diff --git a/tests/testJobSystem.cpp b/tests/testJobSystem.cpp
index d61fea9..d8512a5 100644
--- a/tests/testJobSystem.cpp
+++ b/tests/testJobSystem.cpp
@@ -154,3 +154,162 @@ VF_TEST(graphNeverStartsNodeBeforeItsDependencies) {
     }
 }
 
+// A wide graph: one root, a fan-out layer, and a join node that must observe
+// every fan-out node as finished before it starts.
+VF_TEST(graphJoinNodeWaitsForTheWholeFanOut) {
+    for (unsigned threads : {0u, 2u, 6u}) {
+        JobSystem jobs(threads);
+        std::atomic<int> liveFanOut{0};
+        std::atomic<int> peakLiveFanOut{0};
+        std::atomic<int> doneFanOut{0};
+        std::atomic<int> joinSawLive{-1};
+        std::atomic<int> joinSawDone{-1};
+        std::atomic<int> rootDone{0};
+        std::atomic<int> fanOutSawRoot{0};
+
+        JobGraph graph;
+        std::size_t root = graph.addNode("root", [&] { rootDone.store(1); });
+        std::vector<std::size_t> fanOut;
+        for (int i = 0; i < 12; ++i) {
+            fanOut.push_back(graph.addNode("fan" + std::to_string(i), [&] {
+                fanOutSawRoot.fetch_add(rootDone.load());
+                int live = liveFanOut.fetch_add(1) + 1;
+                int seen = peakLiveFanOut.load();
+                while (live > seen && !peakLiveFanOut.compare_exchange_weak(seen, live)) {
+                }
+                liveFanOut.fetch_sub(1);
+                doneFanOut.fetch_add(1);
+            }));
+        }
+        std::size_t join = graph.addNode("join", [&] {
+            joinSawLive.store(liveFanOut.load());
+            joinSawDone.store(doneFanOut.load());
+        });
+        for (std::size_t node : fanOut) {
+            graph.addDependency(node, root);
+            graph.addDependency(join, node);
+        }
+
+        JobBatch batch = jobs.runGraph(graph);
+        batch.wait();
+        CHECK_EQ(joinSawLive.load(), 0);
+        CHECK_EQ(joinSawDone.load(), 12);      // the join really ran last
+        CHECK_EQ(fanOutSawRoot.load(), 12);    // and every fan-out node ran after the root
+        CHECK(peakLiveFanOut.load() >= 1);
+        if (threads == 0u) CHECK_EQ(peakLiveFanOut.load(), 1);
+        CHECK_EQ(batch.failedCount(), static_cast<std::size_t>(0));
+    }
+}
+
+VF_TEST(graphWithCycleIsRejectedInsteadOfHanging) {
+    for (unsigned threads : {0u, 4u}) {
+        JobSystem jobs(threads);
+        std::atomic<int> ran{0};
+        JobGraph graph;
+        for (int i = 0; i < 3; ++i)
+            graph.addNode("loop" + std::to_string(i), [&ran] { ran.fetch_add(1); });
+        graph.addDependency(0, 2);
+        graph.addDependency(1, 0);
+        graph.addDependency(2, 1);
+
+        CHECK(graph.hasCycle());
+        CHECK_EQ(graph.topologicalOrder().size(), static_cast<std::size_t>(0));
+        JobBatch batch = jobs.runGraph(graph);
+        CHECK(!batch.valid());
+        CHECK(batch.empty());
+        batch.wait();                 // returns at once, nothing was scheduled
+        CHECK_EQ(ran.load(), 0);
+    }
+}
+
+VF_TEST(selfDependentNodeCountsAsACycle) {
+    JobSystem jobs(2);
+    JobGraph graph;
+    std::atomic<int> ran{0};
+    graph.addNode("lonely", [&ran] { ran.fetch_add(1); });
+    graph.addDependency(0, 0);
+    CHECK(graph.hasCycle());
+    JobBatch batch = jobs.runGraph(graph);
+    CHECK(!batch.valid());
+    CHECK_EQ(ran.load(), 0);
+}
+
+VF_TEST(jobExceptionTravelsBackThroughItsHandle) {
+    for (unsigned threads : {0u, 3u}) {
+        JobSystem jobs(threads);
+        JobHandle bad = jobs.submit([] { throw std::runtime_error("chunk generator blew up"); });
+        bad.wait();
+        CHECK(bad.done());
+        CHECK(bad.failed());
+        CHECK(bad.errorMessage() == std::string("chunk generator blew up"));
+
+        bool rethrown = false;
+        try {
+            bad.rethrowError();
+        } catch (const std::runtime_error& ex) {
+            rethrown = std::string(ex.what()) == "chunk generator blew up";
+        }
+        CHECK(rethrown);
+
+        std::atomic<int> after{0};
+        JobBatch survivors;
+        for (int i = 0; i < 32; ++i)
+            survivors.add(jobs.submit([&after] { after.fetch_add(1); }));
+        survivors.wait();
+        CHECK_EQ(after.load(), 32);
+        CHECK_EQ(survivors.failedCount(), static_cast<std::size_t>(0));
+        CHECK_EQ(jobs.stats().failed, static_cast<std::uint64_t>(1));
+    }
+}
+
+VF_TEST(graphNodeWithFailedDependencyIsSkippedNotHung) {
+    for (unsigned threads : {0u, 4u}) {
+        JobSystem jobs(threads);
+        std::atomic<int> ranDependent{0};
+        std::atomic<int> ranIndependent{0};
+        JobGraph graph;
+        std::size_t broken = graph.addNode("broken", [] { throw std::runtime_error("stage lost"); });
+        std::size_t child = graph.addNode("child", [&] { ranDependent.fetch_add(1); });
+        std::size_t grandchild = graph.addNode("grandchild", [&] { ranDependent.fetch_add(1); });
+        std::size_t sibling = graph.addNode("sibling", [&] { ranIndependent.fetch_add(1); });
+        graph.addDependency(child, broken);
+        graph.addDependency(grandchild, child);
+
+        JobBatch batch = jobs.runGraph(graph);
+        CHECK(batch.valid());
+        batch.wait();                 // must not deadlock on the skipped nodes
+
+        CHECK(batch.at(broken).failed());
+        CHECK(batch.at(child).failed());
+        CHECK(batch.at(grandchild).failed());
+        CHECK(!batch.at(sibling).failed());
+        CHECK_EQ(ranDependent.load(), 0);
+        CHECK_EQ(ranIndependent.load(), 1);
+        CHECK_EQ(batch.failedCount(), static_cast<std::size_t>(3));
+        CHECK(batch.firstError() != nullptr);
+    }
+}
+
+VF_TEST(highPriorityJobsAreServedBeforeLowOnesOnOneWorker) {
+    JobSystem jobs(1);
+    Latch started;
+    Latch release;
+    std::mutex orderMutex;
+    std::vector<int> order;
+
+    JobHandle blocker = jobs.submit([&] {
+        started.open();
+        release.await();
+    });
+    started.await();                  // the single worker is now occupied
+
+    auto record = [&](int tag) {
+        return [&, tag] {
+            std::lock_guard<std::mutex> lock(orderMutex);
+            order.push_back(tag);
+        };
+    };
+    JobHandle low = jobs.submit(record(3), JobPriority::Low);
+    JobHandle normal = jobs.submit(record(2), JobPriority::Normal);
+    JobHandle high = jobs.submit(record(1), JobPriority::High);
+