof(bool));

    for (uint32_t i = 0; i < n_idx; i++) {
        ray_op_ext_t* ie = find_ext(g, ext->pivot.index_cols[i]);
        idx_vecs[i] = (ie && ie->base.opcode == OP_SCAN)
                     ? ray_table_get_col(tbl, ie->sym) : NULL;
        if (!idx_vecs[i]) { scratch_free(key_hdr); return ray_error("domain", "pivot: index column not found"); }
    }

    ray_op_ext_t* pe = find_ext(g, ext->pivot.pivot_col);
    ray_t* pcol = (pe && pe->base.opcode == OP_SCAN)
                ? ray_table_get_col(tbl, pe->sym) : NULL;
    if (!pcol) { scratch_free(key_hdr); return ray_error("domain", "pivot: pivot column not found"); }

    ray_op_ext_t* ve = find_ext(g, ext->pivot.value_col);
    ray_t* vcol = (ve && ve->base.opcode == OP_SCAN)
                ? ray_table_get_col(tbl, ve->sym) : NULL;
    if (!vcol) { scratch_free(key_hdr); return ray_error("domain", "pivot: value column not found"); }

    if (nrows == 0) { scratch_free(key_hdr); return ray_table_new(0); }

    /* Combined keys: index_cols + pivot_col.  No count cap — the ght
     * layout below (ght_compute_layout) spills to an owned heap block
     * whenever n_keys exceeds its GHT_INLINE (8) inline capacity; its own
     * key-stride budget (uint16 byte offsets) is the only remaining limit,
     * unreachable at any index-column count a real table could carry. */
    for (uint32_t k = 0; k < n_idx; k++)
        idx_wide[k] = (idx_vecs[k]->type == RAY_GUID);
    bool pvt_wide = (pcol->type == RAY_GUID);

    for (uint32_t k = 0; k < n_idx; k++) {
        key_data[k]  = ray_data(idx_vecs[k]);
        key_types[k] = idx_vecs[k]->type;
        key_attrs[k] = idx_vecs[k]->attrs;
        key_vecs[k]  = idx_vecs[k];
    }
    key_data[n_idx]  = ray_data(pcol);
    key_types[n_idx] = pcol->type;
    key_attrs[n_idx] = pcol->attrs;
    key_vecs[n_idx]  = pcol;

    /* Single agg input: value column */
    ray_t* agg_vecs[1] = { vcol };
    uint16_t agg_ops[1] = { agg_op };

    /* Compute need_flags for the agg op */
    uint8_t need_flags = GHT_NEED_SUM; /* always need sum (used for FIRST/LAST too) */
    if (agg_op == OP_MIN) need_flags |= GHT_NEED_MIN;
    if (agg_op == OP_MAX) need_flags |= GHT_NEED_MAX;
    /* Integer avg divides the exact 128-bit sum (high words in off_sum_hi),
     * like every group engine — never a wrapped int64. */
    if (agg_op == OP_AVG && (vcol->type == RAY_I64 || vcol->type == RAY_TIMESTAMP ||
                             (vcol->type != RAY_F64 && vcol->type != RAY_F32 &&
                              nrows >= ((int64_t)1 << 31))))
        need_flags |= GHT_NEED_SUM128;

    /* n_keys/n_aggs are no longer capped: ght_compute_layout spills to an
     * owned heap block (ly.spill_hdr) whenever n_keys exceeds GHT_INLINE
     * (8) — the same path group.c's exec_group_run already exercises.
     * ght_layout_free(&ly) is therefore no longer always a no-op past this
     * point; every exit below (early return AND pivot_cleanup) must call
     * it exactly once so a spilled layout's heap block is never leaked. */
    ght_layout_t ly;
    /* key_vecs = NULL: pivot must NOT get the null-mask elision.  Its ingest
     * reads the key region's null word RAW (`keys[n_keys + pvt_null_word]`
     * below) to detect a null pivot key, so the layout needs null_words >= 1.
     * Passing NULL is what keeps it. */
    if (!ght_compute_layout(&ly, n_keys, 1, agg_vecs, NULL,
                            need_flags, agg_ops, key_types, NULL)) {
        scratch_free(key_hdr);
        return ray_error("limit", "pivot: key stride budget exceeded");
    }

    /* Hash-aggregate all rows via the shared radix pipeline — parallel
     * across thread-pool workers for n_scan ≥ RAY_PARALLEL_THRESHOLD,
     * sequential single-HT for smaller inputs. */
    ray_progress_update("pivot", "hash-aggregate", 0, (uint64_t)nrows);
    pivot_ingest_t pg;
    if (!pivot_ingest_run(&pg, &ly, key_data, key_types, key_attrs,
                          key_vecs, agg_vecs, nrows)) {
        pivot_ingest_free(&pg);
        ght_layout_free(&ly);
        scratch_free(key_hdr);
        return ray_error("oom", NULL);
    }
    ray_progress_update("pivot", "dedupe", 0, (uint64_t)pg.total_grps);
    if (ray_interrupted()) {
        pivot_ingest_free(&pg);
        ght_layout_free(&ly);
        scratch_free(key_hdr);
        return ray_error("cancel", "interrupted");
    }
    uint32_t grp_count = pg.total_grps;
    if (grp_count == 0) {
        pivot_ingest_free(&pg);
        ght_layout_free(&ly);
        scratch_free(key_hdr);
        return ray_table_new(0);
    }

    /* Pass 2: Collect distinct pivot values and distinct index keys.
     * Each group row layout: [hash:8][key0:8]...[keyN-1:8][null_word0:8]
     * ...[null_wordM-1:8][accum...] where the keys region holds n_idx
     * index keys + 1 pivot key, followed by the ceil(n_keys/64) key-null
     * mask words written by group_rows_range (ly.null_words; Task 3). */

    /* SQL PIVOT treats a null pivot key as "no column" — drop those groups.
     * The pivot key is the LAST of the n_keys keys, at bit position n_idx —
     * word-indexed per the Task-3/4 convention (word k>>6, bit k&63; the
     * unsigned-shift idiom sidesteps UB at bit 63) since n_idx can now
     * exceed 63 (pivot's own admission gate lifts along with the ght-path
     * key gate this migration also lifts — the two are no longer
     * independent facts once BOTH caps are gone). */
    const uint32_t null_words     = ly.null_words;
    const uint32_t pvt_null_word  = n_idx >> 6;
    const int64_t  pvt_null_bit   = (int64_t)((uint64_t)1 << (n_idx & 63));

    /* Radix partitioning is an execution detail: positions are handed out
     * in the order a value's or index tuple's earliest source row appears,
     * so the pivot columns and the index rows come out in first-seen order
     * whatever the partition or task-claim order was (group-by orders its
     * output the same way).  Both walks below stay linear over the
     * partitions; the earliest row per position is tracked as they go and
     * the positions are renumbered once at the end. */
    ray_t* pv_first_hdr = NULL;
    uint32_t pv_first_cap = 64;
    int64_t* pv_first = (int64_t*)scratch_alloc(&pv_first_hdr, pv_first_cap * sizeof(int64_t));
    if (!pv_first) { pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr); return ray_error("oom", NULL); }
    /* Collect distinct pivot values */
    uint32_t pv_cap = 64, pv_count = 0;
    ray_t* pv_hdr = NULL;
    int64_t* pv_vals = (int64_t*)scratch_alloc(&pv_hdr, pv_cap * sizeof(int64_t));
    if (!pv_vals) { scratch_free(pv_first_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr); return ray_error("oom", NULL); }

    const char* pvt_base = pvt_wide ? (const char*)key_data[n_idx] : NULL;
    for (uint32_t _p = 0; _p < pg.n_parts; _p++) {
        group_ht_t* ph = &pg.part_hts[_p];
        uint32_t pcount = ph->grp_count;
        for (uint32_t gi_local = 0; gi_local < pcount; gi_local++) {
            const char* row = ph->rows + (size_t)gi_local * pg.row_stride;
            const int64_t* rkeys = (const int64_t*)(row + 8);
            if (rkeys[n_keys + pvt_null_word] & pvt_null_bit) continue;
            int64_t pval = rkeys[n_idx];
            int64_t gfirst;
            memcpy(&gfirst, row + ly.off_group_first, 8);
            bool found = false;
            for (uint32_t p = 0; p < pv_count; p++) {
                if (pvt_wide) {
                    if (memcmp(pvt_base + (size_t)pv_vals[p] * 16,
                               pvt_base + (size_t)pval * 16, 16) == 0) { found = true; if (gfirst < pv_first[p]) pv_first[p] = gfirst; break; }
                } else {
                    if (pv_vals[p] == pval) { found = true; if (gfirst < pv_first[p]) pv_first[p] = gfirst; break; }
                }
            }
            if (!found) {
                if (pv_count >= pv_cap) {
                    uint32_t new_cap = pv_cap * 2;
                    int64_t* new_pv = (int64_t*)scratch_realloc(&pv_hdr,
                        pv_cap * sizeof(int64_t), new_cap * sizeof(int64_t));
                    int64_t* new_pf = (int64_t*)scratch_realloc(&pv_first_hdr,
                        pv_first_cap * sizeof(int64_t), new_cap * sizeof(int64_t));
                    if (!new_pv || !new_pf) { scratch_free(pv_first_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr); return ray_error("oom", NULL); }
                    pv_vals = new_pv; pv_first = new_pf;
                    pv_cap = new_cap; pv_first_cap = new_cap;
                }
                pv_first[pv_count] = gfirst;
                pv_vals[pv_count++] = pval;
            }
        }
    }

    /* Collect distinct index keys.
     * Flat append-only entry array + secondary open-addressed HT keyed by
     * the hash of (idx_keys + idx_null_words). The HT makes phase2 dedupe
     * O(grp_count) instead of the previous O(grp_count * ix_count)
     * linear scan which hung on large pivots.
     * Entry layout: [hash:8 | idx_keys:8*n_idx | idx_null_words:8*null_words].
     * Widened from a single trailing null word (Task-3 carryover finding):
     * a single word only ever compared/hashed/stored bits 0..63, silently
     * conflating two index tuples that agree on keys/nulls < 64 but differ
     * in a null bit >= 64 (index key 64+) into one dedupe entry — the
     * 65-index cell caught the clobber. */
    uint32_t ix_cap = 256, ix_count = 0;
    ray_t* ix_hdr = NULL;
    size_t ix_entry = 8 + (size_t)n_idx * 8 + (size_t)null_words * 8;
    char* ix_rows = (char*)scratch_alloc(&ix_hdr, ix_cap * ix_entry);
    if (!ix_rows) { scratch_free(pv_first_hdr); scratch_free(pv_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr); return ray_error("oom", NULL); }
    ray_t* ix_first_hdr = NULL;
    int64_t* ix_first = (int64_t*)scratch_alloc(&ix_first_hdr, ix_cap * sizeof(int64_t));
    if (!ix_first) { scratch_free(ix_hdr); scratch_free(pv_first_hdr); scratch_free(pv_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr); return ray_error("oom", NULL); }

    /* Secondary HT: hash slot -> ix_row index; empty = UINT32_MAX. */
    uint32_t ix_ht_cap = 256;
    while (ix_ht_cap < (uint32_t)grp_count * 2 && ix_ht_cap < (1u << 30)) ix_ht_cap <<= 1;
    ray_t* ix_ht_hdr = NULL;
    uint32_t* ix_ht = (uint32_t*)scratch_alloc(&ix_ht_hdr, ix_ht_cap * sizeof(uint32_t));
    if (!ix_ht) {
        scratch_free(ix_hdr); scratch_free(pv_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly);
        scratch_free(key_hdr);
        scratch_free(ix_first_hdr); scratch_free(pv_first_hdr);
        return ray_error("oom", NULL);
    }
    memset(ix_ht, 0xFF, ix_ht_cap * sizeof(uint32_t));
    uint32_t ix_ht_mask = ix_ht_cap - 1;

    /* Map: group_id -> (ix_row, pv_idx) for result cell placement */
    ray_t* map_hdr = NULL;
    uint32_t* grp_ix  = (uint32_t*)scratch_alloc(&map_hdr, grp_count * 2 * sizeof(uint32_t));
    if (!grp_ix) {
        scratch_free(ix_ht_hdr); scratch_free(ix_hdr); scratch_free(pv_hdr);
        pivot_ingest_free(&pg); ght_layout_free(&ly);
        scratch_free(key_hdr);
        scratch_free(ix_first_hdr); scratch_free(pv_first_hdr);
        return ray_error("oom", NULL);
    }
    uint32_t* grp_pv = grp_ix + grp_count;

    for (uint32_t _p = 0; _p < pg.n_parts; _p++) {
        group_ht_t* ph = &pg.part_hts[_p];
        uint32_t pcount = ph->grp_count;
        uint32_t gi_base = pg.part_offsets[_p];
        /* Progress tick at each partition boundary — time-gated so
         * 256 small partitions do not spam the callback. */
        ray_progress_update(NULL, NULL, gi_base, (uint64_t)grp_count);
        for (uint32_t gi_local = 0; gi_local < pcount; gi_local++) {
            uint32_t gi = gi_base + gi_local;
            const char* row = ph->rows + (size_t)gi_local * pg.row_stride;
            const int64_t* keys = (const int64_t*)(row + 8);
            if (keys[n_keys + pvt_null_word] & pvt_null_bit) {
                grp_ix[gi] = UINT32_MAX;
                grp_pv[gi] = UINT32_MAX;
                continue;
            }
            int64_t gfirst;
            memcpy(&gfirst, row + ly.off_group_first, 8);
        /* Index-key null words: keys[n_keys .. n_keys+null_words) as-is.
         * The pivot key's own null bit (position n_idx, word pvt_null_word)
         * is guaranteed 0 here — the check just above already `continue`d
         * on it being set — so these words already carry exactly the
         * index-key null flags with no separate masking needed. */
        const int64_t* idx_nwords = keys + n_keys;

        /* Hash index keys only (exclude pivot key) + null words.
         * Wide keys (GUID) resolve actual bytes via key_data[k]. */
        uint64_t ih = 0;
        for (uint32_t k = 0; k < n_idx; k++) {
            uint64_t kh;
            if (idx_wide[k]) {
                const char* base = (const char*)key_data[k];
                kh = ray_hash_bytes(base + (size_t)keys[k] * 16, 16);
            } else if (key_types[k] == RAY_F64) {
                double kd;                          /* F64 key bits live in the i64 slot; */
                memcpy(&kd, &keys[k], sizeof kd);   /* memcpy bit-cast is aliasing-safe   */
                kh = ray_hash_f64(kd);
            } else {
                kh = ray_hash_i64(keys[k]);
            }
            ih = (k == 0) ? kh : ray_hash_combine(ih, kh);
        }
        ih = pivot_hash_null_words(ih, idx_nwords, null_words);

        /* Open-addressed HT probe. On match, reuse; else insert. */
        uint32_t ix_row = UINT32_MAX;
        uint32_t slot = (uint32_t)(ih & ix_ht_mask);
        for (;;) {
            uint32_t ent = ix_ht[slot];
            if (ent == UINT32_MAX) break; /* empty → insert below */
            const char* ix_entry_p = ix_rows + (size_t)ent * ix_entry;
            if (*(const uint64_t*)ix_entry_p == ih) {
                const int64_t* ekeys = (const int64_t*)(ix_entry_p + 8);
                bool eq = true;
                for (uint32_t k = 0; k < n_idx && eq; k++) {
                    if (idx_wide[k]) {
                        const char* base = (const char*)key_data[k];
                        eq = (memcmp(base + (size_t)ekeys[k] * 16,
                                      base + (size_t)keys[k] * 16, 16) == 0);
                    } else {
                        eq = (ekeys[k] == keys[k]);
                    }
                }
                if (eq) {
                    const int64_t* ent_nwords = (const int64_t*)(ix_entry_p + 8 + (size_t)n_idx * 8);
                    eq = (memcmp(ent_nwords, idx_nwords, (size_t)null_words * 8) == 0);
                }
                if (eq) { ix_row = ent; if (gfirst < ix_first[ent]) ix_first[ent] = gfirst; break; }
            }
            slot = (slot + 1) & ix_ht_mask;
        }
        if (ix_row == UINT32_MAX) {
            if (ix_count >= ix_cap) {
                uint32_t new_cap = ix_cap * 2;
                char* new_rows = (char*)scratch_realloc(&ix_hdr,
                    ix_cap * ix_entry, new_cap * ix_entry);
                if (!new_rows) {
                    scratch_free(map_hdr); scratch_free(ix_ht_hdr);
                    scratch_free(pv_hdr); pivot_ingest_free(&pg); ght_layout_free(&ly);
                    scratch_free(key_hdr);
                    scratch_free(ix_first_hdr); scratch_free(pv_first_hdr);
                    return ray_error("oom", NULL);
                }
                int64_t* new_first = (int64_t*)scratch_realloc(&ix_first_hdr,
                    ix_cap * sizeof(int64_t), new_cap * sizeof(int64_t));
                if (!new_first) {
                    scratch_free(map_hdr); scratch_free(ix_ht_hdr); scratch_free(ix_first_hdr);
                    scratch_free(ix_hdr); scratch_free(pv_first_hdr); scratch_free(pv_hdr);
                    pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr);
                    return ray_error("oom", NULL);
                }
                ix_first = new_first;
                ix_rows = new_rows;
                ix_cap = new_cap;
            }
            ix_row = ix_count++;
            ix_first[ix_row] = gfirst;
            char* dst = ix_rows + (size_t)ix_row * ix_entry;
            *(uint64_t*)dst = ih;
            memcpy(dst + 8, keys, (size_t)n_idx * 8);
            memcpy(dst + 8 + (size_t)n_idx * 8, idx_nwords, (size_t)null_words * 8);
            ix_ht[slot] = ix_row;
        }

        /* Find pivot column index. For wide pivot keys both slot values
         * are source row indices — resolve to actual bytes for compare,
         * otherwise duplicate GUID pivot values map to the wrong column. */
        int64_t pval = keys[n_idx];
        uint32_t pv_idx = UINT32_MAX;
        for (uint32_t p = 0; p < pv_count; p++) {
            if (pvt_wide) {
                if (memcmp(pvt_base + (size_t)pv_vals[p] * 16,
                           pvt_base + (size_t)pval * 16, 16) == 0) { pv_idx = p; break; }
            } else {
                if (pv_vals[p] == pval) { pv_idx = p; break; }
            }
        }

            grp_ix[gi] = ix_row;
            grp_pv[gi] = pv_idx;
        }
    }

    /* Pass 3: Build output table */
    /* Renumber: rank pivot values and index rows by their earliest source
     * row, permute pv_vals / ix_rows into rank order, and point every group
     * at its new positions.  A linear scan over the row domain ranks the
     * index rows when there are many of them; a merge on ids does otherwise. */
    {
        /* 8-byte array first, the 4-byte ranks after it, so every array
         * starts aligned whatever the counts are. */
        ray_t* rank_hdr = NULL;
        int64_t* pv_tmp = (int64_t*)scratch_alloc(&rank_hdr,
            ((size_t)pv_count + 1) * sizeof(int64_t) + ((size_t)pv_count + (size_t)ix_count + 1) * sizeof(uint32_t));
        uint32_t* pv_rank = pv_tmp ? (uint32_t*)(pv_tmp + pv_count + 1) : NULL;
        if (!pv_rank) {
            scratch_free(map_hdr); scratch_free(ix_ht_hdr); scratch_free(ix_first_hdr);
            scratch_free(ix_hdr); scratch_free(pv_first_hdr); scratch_free(pv_hdr);
            pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr);
            return ray_error("oom", NULL);
        }
        uint32_t* ix_rank = pv_rank + pv_count;
        /* pivot values: few — insertion sort of old positions by first row */
        for (uint32_t p = 0; p < pv_count; p++) pv_rank[p] = p;
        for (uint32_t i = 1; i < pv_count; i++) {
            uint32_t v = pv_rank[i]; uint32_t j = i;
            while (j > 0 && pv_first[pv_rank[j - 1]] > pv_first[v]) { pv_rank[j] = pv_rank[j - 1]; j--; }
            pv_rank[j] = v;
        }
        for (uint32_t r = 0; r < pv_count; r++) pv_tmp[r] = pv_vals[pv_rank[r]];
        memcpy(pv_vals, pv_tmp, (size_t)pv_count * sizeof(int64_t));
        for (uint32_t r = 0; r < pv_count; r++) pv_tmp[pv_rank[r]] = r;   /* invert: old -> new */
        for (uint32_t p = 0; p < pv_count; p++) pv_rank[p] = (uint32_t)pv_tmp[p];
        /* index rows */
        bool ranked = false;
        if ((uint64_t)ix_count * 16 >= (uint64_t)nrows && nrows > 0) {
            ray_t* pos_hdr = NULL;
            uint32_t* pos = (uint32_t*)scratch_alloc(&pos_hdr, (size_t)nrows * sizeof(uint32_t));
            if (pos) {
                memset(pos, 0xFF, (size_t)nrows * sizeof(uint32_t));
                bool ok = true;
                for (uint32_t i = 0; i < ix_count && ok; i++) {
                    int64_t f = ix_first[i];
                    if (f < 0 || f >= nrows || pos[f] != UINT32_MAX) ok = false; else pos[f] = i;
                }
                if (ok) {
                    uint32_t next = 0;
                    for (int64_t r = 0; r < nrows; r++)
                        if (pos[r] != UINT32_MAX) ix_rank[pos[r]] = next++;
                    ranked = (next == ix_count);
                }
                scratch_free(pos_hdr);
            }
        }
        if (!ranked && ix_count > 0) {
            ray_t* tmp_hdr = NULL;
            uint32_t* buf = (uint32_t*)scratch_alloc(&tmp_hdr, 2 * (size_t)ix_count * sizeof(uint32_t));
            if (!buf) {
                scratch_free(rank_hdr); scratch_free(map_hdr); scratch_free(ix_ht_hdr); scratch_free(ix_first_hdr);
                scratch_free(ix_hdr); scratch_free(pv_first_hdr); scratch_free(pv_hdr);
                pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr);
                return ray_error("oom", NULL);
            }
            uint32_t* src = buf; uint32_t* dst = buf + ix_count;
            for (uint32_t i = 0; i < ix_count; i++) src[i] = i;
            for (uint32_t w = 1; w < ix_count; w *= 2) {
                for (uint32_t lo = 0; lo < ix_count; lo += 2 * w) {
                    uint32_t mid = lo + w < ix_count ? lo + w : ix_count;
                    uint32_t hi  = lo + 2 * w < ix_count ? lo + 2 * w : ix_count;
                    uint32_t a = lo, b = mid, o = lo;
                    while (a < mid && b < hi) dst[o++] = (ix_first[src[b]] < ix_first[src[a]]) ? src[b++] : src[a++];
                    while (a < mid) dst[o++] = src[a++];
                    while (b < hi)  dst[o++] = src[b++];
                }
                uint32_t* t = src; src = dst; dst = t;
            }
            for (uint32_t r = 0; r < ix_count; r++) ix_rank[src[r]] = r;   /* old -> new */
            scratch_free(tmp_hdr);
        }
        if (ix_count > 0) {
            ray_t* perm_hdr = NULL;
            char* perm = (char*)scratch_alloc(&perm_hdr, (size_t)ix_count * ix_entry);
            if (!perm) {
                scratch_free(rank_hdr); scratch_free(map_hdr); scratch_free(ix_ht_hdr); scratch_free(ix_first_hdr);
                scratch_free(ix_hdr); scratch_free(pv_first_hdr); scratch_free(pv_hdr);
                pivot_ingest_free(&pg); ght_layout_free(&ly); scratch_free(key_hdr);
                return ray_error("oom", NULL);
            }
            for (uint32_t i = 0; i < ix_count; i++)
                memcpy(perm + (size_t)ix_rank[i] * ix_entry, ix_rows + (size_t)i * ix_entry, ix_entry);
            memcpy(ix_rows, perm, (size_t)ix_count * ix_entry);
            scratch_free(perm_hdr);
        }
        for (uint32_t gi = 0; gi < grp_count; gi++) {
            if (grp_ix[gi] != UINT32_MAX) grp_ix[gi] = ix_rank[grp_ix[gi]];
            if (grp_pv[gi] != UINT32_MAX) grp_pv[gi] = pv_rank[grp_pv[gi]];
        }
        scratch_free(rank_hdr);
    }
    scratch_free(ix_first_hdr);
    scratch_free(pv_first_hdr);
    ray_progress_update("pivot", "scatter", 0, (uint64_t)pv_count);
    bool val_is_f64 = vcol->type == RAY_F64;
    int8_t out_agg_type;
    switch (agg_op) {
        case OP_AVG:   out_agg_type = RAY_F64; break;
        case OP_COUNT: out_agg_type = RAY_I64; break;
        case OP_SUM:   out_agg_type = val_is_f64 ? RAY_F64 : RAY_I64; break;
        default:       out_agg_type = vcol->type; break;
    }

    int64_t out_ncols = (int64_t)n_idx + (int64_t)pv_count;
    ray_t* result = ray_table_new(out_ncols);
    if (!result || RAY_IS_ERR(result)) goto pivot_cleanup;

    /* Index columns */
    for (uint32_t k = 0; k < n_idx; k++) {
        ray_t* new_col = col_vec_new(idx_vecs[k], (int64_t)ix_count);
        if (!new_col || RAY_IS_ERR(new_col)) { ray_release(result); result = ray_error("oom", NULL); goto pivot_cleanup; }
        new_col->len = (int64_t)ix_count;
        uint8_t esz = col_esz(idx_vecs[k]);
        int8_t kt = idx_vecs[k]->type;
        const char* src_base = idx_wide[k] ? (const char*)key_data[k] : NULL;
        /* Word-indexed per the Task-3/4 convention: k is fixed in this outer
         * scope, so hoist its word/bit out of the per-row loop below (as
         * the seq emit loop in group.c does) rather than recomputing it
         * ix_count times. */
        const uint32_t null_kw  = k >> 6;
        const int64_t  null_kbit = (int64_t)((uint64_t)1 << (k & 63));
        for (uint32_t r = 0; r < ix_count; r++) {
            const char* ix_entry_p = ix_rows + r * ix_entry;
            int64_t kv = ((const int64_t*)(ix_entry_p + 8))[k];
            const int64_t* ent_nwords = (const int64_t*)(ix_entry_p + 8 + (size_t)n_idx * 8);
            if (ent_nwords[null_kw] & null_kbit) {
                ray_vec_set_null(new_col, (int64_t)r, true);
                /* Fill the correct-width sentinel. */
                switch (kt) {
                    case RAY_F64:
                        ((double*)ray_data(new_col))[r] = NULL_F64; break;
                    case RAY_I64: case RAY_TIMESTAMP:
                        ((int64_t*)ray_data(new_col))[r] = NULL_I64; break;
                    case RAY_I32: case RAY_DATE: case RAY_TIME:
                        ((int32_t*)ray_data(new_col))[r] = NULL_I32; break;
                    case RAY_I16:
                        ((int16_t*)ray_data(new_col))[r] = NULL_I16; break;
                    default: break;
                }
                continue;
            }
            if (idx_wide[k]) {
                /* kv is a source row index; copy the 16 raw bytes. */
                memcpy((char*)ray_data(new_col) + (size_t)r * esz,
                       src_base + (size_t)kv * 16, 16);
            } else if (kt == RAY_F64) {
                memcpy((char*)ray_data(new_col) + (size_t)r * esz, &kv, 8);
            } else {
                write_col_i64(ray_data(new_col), (int64_t)r, kv, kt, new_col->attrs);
            }
        }
        if (idx_vecs[k]->type == RAY_STR)
            col_propagate_str_pool(new_col, idx_vecs[k]);
        /* Output-vec rule (Task-2 review): a SYM index column raw-copies
         * cell ids from its source, so it must resolve over the source's
         * dictionary.  No-op while every domain is the runtime singleton. */
        if (new_col->type == RAY_SYM)
            ray_sym_vec_adopt_domain(new_col, idx_vecs[k]);

        ray_op_ext_t* ie = find_ext(g, ext->pivot.index_cols[k]);
        result = ray_table_add_col(result, ie->sym, new_col);
        ray_release(new_col);
        if (RAY_IS_ERR(result)) goto pivot_cleanup;
    }

    /* Value columns — one per distinct pivot value */
    {
    int8_t s = ly.agg_val_slot[0]; /* single agg input -> slot 0 */
    for (uint32_t p = 0; p < pv_count; p++) {
        ray_t* new_col = (out_agg_type == vcol->type)
                        ? col_vec_new(vcol, (int64_t)ix_count)
                        : ray_vec_new(out_agg_type, (int64_t)ix_count);
        if (!new_col || RAY_IS_ERR(new_col)) { ray_release(result); result = ray_error("oom", NULL); goto pivot_cleanup; }
        new_col->len = (int64_t)ix_count;

        /* Initialize missing cells with the type-correct NULL sentinel —
         * a pivot cell with no source row is "no data", which must stay
         * distinguishable from a real 0 (review 2.10).  Present cells get
         * overwritten in the scatter loop below; whatever remains is null.
         * par_finalize_nulls (after scatter) flips HAS_NULLS if any
         * sentinel survives.  Types without a dedicated initialization arm
         * (SYM/STR/BOOL/U8/GUID) fall back to zero-fill.  For SYM/STR/BOOL/U8
         * that is an ordinary zero/empty value and cannot represent "no
         * data"; GUID is also zero-filled on this path. */
        switch (new_col->type) {
            case RAY_F64: {
                double* d = (double*)ray_data(new_col);
                for (int64_t r = 0; r < (int64_t)ix_count; r++) d[r] = NULL_F64;
                break;
            }
            case RAY_F32: {
                float* d = (float*)ray_data(new_col);
                for (int64_t r = 0; r < (int64_t)ix_count; r++) d[r] = NULL_F32;
                break;
            }
            case RAY_I64: case RAY_TIMESTAMP: {
                int64_t* d = (int64_t*)ray_data(new_col);
                for (int64_t r = 0; r < (int64_t)ix_count; r++) d[r] = NULL_I64;
                break;
            }
            case RAY_I32: case RAY_DATE: case RAY_TIME: {
                int32_t* d = (int32_t*)ray_data(new_col);
                for (int64_t r = 0; r < (int64_t)ix_count; r++) d[r] = NULL_I32;
                break;
            }
            case RAY_I16: {
                int16_t* d = (int16_t*)ray_data(new_col);
                for (int64_t r = 0; r < (int64_t)ix_count; r++) d[r] = NULL_I16;
                break;
            }
            default:
                memset(ray_data(new_col), 0, (size_t)ix_count * (size_t)col_esz(new_col));
                break;
        }

        for (uint32_t _pp = 0; _pp < pg.n_parts; _pp++) {
            group_ht_t* ph = &pg.part_hts[_pp];
            uint32_t pcount = ph->grp_count;
            uint32_t gi_base = pg.part_offsets[_pp];
            for (uint32_t gi_local = 0; gi_local < pcount; gi_local++) {
                uint32_t gi = gi_base + gi_local;
                if (grp_pv[gi] != p) continue;
                uint32_t r = grp_ix[gi];
                const char* row = ph->rows + (size_t)gi_local * pg.row_stride;
                int64_t cnt = *(const int64_t*)(const void*)row;
                /* nn = per-slot non-null count (nullable value column) or the
                 * group row count (null-free — byte-identical to before). */
                /* s < 0 for an agg that reserves no value slot (holistic, or
                 * the now-valueless OP_COUNT): it owns no off_nn slot either,
                 * so never index with it — mirrors the guard in group.c's
                 * radix_phase3_fn / sequential emit.  Safe today only because
                 * off_nn is 0 for every such layout; the guard makes it not
                 * depend on that coincidence. */
                int64_t nn = (ly.off_nn && s >= 0)
                    ? ((const int64_t*)(const void*)(row + ly.off_nn))[s] : cnt;

            if (out_agg_type == RAY_F64) {
                double v;
                switch (agg_op) {
                    case OP_SUM:
                        v = val_is_f64 ? ROW_RD_F64(row, ly.off_sum, s)
                                       : (double)ROW_RD_I64(row, ly.off_sum, s);
                        break;
                    case OP_AVG:
                        if (nn == 0) { v = NULL_F64; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = val_is_f64 ? ROW_RD_F64(row, ly.off_sum, s) / nn
                          : (ly.need_flags & GHT_NEED_SUM128)
                            ? ray_i128_to_f64(ROW_RD_I64(row, ly.off_sum_hi, s),
                                              (uint64_t)ROW_RD_I64(row, ly.off_sum, s)) / nn
                            : (double)ROW_RD_I64(row, ly.off_sum, s) / nn;
                        break;
                    case OP_MIN:
                        if (nn == 0) { v = NULL_F64; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = val_is_f64 ? ROW_RD_F64(row, ly.off_min, s)
                                       : (double)ROW_RD_I64(row, ly.off_min, s);
                        break;
                    case OP_MAX:
                        if (nn == 0) { v = NULL_F64; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = val_is_f64 ? ROW_RD_F64(row, ly.off_max, s)
                                       : (double)ROW_RD_I64(row, ly.off_max, s);
                        break;
                    case OP_FIRST: case OP_LAST:
                        if (nn == 0) { v = NULL_F64; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = val_is_f64 ? ROW_RD_F64(row, ly.off_sum, s)
                                       : (double)ROW_RD_I64(row, ly.off_sum, s);
                        break;
                    default: v = 0.0; break;
                }
                /* Single-null float model: canonicalize non-finite (avg
                 * division, sum overflow) to NULL_F64; HAS_NULLS set by the
                 * scan below before new_col is added to the result. */
                ((double*)ray_data(new_col))[r] = ray_f64_fin(v);
            } else {
                int64_t v;
                int64_t int_null = agg_int_null_sentinel_for(out_agg_type);
                switch (agg_op) {
                    case OP_SUM:   v = ROW_RD_I64(row, ly.off_sum, s); break;
                    case OP_COUNT: v = cnt; break;
                    case OP_MIN:
                        if (nn == 0) { v = int_null; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = ROW_RD_I64(row, ly.off_min, s); break;
                    case OP_MAX:
                        if (nn == 0) { v = int_null; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = ROW_RD_I64(row, ly.off_max, s); break;
                    case OP_FIRST: case OP_LAST:
                        if (nn == 0) { v = int_null; ray_vec_set_null(new_col, (int64_t)r, true); break; }
                        v = ROW_RD_I64(row, ly.off_sum, s); break;
                    default:       v = 0; break;
                }
                    write_col_i64(ray_data(new_col), (int64_t)r, v, out_agg_type, new_col->attrs);
                }
            }
        }

        /* Output-vec rule (sym-domain Phase 2): a SYM value column
         * (MIN/MAX/FIRST/LAST over a SYM input) carries raw cell ids
         * accumulated from `vcol` — it must resolve over vcol's
         * dictionary.  The memset-0 fill for missing cells is the SYM
         * null (id 0) in any domain.  No-op while every domain is the
         * runtime singleton. */
        if (new_col->type == RAY_SYM)
            ray_sym_vec_adopt_domain(new_col, vcol);

        /* Column name from pivot value — match pivot_val_to_sym semantics */
        int64_t pval = pv_vals[p];
        int64_t col_sym;
        if (pcol->type == RAY_SYM) {
            /* pivot value (cell-data) becomes a column NAME — names live
             * in the runtime table, so re-express the id there. */
            col_sym = sym_id_runtime(pcol, pval);
        } else if (pvt_wide) {
            /* GUID: format 16 bytes as xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx.
             * pval is a source row index into pvt_base. */
            static const char hex[] = "0123456789abcdef";
            static const int groups[] = {4, 2, 2, 2, 6};
            char buf[37];
            const uint8_t* bytes = (const uint8_t*)pvt_base + (size_t)pval * 16;
            int pos = 0, bpos = 0;
            for (int g = 0; g < 5; g++) {
                if (g > 0) buf[bpos++] = '-';
                for (int j = 0; j < groups[g]; j++) {
                    buf[bpos++] = hex[bytes[pos] >> 4];
                    buf[bpos++] = hex[bytes[pos] & 0x0F];
                    pos++;
                }
            }
            col_sym = ray_sym_intern(buf, (size_t)bpos);
        } else {
            char buf[128];
            int len = 0;
            int8_t pt = key_types[n_idx];
            if (pt == RAY_F64) {
                double fv;
                memcpy(&fv, &pval, 8);
                fv = clear_neg_zero(fv);
                len = snprintf(buf, sizeof(buf), "%g", fv);
            } else if (pt == RAY_BOOL) {
                len = snprintf(buf, sizeof(buf), "%s", pval ? "true" : "false");
            } else if (pt == RAY_I64 || pt == RAY_I32 || pt == RAY_I16 ||
                       pt == RAY_DATE || pt == RAY_TIME || pt == RAY_TIMESTAMP) {
                len = snprintf(buf, sizeof(buf), "%" PRId64, pval);
            } else {
                len = snprintf(buf, sizeof(buf), "col%" PRId64, pval);
            }
            col_sym = ray_sym_intern(buf, (size_t)len);
        }

        /* Flip HAS_NULLS if any pivot cell carries the type-correct NULL
         * sentinel — either a missing cell (no source row) left as null by
         * the init above, or (F64) a value canonicalized to NULL_F64 by
         * ray_f64_fin (avg division, sum overflow).  par_finalize_nulls is
         * a no-op for types it does not handle (SYM/STR/BOOL/U8/GUID). */
        par_finalize_nulls(new_col);
        result = ray_table_add_col(result, col_sym, new_col);
        ray_release(new_col);
        if (RAY_IS_ERR(result)) goto pivot_cleanup;
    }
    }

pivot_cleanup:
    ght_layout_free(&ly);   /* no-op when inline (n_keys <= 8); frees the
                             * owned spill block otherwise (see above) */
    scratch_free(map_hdr);
    scratch_free(ix_ht_hdr);
    scratch_free(ix_hdr);
    scratch_free(pv_hdr);
    scratch_free(key_hdr); /* idx_vecs/idx_wide/key_data/key_types/key_attrs/
                             * key_vecs consolidated carve, alive since the
                             * top of this function (see key_hdr above) */
    pivot_ingest_free(&pg);
    return result;
}