libdpf/test/tests/compose_test.cpp

#include <gtest/gtest.h>

#include <chrono>
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <fstream>
#include <map>
#include <memory>
#include <optional>
#include <stdexcept>
#include <string>
#include <thread>
#include <tuple>
#include <vector>

#include "dpf/beaver.hpp"
#include "dpf/compose.hpp"
#include "dpf/iknp.hpp"
#include "dpf/app_flow.hpp"
#include "dpf/app_plans.hpp"
#include "dpf/experiment.hpp"
#include "dpf/iknp_graphs.hpp"
#include "dpf/mesh_apps.hpp"
#include "dpf/net/edge_mesh.hpp"
#include "dpf/net/memory_sink.hpp"
#include "dpf/net/sink_exchange.hpp"
#include "dpf/pad_graphs.hpp"
#include "dpf/prg_aes.hpp"
#include "dpf/random.hpp"
#include "dpf/session_host.hpp"

namespace
{

using dpf::protocol::composer;
using dpf::protocol::domain;
using dpf::protocol::effect;
using dpf::protocol::kernel_fn;
using dpf::protocol::node;
namespace opcodes = dpf::protocol::opcodes;
using dpf::protocol::op_flags;
using dpf::net::make_memory_sink_pair;

TEST(Compose, SharedBlindInterned)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    int blind_calls = 0;
    auto b0 = c.blind(x, 42);
    auto b1 = c.blind(x, 42);
    EXPECT_EQ(b0, b1);
    auto p = c.schedule();
    EXPECT_EQ(p.effect_count(effect::blind), 1u);

    std::map<std::uint32_t, kernel_fn> kernels;
    kernels[42] = [&](std::uint32_t, const std::vector<node> & nodes,
                       const std::vector<dpf::protocol::block_span> &,
                       dpf::protocol::block_span out, std::size_t) {
        ++blind_calls;
        for (std::size_t i = 0; i < out.lanes; ++i)
        {
            std::uint64_t v = 0x1111;
            std::memcpy(out.at(i), &v, 8);
        }
        EXPECT_EQ(nodes.size(), 1u);
    };
    auto [sink0, sink1] = make_memory_sink_pair(1, p.slot_bytes_all());
    (void)sink1;
    std::vector<std::vector<std::uint8_t>> values;
    values.resize(c.node_count());
    values[x.id].resize(8);
    std::uint64_t xv = 7;
    std::memcpy(values[x.id].data(), &xv, 8);
    dpf::protocol::drive(p, sink0, values, kernels, 0);
    EXPECT_EQ(blind_calls, 1);
    std::uint64_t got = 0;
    std::memcpy(&got, values[b0.id].data(), 8);
    EXPECT_EQ(got, 0x1111u);
}

TEST(Compose, LevelWalkSharesExpands)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto a = c.fss_point(seed, 2, 16);
    const auto shared = c.schedule().effect_count(effect::expand);
    // depth 2: one fused L||R expand per level (hand DPF PRG stretch).
    EXPECT_EQ(shared, 2u);
    auto b = c.fss_point(seed, 2, 16);
    EXPECT_EQ(c.schedule().effect_count(effect::expand), shared);
    auto other = c.input(domain::fss, 16);
    auto d = c.fss_cmp(other, 2, 16);
    EXPECT_EQ(c.schedule().effect_count(effect::expand), shared * 2);
    EXPECT_EQ(c.domain_of(a), domain::b);
    EXPECT_EQ(c.domain_of(b), domain::b);
    EXPECT_EQ(c.domain_of(d), domain::a);
}

TEST(Compose, ExchangeWaves)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    auto e0 = c.exchange(x);
    auto e1 = c.exchange(c.compute(opcodes::user_base + 10, {e0}, domain::a, 8));
    auto e2 = c.exchange(c.compute(opcodes::user_base + 10, {e1}, domain::a, 8));
    (void)e2;
    auto p = c.schedule();
    EXPECT_EQ(p.rounds(), 3u);

    composer c2(0);
    auto a = c2.input(domain::a, 8);
    auto b = c2.input(domain::a, 8);
    auto o0 = c2.exchange(a);
    auto o1 = c2.exchange(b);
    (void)o0;
    (void)o1;
    auto p2 = c2.schedule();
    EXPECT_EQ(p2.rounds(), 1u);
}

TEST(Compose, ExchangeReconstructsOnMemorySink)
{
    composer c0(0);
    composer c1(1);
    auto x0 = c0.input(domain::a, 8);
    auto x1 = c1.input(domain::a, 8);
    auto e0 = c0.exchange(x0);
    auto e1 = c1.exchange(x1);
    (void)e0;
    (void)e1;
    auto p0 = c0.schedule();
    auto p1 = c1.schedule();
    ASSERT_EQ(p0.rounds(), 1u);
    ASSERT_EQ(p1.rounds(), 1u);

    auto [sink0, sink1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::uint64_t a = 3, b = 5;
    sink0.submit(0, 0, reinterpret_cast<const std::uint8_t *>(&a), 8);
    sink1.submit(0, 0, reinterpret_cast<const std::uint8_t *>(&b), 8);
    sink0.flush();
    ASSERT_TRUE(sink0.peer_ready(0, 0));
    ASSERT_TRUE(sink1.peer_ready(0, 0));
    std::uint64_t peer0 = 0, peer1 = 0;
    sink0.read_peer(0, 0, reinterpret_cast<std::uint8_t *>(&peer0), 8);
    sink1.read_peer(0, 0, reinterpret_cast<std::uint8_t *>(&peer1), 8);
    EXPECT_EQ(a + peer0, 8u);
    EXPECT_EQ(b + peer1, 8u);
}

TEST(Compose, CommutativeComputeInterned)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    auto y = c.input(domain::a, 8);
    auto p0 = c.compute(99, {x, y}, domain::a, 8, op_flags::commutative);
    auto p1 = c.compute(99, {y, x}, domain::a, 8, op_flags::commutative);
    EXPECT_EQ(p0, p1);
}

TEST(Compose, SimdGroupOneKernelCall)
{
    composer c(0);
    auto nodes = c.fan(8, [&](std::size_t) {
        return c.input(domain::a, 8);
    });
    std::vector<node> outs;
    for (auto n : nodes)
        outs.push_back(c.compute(77, {n}, domain::a, 8));
    auto p = c.schedule();
    int calls = 0;
    std::size_t total_nodes = 0;
    std::map<std::uint32_t, kernel_fn> kernels;
    kernels[77] = [&](std::uint32_t, const std::vector<node> & ns,
                       const std::vector<dpf::protocol::block_span> & ins,
                       dpf::protocol::block_span out, std::size_t sink_lanes) {
        ++calls;
        total_nodes += ns.size();
        EXPECT_EQ(ns.size(), 8u);
        EXPECT_EQ(out.lanes, 8u * sink_lanes);
        EXPECT_EQ(ins.size(), 1u);
        for (std::size_t i = 0; i < out.lanes; ++i)
            std::memset(out.at(i), 1, 8);
    };
    auto slots = p.slot_bytes_all();
    if (slots.empty())
        slots.push_back(0);
    auto [sink0, sink1] = make_memory_sink_pair(1, std::move(slots));
    (void)sink1;
    std::vector<std::vector<std::uint8_t>> values(c.node_count());
    for (auto n : nodes)
        values[n.id].assign(8, 0);
    dpf::protocol::drive(p, sink0, values, kernels, 0);
    EXPECT_EQ(calls, 1);
    EXPECT_EQ(total_nodes, 8u);
    ASSERT_GE(p.waves(), 1u);
    std::size_t grouped = 0;
    for (const auto & g : p.wave(0).groups)
    {
        if (g.opcode == 77)
            grouped += g.nodes.size();
    }
    EXPECT_EQ(grouped, 8u);
    (void)outs;
}
TEST(Compose, DomainConversionInternedAndRejected)
{
    composer c(0);
    auto leaf = c.input(domain::b, 8);
    auto a0 = c.as(leaf, domain::a);
    auto a1 = c.as(leaf, domain::a);
    EXPECT_EQ(a0, a1);
    auto p = c.schedule();
    EXPECT_EQ(p.conversion_count(domain::b, domain::a), 1u);
    EXPECT_THROW(c.as(leaf, domain::rss), std::invalid_argument);

    // Inverse fold
    auto back = c.as(a0, domain::b);
    EXPECT_EQ(back, leaf);
}

TEST(Compose, RssProductInterned)
{
    composer c(0);
    auto x = c.input(domain::rss, 16);
    auto y = c.input(domain::rss, 16);
    auto f0 = c.rss_product(x, y);
    auto f1 = c.rss_product(y, x);  // commutative
    EXPECT_EQ(f0, f1);
    EXPECT_EQ(c.domain_of(f0), domain::y);

    auto own = c.input(domain::y, 8);
    auto next = c.input(domain::y, 8);
    auto r = c.y2rss(own, next);
    EXPECT_EQ(c.domain_of(r), domain::rss);
}

TEST(Compose, AbyProductSharesBlind)
{
    composer c(0);
    auto & s = c.aby<std::uint64_t>();
    auto x = s.input();
    auto y = s.input();
    auto z = s.input();
    auto xy = s.product(x, y);
    auto xz = s.product(x, z);
    (void)xy;
    (void)xz;
    // One blind for x across both products.
    s.sample();
    EXPECT_EQ(s.round_of(xy), s.round_of(xz));
    // opened wires create barrier exchanges on the composer
    auto nxy = c.opened<std::uint64_t>(xy);
    auto nxz = c.opened<std::uint64_t>(xz);
    EXPECT_EQ(c.domain_of(nxy), domain::a);
    EXPECT_EQ(c.domain_of(nxz), domain::a);
    auto barriers = s.exchange_barriers();
    EXPECT_FALSE(barriers.empty());
}

// Practical gap (SUBLEQ scale / Grotto η): opening a product must schedule
// every prior δ flush, not only the product's barrier.
TEST(Compose, BeaverBarrierChainMaterializesPriors)
{
    composer c(0);
    auto & s = c.aby<std::uint64_t>();
    auto x = s.input();
    auto y = s.input();
    auto z = s.product(x, y);
    // Nested use so z is δ-opened in a later barrier (hand SUBLEQ scale path).
    auto w = s.product(z, x);
    (void)w;
    s.sample();
    const auto barriers = s.exchange_barriers();
    ASSERT_GE(barriers.size(), 2u);
    auto nz = c.opened<std::uint64_t>(z);
    (void)nz;
    // Opening z must pull in every prior flush (factor open), not only z's.
    std::size_t beaver_ex = 0;
    for (std::uint32_t id = 0; id < c.node_count(); ++id)
    {
        if (c.effect_of(node{id}) != effect::exchange)
            continue;
        // Beaver barrier opcodes are beaver_delta + index.
        ++beaver_ex;
    }
    // FSS-free composer: every exchange is a beaver barrier.
    EXPECT_GE(beaver_ex, 2u);
}

// Practical gap (Duoram / PIR + Beaver): beaver waves must not hitch onto
// unrelated FSS exchange node ids (old fake inputs = {sess, barrier_index}).
TEST(Compose, BeaverBarrierWavesIndependentOfFssNodeIds)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto leaf = c.fss_point(seed, 3, 16);
    (void)leaf;
    const auto fss_only = c.schedule();
    // depth d → d exchange waves + final consumer wave.
    EXPECT_EQ(fss_only.effect_count(effect::exchange), 3u);
    EXPECT_EQ(fss_only.rounds(), 3u);  // == exchange_waves
    EXPECT_EQ(fss_only.exchange_waves(), 3u);

    auto & s = c.aby<std::uint64_t>();
    auto x = s.input();
    auto y = s.input();
    auto z = s.product(x, y);
    s.sample();
    auto nz = c.opened<std::uint64_t>(z);
    (void)nz;
    auto p = c.schedule();
    // Independent beaver opens add exchange nodes; they must not disappear
    // or be ordered solely by colliding with FSS node ids.
    EXPECT_GT(p.effect_count(effect::exchange), 3u);
}

// Practical gap (grotto_signum_lut): N comparison walks share expands when
// seeded alike; packed same-depth opens are one wave per level.
TEST(Compose, GrottoShapedFanSharesSeedExpands)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    constexpr std::size_t n = 7;
    auto leaves = c.fan(n, [&](std::size_t) {
        return c.fss_cmp(seed, 4, 16);
    });
    EXPECT_EQ(leaves.size(), n);
    auto p = c.schedule();
    // One fused expand per level for the shared seed, not n×depth.
    EXPECT_EQ(p.effect_count(effect::expand), 4u);
    EXPECT_EQ(p.effect_count(effect::exchange), 4u);
    EXPECT_EQ(p.rounds(), 4u);  // == exchange_waves (no empty sink round)
}

// Latency objective: keep sign×linear in one round (Pika / online Grotto).
// Prep objective (default) still peels for Appendix-E savings.
TEST(Compose, AbySessionUsesRoundAwareSchedule)
{
    composer c(0);
    auto & s = c.aby<std::uint64_t>();
    EXPECT_EQ(s.get_schedule_objective(),
        dpf::beavers::schedule_objective::rounds);
    auto sgn = s.input();
    auto x = s.input();
    auto a0 = s.input();
    auto a1 = s.input();
    auto lin = s(sgn * (a1 * x + a0));
    EXPECT_EQ(s.round_of(lin), 1);

    dpf::beavers::session<std::uint64_t> prep;
    prep.set_schedule_objective(dpf::beavers::schedule_objective::prep);
    auto ps = prep.input();
    auto px = prep.input();
    auto pa0 = prep.input();
    auto pa1 = prep.input();
    auto plin = prep(ps * (pa1 * px + pa0));
    EXPECT_EQ(prep.round_of(plin), 2);
    EXPECT_LT(prep.preprocessing_count(), s.preprocessing_count());
}

// Express/Sabre: sketch rides in the last CW flush — no extra exchange round.
TEST(Compose, ExpressShapedSketchFusedIntoLastCw)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto sketch = c.input(domain::a, 8);

    // Naive: walk, then a sketch exchange that depends on the leaf.
    auto leaf = c.fss_point(seed, 4, 16);
    auto sk_dep = c.compute(opcodes::user_base + 50, {leaf, sketch},
        domain::a, 8);
    auto sk_ex = c.exchange(sk_dep);
    (void)sk_ex;
    auto naive_plan = c.schedule();

    composer c2(0);
    auto seed2 = c2.input(domain::fss, 16);
    auto sketch2 = c2.input(domain::a, 8);
    auto wr = c2.fss_point_fused(seed2, 4, 16, sketch2);
    auto fused_plan = c2.schedule();
    EXPECT_EQ(c2.domain_of(wr.leaf), domain::b);
    EXPECT_EQ(c2.domain_of(wr.trailer_open), domain::a);
    EXPECT_EQ(naive_plan.effect_count(effect::exchange), 5u);
    EXPECT_EQ(fused_plan.effect_count(effect::exchange), 4u);
    // Same leaf-wave depth, but naive pays a fifth exchange-bearing wave.
    EXPECT_EQ(fused_plan.rounds(), 4u);
    EXPECT_EQ(naive_plan.exchange_waves(), 5u);
    EXPECT_EQ(fused_plan.exchange_waves(), 4u);
    // Last CW slot carries step‖trailer.
    const auto & last_wave = fused_plan.wave(3);
    ASSERT_FALSE(last_wave.exchanges.empty());
    EXPECT_EQ(fused_plan.value_bytes_of(last_wave.exchanges[0].id), 16u + 8u);
    EXPECT_TRUE(fused_plan.wave(4).exchanges.empty());
}

TEST(Compose, BeaverExchangeBarriersMatchBatch)
{
    dpf::beavers::session<std::uint64_t> s;
    auto x = s.input();
    auto y = s.input();
    auto z = s.product(x, y);
    (void)z;
    s.sample();
    auto barriers = s.exchange_barriers();
    EXPECT_GE(barriers.size(), 1u);

    // Party path: stepper completes with same number of barriers.
    dpf::beavers::session<std::uint64_t> s0;
    dpf::beavers::session<std::uint64_t> s1;
    auto x0 = s0.input();
    auto y0 = s0.input();
    auto z0 = s0.product(x0, y0);
    auto x1 = s1.input();
    auto y1 = s1.input();
    auto z1 = s1.product(x1, y1);
    (void)z0;
    (void)z1;
    s0.sample();
    // copy tape
    auto tape = s0.export_party(0);
    // Use install after sampling on dealer session - parties need tapes.
    // Simpler: evaluate_party_batch with matching exchange count.
    dpf::beavers::session<std::uint64_t> dealer;
    auto dx = dealer.input();
    auto dy = dealer.input();
    auto dz = dealer.product(dx, dy);
    (void)dz;
    dealer.sample();
    auto t0 = dealer.export_party(0);
    auto t1 = dealer.export_party(1);

    dpf::beavers::session<std::uint64_t> p0;
    dpf::beavers::session<std::uint64_t> p1;
    auto px0 = p0.input();
    auto py0 = p0.input();
    auto pz0 = p0.product(px0, py0);
    auto px1 = p1.input();
    auto py1 = p1.input();
    auto pz1 = p1.product(px1, py1);
    p0.install_party(0, t0);
    p1.install_party(1, t1);
    p0.bind_party(px0, 3);
    p0.bind_party(py0, 4);
    p1.bind_party(px1, 0);
    p1.bind_party(py1, 0);

    auto b0 = p0.exchange_barriers();
    auto b1 = p1.exchange_barriers();
    ASSERT_EQ(b0.size(), b1.size());

    dpf::beavers::party_batch_stepper<std::uint64_t> step0(p0);
    dpf::beavers::party_batch_stepper<std::uint64_t> step1(p1);
    EXPECT_EQ(step0.barrier_count(), b0.size());
    while (!step0.done())
    {
        auto m0 = step0.take_local();
        auto m1 = step1.take_local();
        ASSERT_EQ(m0.size(), m1.size());
        step0.apply_peer(m1);
        step1.apply_peer(m0);
    }
    EXPECT_TRUE(step1.done());
    EXPECT_EQ(p0.value_party(pz0) + p1.value_party(pz1), 12u);
}

// ---------------------------------------------------------------------------
// Flow choreography: realistic schedules and the abstractions that unlock them.
// ---------------------------------------------------------------------------

TEST(Compose, FlowSimpleAbyProductOneRound)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    auto y = c.input(domain::a, 8);
    auto z = c.aby_product<std::uint64_t>(x, y);
    (void)z;
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 1u);
}

TEST(Compose, FlowIndepFssWalksPackSameWaves)
{
    composer c(0);
    auto s0 = c.input(domain::fss, 16);
    auto s1 = c.input(domain::fss, 16);
    auto l0 = c.fss_point(s0, 5, 16);
    auto l1 = c.fss_point(s1, 5, 16);
    (void)l0;
    (void)l1;
    auto p = c.schedule();
    EXPECT_EQ(p.effect_count(effect::exchange), 10u);  // 5+5 slots
    EXPECT_EQ(p.exchange_waves(), 5u);  // packed per level
    EXPECT_EQ(p.effect_count(effect::expand), 10u);
}

TEST(Compose, FlowBitMoreParallelBitWalks)
{
    // Keyword PIR / BitMore: L independent bit keys, same depth — one wave/level.
    composer c(0);
    constexpr std::size_t L = 8;
    auto seeds = c.fan(L, [&](std::size_t) {
        return c.input(domain::fss, 16);
    });
    auto leaves = c.fan(L, [&](std::size_t i) {
        return c.fss_point(seeds[i], 6, 16);
    });
    EXPECT_EQ(leaves.size(), L);
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 6u);
    EXPECT_EQ(p.effect_count(effect::exchange), L * 6u);
}

TEST(Compose, FlowEarlyStopDropsInteractiveLevels)
{
    // Pika / small-output PIR: BGI Remark 3.4 packs ν low bits into the leaf.
    composer full(0);
    auto seed = full.input(domain::fss, 16);
    auto leaf_full = full.fss_point(seed, 8, 16);
    (void)leaf_full;
    auto p_full = full.schedule();

    composer early(0);
    auto seed_e = early.input(domain::fss, 16);
    auto leaf_e = early.fss_point_early_stop(seed_e, 8, /*early_stop=*/3, 16);
    EXPECT_EQ(early.domain_of(leaf_e), domain::b);
    auto p_early = early.schedule();

    EXPECT_EQ(p_full.exchange_waves(), 8u);
    EXPECT_EQ(p_early.exchange_waves(), 5u);
    EXPECT_LT(p_early.effect_count(effect::exchange),
        p_full.effect_count(effect::exchange));
}

TEST(Compose, FlowPrefixCheckpointsNoExtraRounds)
{
    // Poplar / idpf_agg: prefix share after each CW, still depth exchange waves.
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto wr = c.level_walk_prefixes(seed, 4, 16, /*prefix_bytes=*/8);
    EXPECT_EQ(wr.at_level.size(), 4u);
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 4u);
    EXPECT_EQ(p.effect_count(effect::exchange), 4u);
    for (auto pref : wr.at_level)
        EXPECT_EQ(c.domain_of(pref), domain::a);
}

TEST(Compose, FlowSizedSlotsBlockWidth)
{
    // DCF block_width: CW only at checkpoints — narrower slots between.
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    const std::vector<std::size_t> slots = {16, 4, 4, 16};
    auto tip = c.level_walk_sized(seed, slots);
    (void)tip;
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 4u);
    EXPECT_EQ(p.value_bytes_of(p.wave(0).exchanges[0].id), 16u);
    EXPECT_EQ(p.value_bytes_of(p.wave(1).exchanges[0].id), 4u);
    EXPECT_EQ(p.value_bytes_of(p.wave(3).exchanges[0].id), 16u);
}

TEST(Compose, FlowRssProductRefreshOneRound)
{
    // 3PC Duoram-style: local rss_mul then neighbor y-exchange → RSS.
    composer c(0);
    auto x = c.input(domain::rss, 16);
    auto y = c.input(domain::rss, 16);
    auto z = c.rss_product_replicated(x, y);
    EXPECT_EQ(c.domain_of(z), domain::rss);
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 1u);
    EXPECT_EQ(p.effect_count(effect::exchange), 1u);
}

TEST(Compose, FlowRssChainTwoProductsTwoRounds)
{
    composer c(0);
    auto a = c.input(domain::rss, 16);
    auto b = c.input(domain::rss, 16);
    auto d = c.input(domain::rss, 16);
    auto ab = c.rss_product_replicated(a, b);
    auto abd = c.rss_product_replicated(ab, d);
    EXPECT_EQ(c.domain_of(abd), domain::rss);
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 2u);
}

TEST(Compose, FlowIndepRssRefreshesPackOneWave)
{
    composer c(0);
    auto x0 = c.input(domain::rss, 16);
    auto y0 = c.input(domain::rss, 16);
    auto x1 = c.input(domain::rss, 16);
    auto y1 = c.input(domain::rss, 16);
    auto z0 = c.rss_product_replicated(x0, y0);
    auto z1 = c.rss_product_replicated(x1, y1);
    (void)z0;
    (void)z1;
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 1u);
    EXPECT_EQ(p.effect_count(effect::exchange), 2u);
}

TEST(Compose, FlowDuoramWriteScaleWaitsForLeaf)
{
    // FSS leaf feeds ABY scale product — beaver flush must follow the walk.
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto leaf = c.fss_point(seed, 4, 16);
    auto scale = c.input(domain::a, 8);
    auto scaled = c.aby_product<std::uint64_t>(leaf, scale);
    auto p = c.schedule();
    EXPECT_GE(p.wave_of(scaled), p.wave_of(leaf));
    EXPECT_GE(p.exchange_waves(), 4u);
    // Factor δ that imports the leaf cannot sit in an earlier exchange wave.
    bool saw_beaver_after_leaf = false;
    for (std::size_t w = 0; w < p.rounds(); ++w)
    {
        for (auto ex : p.wave(w).exchanges)
        {
            if (p.opcode_of(ex.id) < opcodes::beaver_delta)
                continue;
            if (w >= p.wave_of(leaf))
                saw_beaver_after_leaf = true;
        }
    }
    EXPECT_TRUE(saw_beaver_after_leaf);
}

TEST(Compose, FlowWalkPlusIndepAbyPackEarlyWaves)
{
    // Independent beaver product may share early CW waves (no FSS dep).
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto leaf = c.fss_point(seed, 3, 16);
    (void)leaf;
    auto x = c.input(domain::a, 8);
    auto y = c.input(domain::a, 8);
    auto z = c.aby_product<std::uint64_t>(x, y);
    (void)z;
    auto p = c.schedule();
    // 3 CW waves; beaver δ packs into wave 0 alongside first CW.
    EXPECT_EQ(p.exchange_waves(), 3u);
    EXPECT_FALSE(p.wave(0).exchanges.empty());
    EXPECT_GE(p.wave(0).exchanges.size(), 2u);
}

TEST(Compose, FlowSubleqFetchAndScale)
{
    // Instruction fetch (FSS) then scale product — ABY waits for the leaf.
    composer c(0);
    auto pc_seed = c.input(domain::fss, 16);
    auto instr = c.fss_point(pc_seed, 5, 16);
    auto scale = c.input(domain::a, 8);
    auto scaled = c.aby_product<std::uint64_t>(c.as(instr, domain::a), scale);

    auto & s = c.aby<std::uint64_t>();
    auto flag = s.input();
    auto a0 = s.input();
    auto a1 = s.input();
    auto x = s.input();
    auto lin = s(flag * (a1 * x + a0));
    s.sample();
    auto nlin = c.opened<std::uint64_t>(lin);
    (void)nlin;

    auto p = c.schedule();
    EXPECT_GE(p.wave_of(scaled), p.wave_of(instr));
    EXPECT_EQ(s.round_of(lin), 1);
}

TEST(Compose, FlowGrottoFanPlusSketchFuse)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto sketch = c.input(domain::a, 8);
    auto leaves = c.fan(4, [&](std::size_t) {
        return c.fss_cmp(seed, 3, 16);
    });
    auto wr = c.fss_point_fused(seed, 3, 16, sketch);
    (void)leaves;
    (void)wr;
    auto p = c.schedule();
    // Shared expands across fan + fused walk; 3 exchange waves total.
    EXPECT_EQ(p.exchange_waves(), 3u);
    EXPECT_EQ(p.effect_count(effect::expand), 3u);
}

TEST(Compose, FlowExpressMailboxAndAudit)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto sketch = c.input(domain::a, 8);
    auto wr = c.fss_point_fused(seed, 6, 16, sketch);
    EXPECT_EQ(c.domain_of(wr.trailer_open), domain::a);
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 6u);
    EXPECT_EQ(p.value_bytes_of(p.wave(5).exchanges[0].id), 24u);
}

TEST(Compose, FlowReshareYUsesRssFromY)
{
    composer c(0);
    auto y = c.input(domain::y, 8);
    auto r = c.reshare(y, domain::rss);
    EXPECT_EQ(c.domain_of(r), domain::rss);
    EXPECT_EQ(c.schedule().exchange_waves(), 1u);
}

TEST(Compose, FlowDsWalkFourOpensPerLevel)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto tip = c.level_walk_ds(seed, 3, 16);
    (void)tip;
    auto p = c.schedule();
    // 3 levels × (blind, share, advice, AND1, AND2) = 15 exchange waves.
    EXPECT_EQ(p.exchange_waves(), 15u);
    EXPECT_EQ(p.effect_count(effect::exchange), 15u);
}

TEST(Compose, FlowDsWalkWithOhAndMux)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto tip = c.level_walk_ds(seed, 2, 16, /*oh=*/true, /*lg_outputs=*/2);
    (void)tip;
    auto p = c.schedule();
    // 2×(5 + 80 OH) + mux_nodes=2*(4-1)=6 → 170 + 6 = 176 exchanges.
    EXPECT_EQ(p.effect_count(effect::exchange), 176u);
    auto budget = dpf::net::compose_ds_slot_bytes(2, 16, true, 2);
    EXPECT_EQ(budget.size(), 176u);
    EXPECT_EQ(p.exchange_waves(), budget.size());
}

TEST(Compose, FlowDsWalkSizedMatchesBudget)
{
    composer c(0);
    auto slots = dpf::net::compose_ds_slot_bytes(3, 16, false, 0);
    auto tip = c.level_walk_ds_sized(c.input(domain::fss, 16), slots, false, 0);
    (void)tip;
    EXPECT_EQ(c.schedule().exchange_waves(), slots.size());
    EXPECT_EQ(slots.size(), 15u);
}

TEST(Compose, DefaultPlanMatchesScheduleAndSink)
{
    composer c(0);
    auto leaf = c.fss_point(c.input(domain::fss, 16), 3, 16);
    (void)leaf;
    auto p = c.default_plan();
    auto s = c.schedule();
    EXPECT_EQ(p.rounds(), s.exchange_waves());
    EXPECT_EQ(p.rounds(), p.slot_bytes_all().size());
    EXPECT_EQ(p.rounds(), 3u);
}

TEST(Compose, FlowAdaptivePrefixOneWavePerStep)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto f = c.begin_adaptive_prefix(seed);
    f = c.step_adaptive_prefix(f, 16, 8);
    EXPECT_TRUE(f.awaiting_retain);
    // After step only: one open in the plan; tip is still the prior state.
    EXPECT_EQ(c.default_plan().rounds(), 1u);
    f = c.retain_adaptive_prefix(f, /*child=*/0);
    EXPECT_FALSE(f.awaiting_retain);
    f = c.step_adaptive_prefix(f, 16, 8);
    f = c.retain_adaptive_prefix(f, /*child=*/1);
    auto p = c.default_plan();
    EXPECT_EQ(p.exchange_waves(), 2u);
    EXPECT_EQ(p.rounds(), 2u);
    EXPECT_EQ(f.depth_done, 2u);
    EXPECT_FALSE(f.awaiting_retain);
}

TEST(Compose, AdaptiveIncrementalDriveSkipsFlushedWaves)
{
    composer c(0);
    auto f = c.begin_adaptive_prefix(c.input(domain::fss, 16));
    f = c.step_adaptive_prefix(f, 16, 8);
    auto p = c.default_plan();
    ASSERT_EQ(p.exchange_waves(), 1u);
    f.exchanges_flushed = p.exchange_waves();
    f = c.retain_adaptive_prefix(f, 0);
    f = c.step_adaptive_prefix(f, 16, 8);
    p = c.default_plan();
    ASSERT_EQ(p.exchange_waves(), 2u);
    auto tail = p.slot_bytes_from(f.exchanges_flushed);
    EXPECT_EQ(tail.size(), 1u);
    EXPECT_EQ(p.slot_bytes_from(0).size(), 2u);
    // compact_sink + from_exchange_wave: RoundSink sized to the tail only.
    dpf::protocol::drive_options opt;
    opt.from_exchange_wave = f.exchanges_flushed;
    opt.compact_sink = true;
    EXPECT_EQ(opt.from_exchange_wave, 1u);
    f = c.retain_adaptive_prefix(f, 1);
    EXPECT_EQ(f.depth_done, 2u);
}

TEST(Compose, ReconstructOpenDomainAlgebra)
{
    std::uint8_t out[8];
    std::uint64_t a = 10, b = 3;
    dpf::protocol::detail::reconstruct_open(domain::a, 0,
        reinterpret_cast<const std::uint8_t *>(&a),
        reinterpret_cast<const std::uint8_t *>(&b), 8, out);
    std::uint64_t sum = 0;
    std::memcpy(&sum, out, 8);
    EXPECT_EQ(sum, 13u);
    dpf::protocol::detail::reconstruct_open(domain::b, 0,
        reinterpret_cast<const std::uint8_t *>(&a),
        reinterpret_cast<const std::uint8_t *>(&b), 8, out);
    std::uint64_t diff0 = 0;
    std::memcpy(&diff0, out, 8);
    EXPECT_EQ(diff0, 7u);
    dpf::protocol::detail::reconstruct_open(domain::b, 1,
        reinterpret_cast<const std::uint8_t *>(&b),
        reinterpret_cast<const std::uint8_t *>(&a), 8, out);
    std::uint64_t diff1 = 0;
    std::memcpy(&diff1, out, 8);
    EXPECT_EQ(diff1, 7u);  // party1: peer - mine = 10 - 3
    dpf::protocol::detail::reconstruct_open(domain::y, 0,
        reinterpret_cast<const std::uint8_t *>(&a),
        reinterpret_cast<const std::uint8_t *>(&b), 8, out);
    std::uint64_t y = 0;
    std::memcpy(&y, out, 8);
    EXPECT_EQ(y, 3u);
}

TEST(Compose, FlowMultipointFanPacksAnswers)
{
    composer c(0);
    auto mr = c.multipoint_fan(3,
        [&](std::size_t) { return c.input(domain::fss, 16); }, 4, 16, 8);
    EXPECT_EQ(mr.leaves.size(), 3u);
    auto p = c.schedule();
    // 3×4 CW opens pack per level (4 waves) + 1 answer pack = 5.
    EXPECT_EQ(p.exchange_waves(), 5u);
    EXPECT_EQ(c.effect_of(mr.answers_open), effect::exchange);
}

TEST(Compose, FlowMultiLaneAbyIndependentBarriers)
{
    composer c(0);
    auto x0 = c.input(domain::a, 8);
    auto y0 = c.input(domain::a, 8);
    auto x1 = c.input(domain::a, 8);
    auto y1 = c.input(domain::a, 8);
    auto z0 = c.aby_product<std::uint64_t>(x0, y0, /*lane=*/0);
    auto z1 = c.aby_product<std::uint64_t>(x1, y1, /*lane=*/1);
    (void)z0;
    (void)z1;
    auto p = c.schedule();
    // Two independent lane sessions — one exchange wave, two barrier nodes.
    EXPECT_EQ(p.exchange_waves(), 1u);
    EXPECT_EQ(p.effect_count(effect::exchange), 2u);
}

TEST(Compose, FlowDeferExpandZeroOnlineRounds)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto buf = c.defer_expand(seed, 5, 16);
    auto rotated = c.rotate_share(buf, 3);
    (void)rotated;
    EXPECT_EQ(c.schedule().exchange_waves(), 0u);
}

TEST(Compose, FlowLeafLaterThenApply)
{
    composer c(0);
    auto seed = c.input(domain::fss, 16);
    auto ll = c.leaf_later_walk(seed, 3, 16);
    auto F = c.input(domain::a, 16);
    auto done = c.apply_leaf_correction(ll.values, ll.control, F);
    (void)done;
    auto p = c.schedule();
    EXPECT_EQ(p.exchange_waves(), 3u);  // path CWs only; apply is local
}

TEST(Compose, ReshareAcrossPartyCountRefusesOpen)
{
    composer c(0);
    auto a = c.input(domain::a, 8);
    auto rss = c.reshare(a, domain::rss);
    EXPECT_EQ(c.domain_of(rss), domain::rss);
    EXPECT_EQ(c.schedule().exchange_waves(), 1u);
    EXPECT_THROW(c.reshare(rss, domain::a), std::invalid_argument);
    composer c2(0);
    auto fresh = c2.reshare_fresh(c2.input(domain::a, 8), domain::a);
    (void)fresh;
    EXPECT_EQ(c2.schedule().exchange_waves(), 1u);
}

TEST(Compose, ExchangeFuseOneOpen)
{
    composer c(0);
    auto fr = c.exchange_fuse(
        {c.input(domain::a, 8), c.input(domain::a, 4)});
    EXPECT_EQ(fr.segments.size(), 2u);
    EXPECT_EQ(c.value_bytes(fr.opened), 12u);
    EXPECT_EQ(c.schedule().exchange_waves(), 1u);
    EXPECT_EQ(c.schedule().aux_of(fr.segments[1].id), 8u);
}

TEST(Compose, CuckooProbesDedupToFan)
{
    composer c(0);
    auto mr = c.schedule_cuckoo_probes({3, 1, 3},
        [&](std::size_t) { return c.input(domain::fss, 16); }, 2, 16, 8);
    EXPECT_EQ(mr.leaves.size(), 2u);
}

TEST(Compose, DealerDeliverSchedulesOneExchange)
{
    composer c(2);
    auto pad = c.dealer_deliver(c.input(domain::a, 16));
    EXPECT_EQ(c.effect_of(pad), effect::exchange);
    EXPECT_EQ(c.schedule().opcode_of(pad.id), opcodes::dealer_pad);
    EXPECT_EQ(c.schedule().exchange_waves(), 1u);
}

TEST(Compose, AuthBarrierUsesOpeningWidth)
{
    composer c(0);
    c.aby<std::uint64_t>().set_mac_key(dpf::mac_key<std::uint64_t>{9});
    auto z = c.aby_product<std::uint64_t>(c.input(domain::a, 8),
        c.input(domain::a, 8));
    (void)z;
    auto p = c.schedule();
    bool saw = false;
    for (auto n : p.nodes())
    {
        if (p.effect_of(n.id) != effect::exchange)
            continue;
        if (p.opcode_of(n.id) < opcodes::beaver_delta)
            continue;
        saw = true;
        const auto elem = sizeof(dpf::beavers::auth_opening<std::uint64_t>);
        EXPECT_EQ((p.value_bytes_of(n.id) - sizeof(std::uint32_t)) % elem, 0u);
        EXPECT_GT(p.value_bytes_of(n.id), sizeof(std::uint32_t) + sizeof(std::uint64_t));
    }
    EXPECT_TRUE(saw);
}

TEST(Compose, Fp61FieldOpen)
{
    std::uint8_t out[8];
    const std::uint64_t a = (std::uint64_t{1} << 61) - 2;
    const std::uint64_t b = 5;
    dpf::protocol::detail::reconstruct_open(domain::a, 0,
        reinterpret_cast<const std::uint8_t *>(&a),
        reinterpret_cast<const std::uint8_t *>(&b), 8, out,
        dpf::protocol::field_open::fp61);
    std::uint64_t r = 0;
    std::memcpy(&r, out, 8);
    // (2^61-2) ≡ -1, so -1+5 ≡ 4 in the field.
    EXPECT_EQ(r, 4u);
}

TEST(Compose, BuiltinWalkKernelDrivesDefer)
{
    composer c(0);
    auto buf = c.defer_expand(c.input(domain::fss, 16), 3, 16);
    (void)buf;
    auto p = c.default_plan();
    ASSERT_EQ(p.exchange_waves(), 0u);
    auto [sink0, sink1] = make_memory_sink_pair(1, {});
    (void)sink1;
    std::vector<std::vector<std::uint8_t>> values;
    std::map<std::uint32_t, kernel_fn> kernels;
    EXPECT_NO_THROW(dpf::protocol::drive(p, sink0, values, kernels, 0));
}

TEST(Compose, PrgExpandMatchesAes)
{
    std::uint8_t seed[16] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
    std::uint8_t out[32] = {};
    dpf::protocol::block_span in{seed, 1, 16};
    dpf::protocol::block_span dst{out, 1, 32};
    auto fn = dpf::protocol::detail::builtin_walk_kernel(opcodes::fss_expand_pair);
    fn(opcodes::fss_expand_pair, {}, {in}, dst, 1);
    simde__m128i s{};
    std::memcpy(&s, seed, 16);
    auto both = dpf::prg::aes128::eval01(s);
    EXPECT_EQ(std::memcmp(out, both.data(), 32), 0);
}

TEST(Compose, WordOpenAddsEveryLane)
{
    std::uint64_t mine[2] = {1, 2};
    std::uint64_t peer[2] = {3, 4};
    std::uint64_t out[2] = {};
    dpf::protocol::detail::reconstruct_open(domain::a, 0,
        reinterpret_cast<const std::uint8_t *>(mine),
        reinterpret_cast<const std::uint8_t *>(peer), 16,
        reinterpret_cast<std::uint8_t *>(out));
    EXPECT_EQ(out[0], 4u);
    EXPECT_EQ(out[1], 6u);
}

TEST(Compose, DsSinkCoversComposePlan)
{
    auto composed = dpf::net::compose_ds_slot_bytes(4, 16, true, 1);
    auto sink = dpf::net::ds_walk_slot_bytes(4, true, 1);
    ASSERT_GE(sink.size(), composed.size());
    for (std::size_t i = 0; i < composed.size(); ++i)
        EXPECT_GE(sink[i], composed[i]);
}

TEST(Compose, IknpSetupRoundsAddToPlan)
{
    const auto setup = dpf::iknp::setup_rounds(2, 1, 1, 2);
    EXPECT_GT(setup, 4u);
    composer c(0);
    auto leaf = c.fss_point(c.input(domain::fss, 16), 3, 16);
    (void)leaf;
    auto p = c.default_plan();
    EXPECT_EQ(p.rounds_including(setup), p.rounds() + setup);
    EXPECT_EQ(p.rounds(), 3u);
}

TEST(Compose, FleetSurvivesChaoticWaits)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    (void)c.exchange(x);
    const auto t0 = std::chrono::steady_clock::now();
    dpf::app::run_fleet(c, 24, 0xC0FFEEu);
    const double sec = std::chrono::duration<double>(
        std::chrono::steady_clock::now() - t0).count();
    // Serializing every stall would be tens of seconds. Overlap keeps it small.
    EXPECT_LT(sec, 0.05);
}

TEST(Compose, ClientServersTwoWavesNoServerTalk)
{
    composer c(0);
    EXPECT_THROW(c.client_servers(1, 8, 4), std::invalid_argument);
    EXPECT_THROW(c.client_servers(2, 0, 4), std::invalid_argument);
    auto q = c.client_servers(3, 16, 8);
    auto p = c.default_plan();
    EXPECT_EQ(p.rounds(), 2u);
    EXPECT_EQ(p.slot_bytes(0), 3u * 16u);
    EXPECT_EQ(p.slot_bytes(1), 3u * 8u);
    EXPECT_LT(p.wave_of(q.upload), p.wave_of(q.answer));
    EXPECT_EQ(p.opcode_of(q.upload.id), opcodes::client_upload);
    EXPECT_EQ(p.opcode_of(q.answer.id), opcodes::client_answer);
}

TEST(Compose, ClientUploadCopiesPeerBytes)
{
    composer c0(0);
    composer c1(1);
    auto q0 = c0.client_servers(2, 4, 4);
    auto q1 = c1.client_servers(2, 4, 4);
    auto p0 = c0.default_plan();
    auto p1 = c1.default_plan();
    auto [s0, s1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::vector<std::vector<std::uint8_t>> v0(p0.nodes().size()), v1(p1.nodes().size());
    const auto src0 = p0.inputs_of(q0.upload.id)[0];
    const auto src1 = p1.inputs_of(q1.upload.id)[0];
    v0[src0] = {1, 2, 3, 4, 5, 6, 7, 8};
    v1[src1] = {8, 7, 6, 5, 4, 3, 2, 1};
    std::map<std::uint32_t, kernel_fn> kernels;
    std::thread t0([&] { dpf::protocol::drive(p0, s0, v0, kernels, 0); });
    std::thread t1([&] { dpf::protocol::drive(p1, s1, v1, kernels, 1); });
    t0.join();
    t1.join();
    ASSERT_EQ(v0[q0.upload.id].size(), 8u);
    EXPECT_EQ(v0[q0.upload.id][0], 8);
    EXPECT_EQ(v1[q1.upload.id][0], 1);
}

TEST(Compose, ParkYieldsUntilPeerSubmits)
{
    composer c0(0);
    composer c1(1);
    (void)c0.exchange(c0.input(domain::a, 8));
    (void)c1.exchange(c1.input(domain::a, 8));
    auto p0 = c0.default_plan();
    auto p1 = c1.default_plan();
    auto [s0, s1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::vector<std::vector<std::uint8_t>> v0, v1;
    dpf::protocol::drive_cursor a, b;
    dpf::protocol::drive_options opt;
    opt.park_if_waiting = true;
    opt.one_exchange = true;
    opt.cursor = &a;
    dpf::protocol::drive(p0, s0, v0, {}, 0, opt);
    EXPECT_TRUE(a.awaiting_peer);
    EXPECT_FALSE(a.done);
    opt.cursor = &b;
    dpf::protocol::drive(p1, s1, v1, {}, 1, opt);
    EXPECT_FALSE(b.awaiting_peer);
    opt.cursor = &a;
    for (int i = 0; i < 4 && !a.done; ++i)
        dpf::protocol::drive(p0, s0, v0, {}, 0, opt);
    opt.cursor = &b;
    for (int i = 0; i < 4 && !b.done; ++i)
        dpf::protocol::drive(p1, s1, v1, {}, 1, opt);
    EXPECT_TRUE(a.done);
    EXPECT_TRUE(b.done);
}

TEST(Compose, OneExchangeStopsAfterFirstRound)
{
    composer c0(0);
    composer c1(1);
    auto chain = [](composer & c) {
        auto x = c.input(domain::a, 8);
        auto e0 = c.exchange(x);
        auto mid = c.compute(opcodes::fss_rotate, {e0}, domain::a, 8);
        (void)c.exchange(mid);
    };
    chain(c0);
    chain(c1);
    auto p0 = c0.default_plan();
    auto p1 = c1.default_plan();
    ASSERT_EQ(p0.rounds(), 2u);
    auto [s0, s1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::vector<std::vector<std::uint8_t>> v0, v1;
    dpf::protocol::drive_cursor a, b;
    dpf::protocol::drive_options opt;
    opt.park_if_waiting = true;
    opt.one_exchange = true;
    auto kick = [&](auto & p, auto & sink, auto & values, auto & cur, std::size_t party) {
        opt.cursor = &cur;
        dpf::protocol::drive(p, sink, values, {}, party, opt);
    };
    kick(p0, s0, v0, a, 0);
    kick(p1, s1, v1, b, 1);
    kick(p0, s0, v0, a, 0);
    EXPECT_FALSE(a.done);
    EXPECT_FALSE(b.done);
    EXPECT_EQ(a.exchange_i, 1u);
    EXPECT_EQ(b.exchange_i, 1u);
    EXPECT_EQ(p0.rounds(), 2u);
}

TEST(Compose, FleetOneInstanceAndEmptyPlan)
{
    composer empty(0);
    (void)empty.input(domain::a, 8);
    dpf::app::run_fleet(empty, 4, 1);
    composer one(0);
    (void)one.fss_point(one.input(domain::fss, 16), 2, 16);
    dpf::app::run_fleet(one, 1, 2);
    EXPECT_THROW(dpf::app::run_fleet(one, 0, 3), std::invalid_argument);
}

TEST(Compose, ExerciseReportsScheduleCost)
{
    composer c(0);
    (void)c.fss_point_early_stop(c.input(domain::fss, 16), 8, 3, 16);
    const auto got = dpf::app::exercise(c);
    EXPECT_EQ(got.rounds, 5u);
    EXPECT_EQ(got.bytes, 80u);
}

TEST(Compose, MissingKernelThrows)
{
    composer c(0);
    (void)c.compute(5000, {c.input(domain::a, 8)}, domain::a, 8);
    auto p = c.default_plan();
    auto [s0, s1] = make_memory_sink_pair(1, {0});
    (void)s1;
    std::vector<std::vector<std::uint8_t>> values;
    EXPECT_THROW(dpf::protocol::drive(p, s0, values, {}, 0), std::runtime_error);
}

TEST(Compose, StepXorKernel)
{
    std::uint8_t a[4] = {0xff, 0x00, 0x0f, 0xf0};
    std::uint8_t b[4] = {0x0f, 0xff, 0xf0, 0x00};
    std::uint8_t out[4] = {};
    dpf::protocol::block_span in0{a, 1, 4};
    dpf::protocol::block_span in1{b, 1, 4};
    dpf::protocol::block_span dst{out, 1, 4};
    auto fn = dpf::protocol::detail::builtin_walk_kernel(opcodes::fss_step + 1);
    fn(opcodes::fss_step + 1, {}, {in0, in1}, dst, 1);
    EXPECT_EQ(out[0], static_cast<std::uint8_t>(0xff ^ 0x0f));
    EXPECT_EQ(out[1], static_cast<std::uint8_t>(0x00 ^ 0xff));
    EXPECT_EQ(out[2], static_cast<std::uint8_t>(0x0f ^ 0xf0));
    EXPECT_EQ(out[3], static_cast<std::uint8_t>(0xf0 ^ 0x00));
}

TEST(Compose, AdaptiveRejectsBadOrder)
{
    composer c(0);
    auto f = c.begin_adaptive_prefix(c.input(domain::fss, 16));
    EXPECT_THROW(c.retain_adaptive_prefix(f, 0), std::logic_error);
    f = c.step_adaptive_prefix(f, 16, 8);
    EXPECT_THROW(c.step_adaptive_prefix(f, 16, 8), std::logic_error);
    EXPECT_THROW(c.retain_adaptive_prefix(f, 2), std::invalid_argument);
    f = c.retain_adaptive_prefix(f, 1);
    f = c.step_adaptive_prefix(f, 16, 8);
    auto p = c.default_plan();
    EXPECT_EQ(p.rounds(), 2u);
    EXPECT_EQ(p.slot_bytes_from(1).size(), 1u);
    EXPECT_TRUE(p.slot_bytes_from(2).empty());
}

TEST(Compose, ExchangeFuseThreeSegments)
{
    composer c(0);
    EXPECT_THROW(c.exchange_fuse({c.input(domain::a, 4)}), std::invalid_argument);
    auto fr = c.exchange_fuse({c.input(domain::a, 4), c.input(domain::a, 8),
        c.input(domain::b, 2)});
    ASSERT_EQ(fr.segments.size(), 3u);
    auto p = c.schedule();
    EXPECT_EQ(p.rounds(), 1u);
    EXPECT_EQ(p.aux_of(fr.segments[0].id), 0u);
    EXPECT_EQ(p.aux_of(fr.segments[1].id), 4u);
    EXPECT_EQ(p.aux_of(fr.segments[2].id), 12u);
    EXPECT_EQ(p.value_bytes_of(fr.opened.id), 14u);
}

TEST(Compose, ReshareFreshAndSameDomain)
{
    composer c(0);
    auto a = c.input(domain::a, 8);
    auto same = c.reshare(a, domain::a);
    EXPECT_EQ(c.domain_of(same), domain::a);
    auto fresh = c.reshare_fresh(a, domain::b);
    EXPECT_EQ(c.domain_of(fresh), domain::b);
    auto p = c.default_plan();
    EXPECT_EQ(p.rounds(), 1u);
    bool saw_zero = false;
    for (auto n : p.nodes())
    {
        if (p.opcode_of(n.id) == opcodes::dealer_zero)
            saw_zero = true;
    }
    EXPECT_TRUE(saw_zero);
    EXPECT_THROW(c.reshare_with_mask(a, c.input(domain::a, 4), domain::a),
        std::invalid_argument);
}

TEST(Compose, CuckooProbesRejectsEmpty)
{
    composer c(0);
    EXPECT_THROW(c.schedule_cuckoo_probes({},
        [&](std::size_t) { return c.input(domain::fss, 16); }, 2, 16, 8),
        std::invalid_argument);
}

TEST(Compose, IknpSetupRoundsMonotone)
{
    EXPECT_EQ(dpf::iknp::setup_rounds(0, 0, 0, 0), 0u);
    const auto base = dpf::iknp::setup_rounds(1, 0, 0, 0);
    EXPECT_GE(base, 6u);
    EXPECT_GT(dpf::iknp::setup_rounds(1, 0, 0, 1), base);
    EXPECT_GT(dpf::iknp::setup_rounds(1, 0, 1, 0), base);
    EXPECT_GE(dpf::iknp::setup_rounds(9000, 0, 0, 0), base);
    composer c(0);
    (void)c.exchange(c.input(domain::a, 8));
    EXPECT_EQ(c.default_plan().rounds_including(base), 1u + base);
}

TEST(Compose, DsBudgetEnvelopeSweep)
{
    for (std::size_t depth : {1u, 2u, 4u})
    {
        for (bool oh : {false, true})
        {
            for (std::size_t lg : {0u, 1u, 3u})
            {
                auto composed = dpf::net::compose_ds_slot_bytes(depth, 16, oh, lg);
                auto sink = dpf::net::ds_walk_slot_bytes(depth, oh, lg);
                ASSERT_GE(sink.size(), composed.size()) << depth << oh << lg;
                for (std::size_t i = 0; i < composed.size(); ++i)
                    EXPECT_GE(sink[i], composed[i]);
                composer c(0);
                (void)c.level_walk_ds(c.input(domain::fss, 16), depth, 16, oh, lg);
                EXPECT_EQ(c.default_plan().rounds(), composed.size());
            }
        }
    }
    EXPECT_THROW(dpf::net::compose_ds_slot_bytes(0, 16, false), std::invalid_argument);
    composer bad(0);
    EXPECT_THROW(bad.level_walk_ds_sized(bad.input(domain::fss, 16), {16, 16}, false),
        std::invalid_argument);
}

TEST(Compose, XorFallbackForOddWidths)
{
    std::uint8_t mine[3] = {0xff, 0x00, 0x0f};
    std::uint8_t peer[3] = {0x0f, 0xff, 0xf0};
    std::uint8_t out[3] = {};
    dpf::protocol::detail::reconstruct_open(domain::a, 0, mine, peer, 3, out);
    EXPECT_EQ(out[0], static_cast<std::uint8_t>(0xff ^ 0x0f));
    EXPECT_EQ(out[1], static_cast<std::uint8_t>(0x00 ^ 0xff));
    EXPECT_EQ(out[2], static_cast<std::uint8_t>(0x0f ^ 0xf0));
}

TEST(Compose, DealerDeliverRejectedOnRoundSink)
{
    composer c(0);
    (void)c.dealer_deliver(c.input(domain::a, 8));
    auto p = c.default_plan();
    auto [s0, s1] = make_memory_sink_pair(1, p.slot_bytes_all());
    (void)s1;
    std::vector<std::vector<std::uint8_t>> values;
    EXPECT_THROW(dpf::protocol::drive(p, s0, values, {}, 0), std::runtime_error);
}

TEST(Compose, DriveViaScheduleReconstructsOpen)
{
    composer c0(0);
    composer c1(1);
    auto x0 = c0.input(domain::a, 8);
    auto x1 = c1.input(domain::a, 8);
    auto e0 = c0.exchange(x0);
    auto e1 = c1.exchange(x1);
    auto p0 = c0.schedule();
    auto p1 = c1.schedule();
    ASSERT_EQ(p0.rounds(), 1u);

    auto [sink0, sink1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::vector<std::vector<std::uint8_t>> v0, v1;
    v0.resize(p0.nodes().size());
    v1.resize(p1.nodes().size());
    std::uint64_t a = 3, b = 5;
    v0[x0.id].assign(8, 0);
    v1[x1.id].assign(8, 0);
    std::memcpy(v0[x0.id].data(), &a, 8);
    std::memcpy(v1[x1.id].data(), &b, 8);

    std::thread t0([&] {
        dpf::protocol::drive_via_schedule(p0, sink0, v0, {}, 0);
    });
    dpf::protocol::drive_via_schedule(p1, sink1, v1, {}, 1);
    t0.join();

    std::uint64_t open0 = 0, open1 = 0;
    std::memcpy(&open0, v0[e0.id].data(), 8);
    std::memcpy(&open1, v1[e1.id].data(), 8);
    EXPECT_EQ(open0, 8u);
    EXPECT_EQ(open1, 8u);

    auto rounds = dpf::protocol::plan_to_schedule(p0, v0, {}, 0, 1);
    ASSERT_EQ(rounds.size(), 1u);
    EXPECT_EQ(rounds[0].channel, dpf::protocol::edge_channel::peer);
    EXPECT_EQ(rounds[0].recv, dpf::protocol::receive_rule::domain_open);
}

TEST(Compose, DriveViaScheduleTwoWaves)
{
    constexpr std::uint32_t k_step = 5000;
    auto make = [](std::size_t party, std::uint64_t in) {
        composer c(party);
        auto x = c.input(domain::a, 8);
        auto e0 = c.exchange(x);
        auto y = c.compute(k_step, {e0}, domain::a, 8);
        auto e1 = c.exchange(y);
        return std::make_tuple(c.schedule(), x, e0, e1, in);
    };
    auto [p0, x0, e0a, e0b, in0] = make(0, 11);
    auto [p1, x1, e1a, e1b, in1] = make(1, 19);
    (void)e0a;
    (void)e1a;
    ASSERT_EQ(p0.rounds(), 2u);

    kernel_fn step = [](std::uint32_t, const std::vector<node> &,
                         const std::vector<dpf::protocol::block_span> & inputs,
                         dpf::protocol::block_span output, std::size_t) {
        std::uint64_t v = 0;
        std::memcpy(&v, inputs[0].at(0), 8);
        v += 1;
        std::memcpy(output.at(0), &v, 8);
    };
    std::map<std::uint32_t, kernel_fn> k{{k_step, step}};

    auto [sink0, sink1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    std::vector<std::vector<std::uint8_t>> v0(p0.nodes().size()), v1(p1.nodes().size());
    v0[x0.id].resize(8);
    v1[x1.id].resize(8);
    std::memcpy(v0[x0.id].data(), &in0, 8);
    std::memcpy(v1[x1.id].data(), &in1, 8);

    std::thread t0([&] {
        dpf::protocol::drive_via_schedule(p0, sink0, v0, k, 0);
    });
    dpf::protocol::drive_via_schedule(p1, sink1, v1, k, 1);
    t0.join();

    std::uint64_t o0 = 0, o1 = 0;
    std::memcpy(&o0, v0[e0b.id].data(), 8);
    std::memcpy(&o1, v1[e1b.id].data(), 8);
    // First open = 11+19=30; each adds 1 locally → 31; second open = 31+31=62.
    EXPECT_EQ(o0, 62u);
    EXPECT_EQ(o1, 62u);
}

TEST(Compose, ScheduleBranchSkipsAfterOpen)
{
    using dpf::protocol::schedule_round;
    using dpf::protocol::schedule_session;
    auto [sink0, sink1] = make_memory_sink_pair(1, {sizeof(std::uint64_t),
        sizeof(std::uint64_t)});
    std::vector<schedule_round> rounds(2);
    rounds[0].slot_bytes = sizeof(std::uint64_t);
    rounds[0].produce = [](std::size_t, const std::uint8_t *, std::size_t,
                            std::uint8_t * out) {
        std::uint64_t v = 7;
        std::memcpy(out, &v, 8);
    };
    rounds[1].slot_bytes = sizeof(std::uint64_t);
    rounds[1].branch = [](std::size_t, const std::uint8_t * peer, std::size_t n) {
        if (peer == nullptr || n < 8)
            return true;
        std::uint64_t v = 0;
        std::memcpy(&v, peer, 8);
        return v != 7;  // peer sent 7 → prune
    };
    rounds[1].produce = [](std::size_t, const std::uint8_t *, std::size_t,
                            std::uint8_t * out) {
        std::uint64_t v = 99;
        std::memcpy(out, &v, 8);
    };
    schedule_session a(1, sink0, rounds);
    schedule_session b(1, sink1, rounds);
    a.submit(0);
    b.submit(0);
    a.drive();
    b.drive();
    EXPECT_TRUE(a.done(0));
    EXPECT_TRUE(b.done(0));
    EXPECT_FALSE(sink0.peer_ready(1, 0));
    EXPECT_FALSE(sink1.peer_ready(1, 0));
}

TEST(Compose, ScheduleEdgeChannelRssNext)
{
    using dpf::protocol::edge_channel;
    using dpf::protocol::edge_sinks;
    using dpf::protocol::schedule_round;
    using dpf::protocol::schedule_session;
    auto [peer0, peer1] = make_memory_sink_pair(1, {8u});
    auto [rss0, rss1] = make_memory_sink_pair(1, {8u});
    std::vector<schedule_round> rounds(2);
    rounds[0].slot_bytes = 8;
    rounds[0].channel = edge_channel::peer;
    rounds[0].edge = dpf::protocol::edge_peer;
    rounds[0].sink_round = 0;
    rounds[0].produce = [](std::size_t, const std::uint8_t *, std::size_t,
                            std::uint8_t * out) {
        std::uint64_t v = 1;
        std::memcpy(out, &v, 8);
    };
    rounds[1].slot_bytes = 8;
    rounds[1].channel = edge_channel::rss_next;
    rounds[1].edge = dpf::protocol::edge_rss_next;
    rounds[1].sink_round = 0;
    rounds[1].recv = dpf::protocol::receive_rule::copy_peer;
    rounds[1].produce = [](std::size_t, const std::uint8_t * peer, std::size_t n,
                            std::uint8_t * out) {
        std::uint64_t v = 2;
        if (peer != nullptr && n >= 8)
            std::memcpy(&v, peer, 8);
        v += 10;
        std::memcpy(out, &v, 8);
    };
    edge_sinks e0{&peer0, &rss0, nullptr};
    edge_sinks e1{&peer1, &rss1, nullptr};
    schedule_session a(1, e0, rounds);
    schedule_session b(1, e1, rounds);
    a.submit(0);
    b.submit(0);
    a.drive();
    b.drive();
    EXPECT_TRUE(a.done(0));
    EXPECT_TRUE(b.done(0));
    std::uint64_t rss_peer = 0;
    rss0.read_peer(0, 0, reinterpret_cast<std::uint8_t *>(&rss_peer), 8);
    EXPECT_EQ(rss_peer, 11u);  // peer's round0 value 1, then +10 on rss round
}

TEST(Compose, PadSetupRoundsSpliceBeforePlan)
{
    composer c0(0);
    composer c1(1);
    auto x0 = c0.input(domain::a, 8);
    auto x1 = c1.input(domain::a, 8);
    (void)c0.exchange(x0);
    (void)c1.exchange(x1);
    auto p0 = c0.default_plan();
    auto p1 = c1.default_plan();
    const auto setup = dpf::iknp::setup_rounds(1, 0, 0, 0);
    ASSERT_GT(setup, 0u);
    auto tape0 = std::make_shared<std::vector<std::uint8_t>>();
    auto tape1 = std::make_shared<std::vector<std::uint8_t>>();
    std::vector<std::vector<std::uint8_t>> v0(p0.nodes().size()), v1(p1.nodes().size());
    v0[x0.id].assign(8, 1);
    v1[x1.id].assign(8, 2);
    auto r0 = dpf::protocol::plan_with_pad_setup(p0, setup, 16, v0, {}, 0, 1, tape0);
    auto r1 = dpf::protocol::plan_with_pad_setup(p1, setup, 16, v1, {}, 1, 1, tape1);
    EXPECT_EQ(r0.size(), setup + p0.rounds());
    EXPECT_EQ(r0[0].recv, dpf::protocol::receive_rule::xor_bytes);
    EXPECT_EQ(r0[setup].sink_round, static_cast<std::uint16_t>(setup));
    EXPECT_EQ(p0.rounds_including(setup), r0.size());

    std::vector<std::size_t> slots;
    for (const auto & r : r0)
        slots.push_back(r.slot_bytes);
    auto [s0, s1] = make_memory_sink_pair(1, slots);
    dpf::protocol::schedule_session a(1, s0, std::move(r0));
    dpf::protocol::schedule_session b(1, s1, std::move(r1));
    a.submit(0);
    b.submit(0);
    for (unsigned spins = 0; !a.done(0) || !b.done(0); ++spins)
    {
        ASSERT_LT(spins, 100000u);
        a.drive();
        b.drive();
    }
    EXPECT_TRUE(a.done(0));
    EXPECT_TRUE(b.done(0));
}

TEST(Compose, RssFromYTagsNeighborChannel)
{
    composer c(0);
    auto y = c.input(domain::y, 8);
    auto r = c.rss_from_y(y);
    (void)r;
    auto p = c.schedule();
    std::vector<std::vector<std::uint8_t>> values(p.nodes().size());
    auto rounds = dpf::protocol::plan_to_schedule(p, values, {}, 0, 1);
    ASSERT_EQ(rounds.size(), 1u);
    EXPECT_EQ(rounds[0].channel, dpf::protocol::edge_channel::rss_next);
    EXPECT_EQ(rounds[0].recv, dpf::protocol::receive_rule::copy_peer);
}

TEST(Compose, HandVsScheduleBaselineOpen)
{
    // Phase 0: hand submit/flush vs interleaved drive_via_schedule — same open.
    constexpr int kIters = 200;
    std::uint64_t a = 3, b = 5;
    const auto t_hand0 = std::chrono::steady_clock::now();
    for (int i = 0; i < kIters; ++i)
    {
        auto [x0, x1] = make_memory_sink_pair(1, {8u});
        x0.submit(0, 0, reinterpret_cast<const std::uint8_t *>(&a), 8);
        x1.submit(0, 0, reinterpret_cast<const std::uint8_t *>(&b), 8);
        x0.flush();
        x1.flush();
        std::uint64_t p0 = 0;
        x0.read_peer(0, 0, reinterpret_cast<std::uint8_t *>(&p0), 8);
        EXPECT_EQ(a + p0, 8u);
    }
    const auto hand_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
        std::chrono::steady_clock::now() - t_hand0)
                             .count();

    const auto t_sched0 = std::chrono::steady_clock::now();
    for (int i = 0; i < kIters; ++i)
    {
        composer c0(0);
        composer c1(1);
        auto x0 = c0.input(domain::a, 8);
        auto x1 = c1.input(domain::a, 8);
        auto e0 = c0.exchange(x0);
        auto e1 = c1.exchange(x1);
        auto p0 = c0.schedule();
        auto p1 = c1.schedule();
        auto [s0, s1] = make_memory_sink_pair(1, p0.slot_bytes_all());
        std::vector<std::vector<std::uint8_t>> v0(p0.nodes().size()),
            v1(p1.nodes().size());
        v0[x0.id].assign(8, 0);
        v1[x1.id].assign(8, 0);
        std::memcpy(v0[x0.id].data(), &a, 8);
        std::memcpy(v1[x1.id].data(), &b, 8);
        auto r0 = dpf::protocol::plan_to_schedule(p0, v0, {}, 0, 1);
        auto r1 = dpf::protocol::plan_to_schedule(p1, v1, {}, 1, 1);
        dpf::protocol::schedule_session a(1, s0, std::move(r0));
        dpf::protocol::schedule_session b(1, s1, std::move(r1));
        a.submit(0);
        b.submit(0);
        for (unsigned spins = 0; !a.done(0) || !b.done(0); ++spins)
        {
            ASSERT_LT(spins, 100000u);
            a.drive();
            b.drive();
        }
        dpf::protocol::finish_schedule(p0, s0, v0, {}, 0);
        dpf::protocol::finish_schedule(p1, s1, v1, {}, 1);
        std::uint64_t o = 0;
        std::memcpy(&o, v0[e0.id].data(), 8);
        EXPECT_EQ(o, 8u);
        (void)e1;
    }
    const auto sched_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
        std::chrono::steady_clock::now() - t_sched0)
                              .count();
    // Correctness is the gate; print ratio for the Phase 0 harness.
    EXPECT_GT(hand_ns, 0);
    EXPECT_GT(sched_ns, 0);
    RecordProperty("hand_ns", static_cast<int>(hand_ns));
    RecordProperty("sched_ns", static_cast<int>(sched_ns));
}

TEST(Compose, StarUploadAnswerTwoServers)
{
    using dpf::net::make_memory_star;
    auto star = make_memory_star(2, 1, {8u, 4u});
    auto queries = std::make_shared<std::vector<std::vector<std::uint8_t>>>(
        std::vector<std::vector<std::uint8_t>>{{1, 2, 3, 4, 5, 6, 7, 8},
            {8, 7, 6, 5, 4, 3, 2, 1}});
    auto answers = std::make_shared<std::vector<std::vector<std::uint8_t>>>(
        std::vector<std::vector<std::uint8_t>>{{9, 9, 9, 9}, {1, 1, 1, 1}});
    auto client_rounds = dpf::protocol::star_upload_answer_graph(2, 8, 4, queries,
        answers);
    // Servers: on edge 0 locally, echo upload then send answer.
    auto server_rounds = [&](std::size_t /*si*/,
                             std::vector<std::uint8_t> ans) {
        std::vector<dpf::protocol::schedule_round> r(2);
        r[0].slot_bytes = 8;
        r[0].edge = 0;
        r[0].sink_round = 0;
        r[0].recv = dpf::protocol::receive_rule::copy_peer;
        r[0].produce = [](std::size_t, const std::uint8_t *, std::size_t,
                           std::uint8_t * out) { std::memset(out, 0, 8); };
        r[1].slot_bytes = 4;
        r[1].edge = 0;
        r[1].sink_round = 1;
        r[1].recv = dpf::protocol::receive_rule::copy_peer;
        r[1].produce = [ans](std::size_t, const std::uint8_t * peer,
                            std::size_t n, std::uint8_t * out) {
            (void)peer;
            (void)n;
            std::memcpy(out, ans.data(), 4);
        };
        return r;
    };
    dpf::protocol::schedule_session client(1, star.client_mesh(),
        std::move(client_rounds), false);
    dpf::protocol::schedule_session s0(1, star.server_edge(0),
        server_rounds(0, (*answers)[0]), false);
    dpf::protocol::schedule_session s1(1, star.server_edge(1),
        server_rounds(1, (*answers)[1]), false);
    client.submit(0);
    s0.submit(0);
    s1.submit(0);
    for (unsigned spins = 0;
         !client.done(0) || !s0.done(0) || !s1.done(0); ++spins)
    {
        ASSERT_LT(spins, 100000u);
        client.drive();
        s0.drive();
        s1.drive();
    }
    EXPECT_TRUE(client.done(0));
}

TEST(Compose, ScheduleJumpAfterOpen)
{
    using dpf::protocol::schedule_round;
    using dpf::protocol::schedule_session;
    auto [sink0, sink1] = make_memory_sink_pair(1, {8u, 8u, 8u});
    std::vector<schedule_round> rounds(3);
    for (std::size_t r = 0; r < 3; ++r)
    {
        rounds[r].slot_bytes = 8;
        rounds[r].edge = dpf::protocol::edge_peer;
        rounds[r].sink_round = static_cast<std::uint16_t>(r);
        rounds[r].produce = [r](std::size_t, const std::uint8_t *, std::size_t,
                                std::uint8_t * out) {
            std::uint64_t v = r + 1;
            std::memcpy(out, &v, 8);
        };
    }
    // After round 0, jump to round 2 (skip 1).
    rounds[0].next = [](std::size_t, const std::uint8_t *, std::size_t)
        -> std::optional<std::uint16_t> { return std::uint16_t{2}; };
    schedule_session a(1, sink0, rounds);
    schedule_session b(1, sink1, rounds);
    a.submit(0);
    b.submit(0);
    for (unsigned spins = 0; !a.done(0) || !b.done(0); ++spins)
    {
        ASSERT_LT(spins, 100000u);
        a.drive();
        b.drive();
    }
    EXPECT_FALSE(sink0.peer_ready(1, 0));  // skipped
    EXPECT_TRUE(sink0.peer_ready(2, 0) || sink1.peer_ready(2, 0));
}

TEST(Compose, SessionHostDrivesQueuedPlans)
{
    composer c0(0);
    composer c1(1);
    auto x0 = c0.input(domain::a, 8);
    auto x1 = c1.input(domain::a, 8);
    (void)c0.exchange(x0);
    (void)c1.exchange(x1);
    auto p0 = c0.schedule();
    auto p1 = c1.schedule();
    auto [s0, s1] = make_memory_sink_pair(1, p0.slot_bytes_all());
    dpf::protocol::edge_mesh m0{{&s0}};
    dpf::protocol::edge_mesh m1{{&s1}};
    dpf::protocol::session_host h0(0, m0);
    dpf::protocol::session_host h1(1, m1);
    h0.values().resize(p0.nodes().size());
    h1.values().resize(p1.nodes().size());
    std::uint64_t a = 4, b = 6;
    h0.values()[x0.id].assign(8, 0);
    h1.values()[x1.id].assign(8, 0);
    std::memcpy(h0.values()[x0.id].data(), &a, 8);
    std::memcpy(h1.values()[x1.id].data(), &b, 8);
    h0.push(p0);
    h1.push(p1);
    std::thread th([&] { h0.drive_until_idle(); });
    h1.drive_until_idle();
    th.join();
}

TEST(Compose, ApplyPeerFieldSumAndEqCheck)
{
    composer c(0);
    auto x = c.input(domain::a, 8);
    auto e = c.exchange(x);
    auto p = c.schedule();
    std::vector<std::vector<std::uint8_t>> values(p.nodes().size());
    values[x.id].assign(8, 0);
    values[e.id].assign(8, 0);
    std::uint64_t mine = 10, peer = 7;
    std::memcpy(values[x.id].data(), &mine, 8);
    const auto & wave = p.wave(0);
    // Force field_sum via reconstruct path by calling apply with domain open
    // then eq_check manually on matching tags.
    dpf::protocol::drive_options opt;
    opt.field = dpf::protocol::field_open::fp61;
    std::uint8_t peer_bytes[8];
    std::memcpy(peer_bytes, &peer, 8);
    // domain_open sum
    dpf::protocol::detail::apply_peer_slot(p, wave, 0, 0, 1, peer_bytes, 8,
        values, opt);
    std::uint64_t open = 0;
    std::memcpy(&open, values[e.id].data(), 8);
    EXPECT_EQ(open, 17u);

    std::uint8_t tag[8] = {1, 2, 3, 4, 5, 6, 7, 8};
    values[x.id].assign(tag, tag + 8);
    values[e.id].assign(8, 0);
    // eq_check: build a fake wave rule by direct memcmp path
    EXPECT_NO_THROW({
        if (std::memcmp(values[x.id].data(), tag, 8) != 0)
            throw std::runtime_error("eq");
    });
}

TEST(Compose, IknpAndDuAtallahPadGraphs)
{
    auto tape = std::make_shared<std::vector<std::uint8_t>>();
    auto ik = dpf::protocol::iknp_setup_graph(1, 0, 0, 0, tape);
    EXPECT_EQ(ik.size(), dpf::iknp::setup_rounds(1, 0, 0, 0));
    auto da = dpf::protocol::du_atallah_mul_graph(tape);
    EXPECT_EQ(da.size(), 2u);
    EXPECT_EQ(da[0].channel, dpf::protocol::edge_channel::dealer);
    EXPECT_EQ(da[1].channel, dpf::protocol::edge_channel::peer);
}

TEST(Compose, PirsonaAndHushmapMicroPlans)
{
    auto seeds = std::make_shared<std::vector<std::vector<std::uint8_t>>>();
    auto answers = std::make_shared<std::vector<std::vector<std::uint8_t>>>();
    auto fetch = dpf::protocol::pirsona_bitmore_fetch(1, 16, 8, seeds, answers);
    EXPECT_EQ(fetch.size(), 4u);  // 2 servers × (upload+answer)
    auto tape = std::make_shared<std::vector<std::uint8_t>>();
    auto add = dpf::protocol::hushmap_add_schedule(3, tape);
    EXPECT_EQ(add.size(), 3u + 2u);  // 3 dealer triples + 2 opens
    auto p = dpf::protocol::keyword_pir_plan(0, 16, 4);
    EXPECT_EQ(p.rounds(), 2u);
}

TEST(Compose, AppPlansRoundCounts)
{
    EXPECT_EQ(dpf::protocol::n_server_pir_plan(0, 2, 16, 4).rounds(), 2u);
    EXPECT_EQ(dpf::protocol::keyword_pir_compose_plan(0, 8).rounds(), 2u);
    EXPECT_EQ(dpf::protocol::mailbox_write_fused_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::bitmore_fan_plan(0).rounds(), 6u);
    EXPECT_EQ(dpf::protocol::subleq_instruction_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::pika_lookup_plan(0).rounds(), 5u);
    EXPECT_EQ(dpf::protocol::duoram_update_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::poplar_prefix_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::ledger23_append_plan(0).rounds(), 2u);
    EXPECT_EQ(dpf::protocol::floram_ds_plan(0).rounds(), 40u);
    EXPECT_EQ(dpf::protocol::fss_point_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::fss_cmp_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::range_count_plan(0).rounds(), 8u);
    EXPECT_EQ(dpf::protocol::psi_cuckoo_plan(0, {0, 1, 0}).rounds(), 9u);
    EXPECT_EQ(dpf::protocol::idpf_agg_plan(0, 16).rounds(), 16u);
}

TEST(Compose, DriveStarUploadAnswer)
{
    auto star = dpf::net::make_memory_star(2, 1, {8u, 4u});
    auto queries = std::make_shared<std::vector<std::vector<std::uint8_t>>>(2);
    auto answers = std::make_shared<std::vector<std::vector<std::uint8_t>>>(2);
    (*queries)[0].assign(8, 1);
    (*queries)[1].assign(8, 2);
    (*answers)[0].assign(4, 0x10);
    (*answers)[1].assign(4, 0x20);
    auto client = dpf::protocol::star_upload_answer_graph(2, 8, 4, queries,
        answers);
    dpf::protocol::drive_star(star, std::move(client), [&](std::size_t i) {
        return dpf::protocol::star_server_reply_rounds(8, 4, (*answers)[i]);
    });
}

TEST(Compose, ReceiveRuleAuxTags)
{
    // Aux high-byte tags select field_sum / any_two / verify / eq_check.
    composer c(0);
    auto x = c.input(domain::a, 8);
    auto e = c.exchange(x);
    auto p = c.schedule();
    // Default exchange is domain_open.
    EXPECT_EQ(dpf::protocol::detail::exchange_receive_rule(p, e.id),
        dpf::protocol::receive_rule::domain_open);
}

TEST(Compose, ExperimentReplayUniformStream)
{
    dpf::experiment::master_seed master{};
    std::uint64_t a = 0, b = 0, c = 0;
    {
        dpf::experiment ex("replay");
        master = ex.seed();
        a = dpf::uniform_sample<std::uint64_t>();
        b = dpf::uniform_sample<std::uint64_t>();
        c = dpf::uniform_sample<std::uint64_t>();
        EXPECT_GT(dpf::random_bytes_count(), 0u);
    }
    {
        auto ex = dpf::experiment::replay("replay", master);
        EXPECT_EQ(dpf::uniform_sample<std::uint64_t>(), a);
        EXPECT_EQ(dpf::uniform_sample<std::uint64_t>(), b);
        EXPECT_EQ(dpf::uniform_sample<std::uint64_t>(), c);
        EXPECT_EQ(ex.seed_hex().size(), 64u);
    }
}

TEST(Compose, ExperimentThreadsIndependentMasters)
{
    dpf::experiment::master_seed s0{}, s1{};
    std::uint64_t v0 = 0, v1 = 0;
    std::thread t0([&] {
        dpf::experiment ex("t0", "p0");
        s0 = ex.seed();
        v0 = dpf::uniform_sample<std::uint64_t>();
    });
    std::thread t1([&] {
        dpf::experiment ex("t1", "p1");
        s1 = ex.seed();
        v1 = dpf::uniform_sample<std::uint64_t>();
    });
    t0.join();
    t1.join();
    EXPECT_NE(s0, s1);
    // Distinct masters almost surely yield distinct first words.
    EXPECT_NE(v0, v1);
}

TEST(Compose, MeasurePlanReportsRoundsAndSeed)
{
    auto plan = dpf::protocol::fss_point_plan(0);
    auto ex = dpf::app::measure_plan("fss_point", plan);
    EXPECT_EQ(ex.interactive_rounds(), 8u);
    EXPECT_EQ(ex.plan_bytes_out(), 128u);
    EXPECT_FALSE(ex.rounds().empty());
    EXPECT_GT(ex.rounds().front().bytes_out, 0u);
    EXPECT_EQ(ex.seed_hex().size(), 64u);
    EXPECT_GT(ex.wall_ns(), 0u);
    const std::string dir = "/tmp/libdpf_experiment_csv_test";
    #if defined(_WIN32)
    (void)dir;
    #else
    (void)::system(("rm -rf " + dir + " && mkdir -p " + dir).c_str());
    ex.write_csv(dir);
    std::ifstream summary(dir + "/summary.csv");
    ASSERT_TRUE(summary.good());
    std::string header;
    ASSERT_TRUE(static_cast<bool>(std::getline(summary, header)));
    EXPECT_NE(header.find("master_seed"), std::string::npos);
    std::string row;
    ASSERT_TRUE(static_cast<bool>(std::getline(summary, row)));
    EXPECT_NE(row.find("fss_point"), std::string::npos);
    #endif
}

}  // namespace