yams/tests/integration/daemon/onnx_tuning_integration_test.cpp

// ONNX Concurrency Integration Tests
// Tests that ONNX pool and embedding loading respect TuneAdvisor and TuningManager
// settings throughout the full configuration flow:
//   Environment Variables → TuneAdvisor → TuningManager::configureOnnxConcurrencyRegistry()
//       → OnnxConcurrencyRegistry → Embedding/Reranker services respect limits

#include "../../common/env_compat.h"
#include "../../common/test_helpers_catch2.h"
#include "test_daemon_harness.h"
#include <catch2/catch_test_macros.hpp>

#include <yams/cli/cli_sync.h>
#include <yams/compat/unistd.h>
#include <yams/daemon/client/daemon_client.h>
#include <yams/daemon/components/ConfigResolver.h>
#include <yams/daemon/components/TuneAdvisor.h>
#include <yams/daemon/resource/OnnxConcurrencyRegistry.h>

#include <algorithm>
#include <atomic>
#include <chrono>
#include <cstdlib>
#include <filesystem>
#include <fstream>
#include <thread>
#include <vector>

using namespace yams::daemon;
using namespace yams::test;
using namespace std::chrono_literals;

namespace {

using yams::test::ScopedEnvVar;

constexpr auto kHarnessStartTimeout = 30s;

/// Sanitizer startup is instrumented heavily; registry values are asserted after startup using
/// each section's explicit configuration contract rather than live compatibility-environment reads.
bool startOnnxHarness(DaemonHarness& harness, std::chrono::milliseconds timeout) {
    return harness.start(timeout);
}

} // namespace

// =============================================================================
// Test 1: TuneAdvisor Settings Flow to OnnxConcurrencyRegistry
// =============================================================================

TEST_CASE("OnnxConcurrencyRegistry respects TuneAdvisor env settings",
          "[daemon][onnx][tuning][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    SECTION("YAMS_ONNX_MAX_CONCURRENT configures total slots") {
        // Set env vars before daemon starts (TuneAdvisor reads at startup)
        ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "6");
        ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced"); // scale=0.5

        DaemonHarness::Options opts;
        opts.useMockModelProvider = true;
        DaemonHarness harness(opts);
        REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

        REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

        auto& registry = OnnxConcurrencyRegistry::instance();
        // Balanced profile uses 0.5x scale, so 6 * 0.5 = 3.
        // But totalReserved = gliner(1) + embed(1) + reranker(1) = 3, and
        // scaledMax = max(3, 3+1) = 4 to ensure at least 1 shared slot.
        INFO("totalSlots=" << registry.totalSlots());
        CHECK(registry.totalSlots() == 4);

        harness.stop();
    }

    SECTION("Reserved slots are configured per lane") {
        ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "10");
        ScopedEnvVar embedEnv("YAMS_ONNX_EMBED_RESERVED", "3");
        ScopedEnvVar glinerEnv("YAMS_ONNX_GLINER_RESERVED", "2");
        ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

        DaemonHarness::Options opts;
        opts.useMockModelProvider = true;
        DaemonHarness harness(opts);
        REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

        REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

        auto& registry = OnnxConcurrencyRegistry::instance();
        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        auto glinerMetrics = registry.laneMetrics(OnnxLane::Gliner);

        INFO("embedMetrics.reserved=" << embedMetrics.reserved);
        INFO("glinerMetrics.reserved=" << glinerMetrics.reserved);
        CHECK(embedMetrics.reserved == 3);
        CHECK(glinerMetrics.reserved == 2);

        harness.stop();
    }

    SECTION("Reranker reserved slots are configured") {
        ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "8");
        ScopedEnvVar rerankerEnv("YAMS_ONNX_RERANKER_RESERVED", "2");
        ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

        DaemonHarness::Options opts;
        opts.useMockModelProvider = true;
        DaemonHarness harness(opts);
        REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

        REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

        auto& registry = OnnxConcurrencyRegistry::instance();
        auto rerankerMetrics = registry.laneMetrics(OnnxLane::Reranker);

        INFO("rerankerMetrics.reserved=" << rerankerMetrics.reserved);
        CHECK(rerankerMetrics.reserved == 2);

        harness.stop();
    }

    SECTION("Efficient profile scales down max slots") {
        ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "8");
        ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "efficient"); // scale=0.0

        DaemonHarness::Options opts;
        opts.useMockModelProvider = true;
        DaemonHarness harness(opts);
        REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

        REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

        auto& registry = OnnxConcurrencyRegistry::instance();
        // Efficient profile uses 0.0x scale, so 8 * 0.0 = 0
        // Minimum is 2, but totalReserved = gliner(1) + embed(1) + reranker(0, efficient) = 2,
        // so scaledMax = max(2, 2+1) = 3 to ensure at least 1 shared slot.
        INFO("totalSlots=" << registry.totalSlots());
        CHECK(registry.totalSlots() == 3);

        harness.stop();
    }

    SECTION("Aggressive profile scales up max slots") {
        ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "4");
        ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "aggressive"); // scale=1.0

        DaemonHarness::Options opts;
        opts.useMockModelProvider = true;
        DaemonHarness harness(opts);
        REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

        REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

        auto& registry = OnnxConcurrencyRegistry::instance();
        // Aggressive profile uses 1.0x scale, so 4 * 1.0 = 4
        INFO("totalSlots=" << registry.totalSlots());
        CHECK(registry.totalSlots() == 4);

        harness.stop();
    }
}

// =============================================================================
// Test 2: ONNX Registry Configuration with Embedding Slots
// =============================================================================

TEST_CASE("Embedding lane slots are configured correctly",
          "[daemon][onnx][embedding][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    // Configure slots with embedding-specific reservations
    ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "6");
    ScopedEnvVar embedEnv("YAMS_ONNX_EMBED_RESERVED", "2");
    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

    DaemonHarness::Options opts;
    opts.useMockModelProvider = true;
    opts.enableModelProvider = true;
    DaemonHarness harness(opts);
    REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

    REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

    auto& registry = OnnxConcurrencyRegistry::instance();

    SECTION("Registry is configured with correct total and embedding slots") {
        // Balanced profile uses 0.5x scale, so 6 * 0.5 = 3.
        // But totalSlots is clamped to at least totalReserved + 1 for shared capacity.
        // Defaults: gliner=1, reranker=1; with embed=2 => reserved total = 4.
        // scaledMax = max(3, 4+1) = 5.
        INFO("totalSlots=" << registry.totalSlots());
        CHECK(registry.totalSlots() == 5);

        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        INFO("embedding reserved=" << embedMetrics.reserved);
        CHECK(embedMetrics.reserved == 2);
    }

    SECTION("Slots can be manually acquired for embedding lane") {
        uint32_t beforeUsed = registry.usedSlots();

        // Acquire a slot for embedding lane
        OnnxConcurrencyRegistry::SlotGuard guard(registry, OnnxLane::Embedding, 1s);
        REQUIRE(guard.acquired());

        CHECK(registry.usedSlots() == beforeUsed + 1);

        // Metrics should show embedding lane is in use
        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        CHECK(embedMetrics.used >= 1);
    }

    SECTION("Multiple embedding slots can be acquired up to limit") {
        uint32_t beforeUsed = registry.usedSlots();

        // Acquire multiple slots for embedding
        OnnxConcurrencyRegistry::SlotGuard slot1(registry, OnnxLane::Embedding, 1s);
        OnnxConcurrencyRegistry::SlotGuard slot2(registry, OnnxLane::Embedding, 1s);

        REQUIRE(slot1.acquired());
        REQUIRE(slot2.acquired());
        CHECK(registry.usedSlots() >= beforeUsed + 2);
    }

    SECTION("Slots are released when guards go out of scope") {
        uint32_t beforeUsed = registry.usedSlots();

        {
            OnnxConcurrencyRegistry::SlotGuard guard(registry, OnnxLane::Embedding, 1s);
            REQUIRE(guard.acquired());
            CHECK(registry.usedSlots() == beforeUsed + 1);
        }

        // Slot should be released after scope exit
        CHECK(registry.usedSlots() == beforeUsed);
    }

    harness.stop();
}

// =============================================================================
// Test 3: Metrics Observability
// =============================================================================

TEST_CASE("FSM metrics reflect ONNX configuration", "[daemon][onnx][metrics][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "5");
    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

    DaemonHarness::Options opts;
    opts.useMockModelProvider = true;
    DaemonHarness harness(opts);
    REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

    REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

    SECTION("Registry snapshot contains expected metrics") {
        auto& registry = OnnxConcurrencyRegistry::instance();
        auto snap = registry.snapshot();

        // Balanced profile uses 0.5x scale, so 5 * 0.5 = 2.5 -> 2.
        // But totalReserved = gliner(1) + embed(1) + reranker(1) = 3,
        // so scaledMax = max(2, 3+1) = 4 to ensure shared capacity.
        INFO("snapshot totalSlots=" << snap.totalSlots);
        INFO("snapshot usedSlots=" << snap.usedSlots);
        INFO("snapshot availableSlots=" << snap.availableSlots);

        CHECK(snap.totalSlots == 4);
        CHECK(snap.usedSlots == 0);
        CHECK(snap.availableSlots == 4);
    }

    SECTION("Status response includes ONNX metrics via client") {
        ClientConfig clientCfg;
        clientCfg.socketPath = harness.socketPath();
        clientCfg.requestTimeout = 10s;
        clientCfg.autoStart = false;
        DaemonClient client(clientCfg);
        auto connected = yams::cli::run_sync(client.connect(), 5s);
        REQUIRE(connected.has_value());

        auto statusResult = yams::cli::run_sync(client.status(), 10s);
        REQUIRE(statusResult.has_value());

        // Status response should be valid and daemon ready
        const auto& status = statusResult.value();
        INFO("overallStatus=" << status.overallStatus);
        INFO("lifecycleState=" << status.lifecycleState);
        INFO("onnxTotalSlots=" << status.onnxTotalSlots);
        INFO("onnxUsedSlots=" << status.onnxUsedSlots);
        CHECK(status.running);
        CHECK((status.onnxTotalSlots == 0 || status.onnxTotalSlots >= 4));
        CHECK(status.onnxTotalSlots >= status.onnxUsedSlots);
        CHECK((status.ready || status.overallStatus == "ready" || status.overallStatus == "Ready" ||
               status.overallStatus == "degraded" || status.overallStatus == "Degraded"));
    }

    harness.stop();
}

// =============================================================================
// Test 4: Registry State Isolation Between Daemon Restarts
// =============================================================================

TEST_CASE("OnnxConcurrencyRegistry state resets on daemon restart",
          "[daemon][onnx][lifecycle][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    SECTION("Different configs apply correctly on daemon restart") {
        // First daemon with 4 slots
        {
            ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "4");
            ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

            DaemonHarness::Options opts;
            opts.useMockModelProvider = true;
            DaemonHarness harness(opts);
            REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));
            REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

            auto& registry = OnnxConcurrencyRegistry::instance();
            // Balanced profile uses 0.5x scale, so 4 * 0.5 = 2.
            // But totalReserved = 3, so scaledMax = max(2, 3+1) = 4.
            INFO("first daemon totalSlots=" << registry.totalSlots());
            CHECK(registry.totalSlots() == 4);

            harness.stop();
        }

        // Second daemon with 8 slots - should reconfigure
        {
            ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "8");
            ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

            DaemonHarness::Options opts;
            opts.useMockModelProvider = true;
            DaemonHarness harness(opts);
            REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));
            REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

            auto& registry = OnnxConcurrencyRegistry::instance();
            // Balanced profile uses 0.5x scale, so 8 * 0.5 = 4
            INFO("second daemon totalSlots=" << registry.totalSlots());
            CHECK(registry.totalSlots() == 4);

            harness.stop();
        }
    }
}

// =============================================================================
// Test 5: Lane Metrics Update Correctly
// =============================================================================

TEST_CASE("Lane metrics update during slot acquisition", "[daemon][onnx][lanes][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "8");
    ScopedEnvVar embedEnv("YAMS_ONNX_EMBED_RESERVED", "2");
    ScopedEnvVar glinerEnv("YAMS_ONNX_GLINER_RESERVED", "1");
    ScopedEnvVar rerankerEnv("YAMS_ONNX_RERANKER_RESERVED", "1");
    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

    DaemonHarness::Options opts;
    opts.useMockModelProvider = true;
    DaemonHarness harness(opts);
    REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

    REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

    auto& registry = OnnxConcurrencyRegistry::instance();

    SECTION("All lanes have correct reserved slots") {
        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        auto glinerMetrics = registry.laneMetrics(OnnxLane::Gliner);
        auto rerankerMetrics = registry.laneMetrics(OnnxLane::Reranker);
        auto otherMetrics = registry.laneMetrics(OnnxLane::Other);

        CHECK(embedMetrics.reserved == 2);
        CHECK(glinerMetrics.reserved == 1);
        CHECK(rerankerMetrics.reserved == 1);
        CHECK(otherMetrics.reserved == 0);
    }

    SECTION("Manual slot acquisition updates used count") {
        uint32_t beforeUsed = registry.usedSlots();

        // Acquire a slot manually
        OnnxConcurrencyRegistry::SlotGuard guard(registry, OnnxLane::Embedding, 1s);
        REQUIRE(guard.acquired());

        CHECK(registry.usedSlots() == beforeUsed + 1);

        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        CHECK(embedMetrics.used >= 1);
    }

    harness.stop();
}

// =============================================================================
// Test 6: AbiModelProviderAdapter SlotGuard Verification (Real ONNX Plugin)
// =============================================================================

TEST_CASE("AbiModelProviderAdapter acquires SlotGuard during real embedding",
          "[daemon][onnx][adapter][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();
    if (ConfigResolver::envTruthy(std::getenv("YAMS_DISABLE_VECTORS")) ||
        ConfigResolver::envTruthy(std::getenv("YAMS_SQLITE_VEC_SKIP_INIT"))) {
        SKIP("Real ONNX adapter test requires vector support enabled");
    }

    // Configure conservative slot limits to observe contention
    ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "4");
    ScopedEnvVar embedEnv("YAMS_ONNX_EMBED_RESERVED", "1");
    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

    DaemonHarness::Options opts;
    opts.useMockModelProvider = false; // CRITICAL: Use real ONNX plugin
    opts.enableModelProvider = true;
    opts.autoLoadPlugins = true; // Load ONNX plugin
    opts.configureModelPool = true;
    opts.modelPoolLazyLoading = false; // Preload model on startup
    opts.preloadModels = {"all-MiniLM-L6-v2"};

    // Set plugin directory to builddir/plugins
    opts.pluginDir = std::filesystem::current_path() / "builddir" / "plugins";

    // Configure ONNX plugin with preferred model (uses plugin's model resolution)
    opts.pluginConfigs["onnx_plugin"] = R"({"preferred_model": "all-MiniLM-L6-v2"})";

    DaemonHarness harness(opts);

    INFO("Starting daemon with real ONNX plugin (model preloading may take time)...");
    // If model not found, start() will fail - this makes it clear the model must be installed
    if (!startOnnxHarness(harness, std::chrono::seconds(60))) {
        SKIP("Failed to start daemon with ONNX plugin - model may not be available");
    }

    REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

    ClientConfig clientCfg;
    clientCfg.socketPath = harness.socketPath();
    clientCfg.requestTimeout = 30s;
    clientCfg.autoStart = false;
    DaemonClient client(clientCfg);
    auto connected = yams::cli::run_sync(client.connect(), 10s);
    REQUIRE(connected.has_value());

    auto& registry = OnnxConcurrencyRegistry::instance();

    SECTION("Single embedding acquires and releases slot") {
        uint32_t baselineUsed = registry.usedSlots();
        auto beforeSnap = registry.snapshot();

        GenerateEmbeddingRequest req;
        req.text = "test embedding via real ONNX plugin";
        req.modelName = "all-MiniLM-L6-v2";
        req.normalize = true;

        INFO("Issuing embedding request...");
        auto result = yams::cli::run_sync(client.generateEmbedding(req), 30s);

        if (!result.has_value()) {
            INFO("Embedding request failed: " << result.error().message);
            // Skip rather than fail - model may not be fully loaded
            SKIP("Embedding request failed - model may not be available");
        }

        REQUIRE(!result.value().embedding.empty());
        INFO("Embedding dimension: " << result.value().embedding.size());

        // Verify slot released after completion
        auto afterSnap = registry.snapshot();
        INFO("Slots before: " << beforeSnap.usedSlots << ", after: " << afterSnap.usedSlots);
        CHECK(afterSnap.usedSlots == baselineUsed);

        // Verify no timeouts occurred during this request
        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        INFO("Embedding lane timeouts: " << embedMetrics.timeouts);
        CHECK(embedMetrics.timeouts == 0);
    }

    SECTION("Concurrent embeddings respect slot limits") {
        // WARMUP: Run a single embedding first to ensure ONNX session is fully initialized
        // This prevents the "Session invalid during warmup" race condition that occurs
        // when multiple threads attempt to use the session before warmup completes.
        INFO("Warmup: Running initial embedding to initialize ONNX session...");
        {
            GenerateEmbeddingRequest warmupReq;
            warmupReq.text = "warmup request to initialize session";
            warmupReq.modelName = "all-MiniLM-L6-v2";
            warmupReq.normalize = true;

            auto warmupResult = yams::cli::run_sync(client.generateEmbedding(warmupReq), 60s);
            if (!warmupResult.has_value() || warmupResult.value().embedding.empty()) {
                INFO("Warmup embedding failed - ONNX session may not be ready");
                SKIP("ONNX session warmup failed - skipping concurrent test");
            }
            INFO("Warmup complete, session initialized");
        }

        // Brief pause to ensure session is stable after warmup
        std::this_thread::sleep_for(100ms);

        std::atomic<int> completed{0};
        std::atomic<int> failed{0};
        std::atomic<uint32_t> peakUsed{0};

        // Track peak slot usage during concurrent requests
        auto worker = [&](int id) {
            GenerateEmbeddingRequest req;
            req.text = "concurrent test embedding " + std::to_string(id);
            req.modelName = "all-MiniLM-L6-v2";
            req.normalize = true;

            // Capture peak usage during execution
            uint32_t current = registry.usedSlots();
            uint32_t expected = peakUsed.load();
            while (current > expected && !peakUsed.compare_exchange_weak(expected, current)) {
            }

            auto result = yams::cli::run_sync(client.generateEmbedding(req), 60s);
            if (result.has_value() && !result.value().embedding.empty()) {
                ++completed;
            } else {
                ++failed;
            }
        };

        constexpr int numRequests = 4; // Match max slots (4) for more reliable test
        std::vector<std::thread> threads;
        threads.reserve(numRequests);

        INFO("Launching " << numRequests << " concurrent embedding requests...");
        for (int i = 0; i < numRequests; ++i) {
            threads.emplace_back(worker, i);
        }

        for (auto& t : threads) {
            t.join();
        }

        INFO("Completed: " << completed << ", Failed: " << failed);
        INFO("Peak slots used: " << peakUsed);

        // At least some should complete (with warmup, most should succeed)
        CHECK(completed >= 1);

        // After all complete, slots should be released
        CHECK(registry.usedSlots() == 0);

        // Peak usage should not exceed configured max (4)
        CHECK(peakUsed <= 4);
    }

    harness.stop();
}

TEST_CASE("Config selecting onnx backend uses ONNX provider for embeddings",
          "[daemon][onnx][config][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");
    ScopedEnvVar forceCpuEnv("YAMS_ONNX_FORCE_CPU", "1");

    const auto now = std::chrono::steady_clock::now().time_since_epoch().count();
    const auto configRoot =
        std::filesystem::temp_directory_path() /
        ("yams_onnx_backend_config_" + std::to_string(::getpid()) + "_" + std::to_string(now));
    std::filesystem::create_directories(configRoot);
    const auto configPath = configRoot / "config.toml";
    {
        std::ofstream out(configPath, std::ios::trunc);
        REQUIRE(out.is_open());
        out << "[version]\n";
        out << "config_version = 3\n";
        out << "\n[embeddings]\n";
        out << "backend = \"onnx\"\n";
        out << "preferred_model = \"all-MiniLM-L6-v2\"\n";
    }

    DaemonHarness::Options opts;
    opts.useMockModelProvider = false;
    opts.enableModelProvider = true;
    opts.autoLoadPlugins = true;
    opts.configureModelPool = true;
    opts.modelPoolLazyLoading = false;
    opts.preloadModels = {"all-MiniLM-L6-v2"};
    opts.pluginDir = std::filesystem::current_path() / "builddir" / "plugins";
    opts.pluginConfigs["onnx_plugin"] = R"({"preferred_model": "all-MiniLM-L6-v2"})";
    opts.configPath = configPath;

    DaemonHarness harness(opts);
    if (!startOnnxHarness(harness, std::chrono::seconds(60))) {
        std::error_code ec;
        std::filesystem::remove_all(configRoot, ec);
        SKIP(
            "Failed to start daemon with ONNX backend config - plugin or model may be unavailable");
    }

    ClientConfig clientCfg;
    clientCfg.socketPath = harness.socketPath();
    clientCfg.requestTimeout = 30s;
    clientCfg.autoStart = false;
    DaemonClient client(clientCfg);
    auto connected = yams::cli::run_sync(client.connect(), 10s);
    REQUIRE(connected.has_value());

    GenerateEmbeddingRequest req;
    req.text = "config-selected onnx embedding";
    req.modelName.clear();
    req.normalize = true;

    auto embedResult = yams::cli::run_sync(client.generateEmbedding(req), 30s);
    if (!embedResult.has_value()) {
        INFO("Embedding request failed: " << embedResult.error().message);
        harness.stop();
        std::error_code ec;
        std::filesystem::remove_all(configRoot, ec);
        SKIP("Embedding request failed - ONNX provider or model may be unavailable");
    }
    CHECK_FALSE(embedResult.value().embedding.empty());
    CHECK(embedResult.value().embedding.size() > 1);

    harness.stop();
    std::error_code ec;
    std::filesystem::remove_all(configRoot, ec);
}

// =============================================================================
// Test 7: Slot Exhaustion and Timeout Behavior
// =============================================================================

TEST_CASE("ONNX slot exhaustion causes graceful timeout", "[daemon][onnx][timeout][integration]") {
    SKIP_DAEMON_TEST_ON_WINDOWS();

    // Configure very limited slots AND zero reserved slots to force contention
    // Without zeroing reserved slots, lanes can acquire beyond totalSlots via reserved pool
    ScopedEnvVar maxEnv("YAMS_ONNX_MAX_CONCURRENT", "3");
    ScopedEnvVar embedEnv("YAMS_ONNX_EMBED_RESERVED", "0");
    ScopedEnvVar glinerEnv("YAMS_ONNX_GLINER_RESERVED", "0");
    ScopedEnvVar rerankerEnv("YAMS_ONNX_RERANKER_RESERVED", "0");
    ScopedEnvVar profileEnv("YAMS_TUNING_PROFILE", "balanced");

    DaemonHarness::Options opts;
    opts.useMockModelProvider = true;
    opts.enableModelProvider = true;
    DaemonHarness harness(opts);
    REQUIRE(startOnnxHarness(harness, kHarnessStartTimeout));

    REQUIRE(OnnxConcurrencyRegistry::instance().totalSlots() >= 2);

    auto& registry = OnnxConcurrencyRegistry::instance();

    // Verify registry is configured as expected
    INFO("Registry totalSlots: " << registry.totalSlots());
    INFO("Embedding reserved: " << registry.laneMetrics(OnnxLane::Embedding).reserved);
    INFO("Gliner reserved: " << registry.laneMetrics(OnnxLane::Gliner).reserved);
    INFO("Reranker reserved: " << registry.laneMetrics(OnnxLane::Reranker).reserved);
    REQUIRE(registry.totalSlots() >= 2);

    SECTION("Slot acquisition is tracked in metrics") {
        // This test verifies that slot acquisition/release is properly tracked,
        // without making assumptions about exact capacity limits (which can vary
        // based on semaphore implementation details).
        uint32_t beforeUsed = registry.usedSlots();
        INFO("Slots used before: " << beforeUsed);

        // Acquire a slot
        OnnxConcurrencyRegistry::SlotGuard guard(registry, OnnxLane::Other, 1s);
        REQUIRE(guard.acquired());

        // Verify usage increased
        CHECK(registry.usedSlots() == beforeUsed + 1);

        // Lane metrics should reflect the acquisition
        auto otherMetrics = registry.laneMetrics(OnnxLane::Other);
        INFO("Other lane used: " << otherMetrics.used);
        CHECK(otherMetrics.used >= 1);
    }

    SECTION("Slot metrics track reserved slots correctly") {
        // Configure reserved slot for embedding
        uint32_t originalReserved = registry.laneMetrics(OnnxLane::Embedding).reserved;
        registry.setReservedSlots(OnnxLane::Embedding, 1);

        auto embedMetrics = registry.laneMetrics(OnnxLane::Embedding);
        INFO("Embedding reserved slots: " << embedMetrics.reserved);
        CHECK(embedMetrics.reserved == 1);

        // Restore original
        registry.setReservedSlots(OnnxLane::Embedding, originalReserved);
    }

    harness.stop();
}