fix(threadpool): batches run every task once, and Stop ends the whole pool

- Enqueue(vector) dropped tasks when the count didn't divide by the
  workers: the last one got only the remainder. Tasks are now split as
  evenly as the count allows, each worker's taskCount grows by what it
  was given (the single-task Enqueue picks by it), and Start(1), which
  has no workers, leaves the batch for the caller's JoinPool.
- Stop wakes every slot, JoinPool's too, after joining the workers: a
  thread that had joined the pool slept on. It takes each slot's mutex
  before notifying, so a waiter between its check and its sleep can't miss
  the wake-up, and works on the pool it started with. Again, or before
  Start, it does nothing; joining a thread twice threw. Its documentation
  says what it doesn't do: stop from a pool task, or from two threads at
  once, run tasks enqueued once it has begun, or have JoinPool wait for it.
- Tests: EnqueueBatch, StopJoinPool.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/EnqueueBatch/main.cpp

@@ -0,0 +1,77 @@
+//SPDX-License-Identifier: LGPL-3.0-only
+//SPDX-FileCopyrightText: Copyright (C) 2026 Catcrafts®
+
+// ThreadPool::Enqueue(vector) runs every task exactly once, whatever the
+// task count. When the count didn't divide by the number of worker threads,
+// the last worker got only the remainder and tasksPerThread tasks were
+// dropped. The counts here cover fewer tasks than workers, an exact multiple,
+// one more and one less than a multiple, and a large uneven batch, then a
+// batch of single Enqueue calls, which rely on the per-thread task counts the
+// batch path also kept wrong.
+
+#include <cstdlib>
+import Crafter.Thread;
+import std;
+using namespace Crafter;
+
+namespace {
+
+int failures = 0;
+
+void Check(bool ok, std::string_view what) {
+    std::println("{}  {}", ok ? "PASS" : "FAIL", what);
+    if (!ok) ++failures;
+}
+
+// Waits until `done` reaches `expected` or a few seconds pass.
+bool WaitFor(const std::atomic<std::uint32_t>& done, std::uint32_t expected) {
+    const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(5);
+    while (done.load() < expected) {
+        if (std::chrono::steady_clock::now() > deadline) return false;
+        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+    return true;
+}
+
+void RunBatch(std::uint32_t count) {
+    std::atomic<std::uint32_t> done = 0;
+    std::vector<std::atomic<std::uint32_t>> runs(count);
+    std::vector<std::function<void(void)>> tasks;
+    tasks.reserve(count);
+    for (std::uint32_t i = 0; i < count; ++i) {
+        tasks.emplace_back([&runs, &done, i] {
+            runs[i]++;
+            done++;
+        });
+    }
+    ThreadPool::Enqueue(std::move(tasks));
+    const bool finished = WaitFor(done, count);
+    // A little longer, in case something runs twice.
+    std::this_thread::sleep_for(std::chrono::milliseconds(20));
+    const bool once = std::ranges::all_of(runs, [](const std::atomic<std::uint32_t>& r) { return r.load() == 1; });
+    Check(finished && once, std::format("a batch of {} tasks runs every task exactly once ({} of {} done)", count, done.load(), count));
+}
+
+} // namespace
+
+int main() {
+    ThreadPool::Start(4);
+    const std::uint32_t workers = ThreadPool::activeThreads;
+    std::println("      {} worker threads", workers);
+
+    for (std::uint32_t count : {1u, workers - 1, workers, workers + 1, 3 * workers + 2, 1000u}) {
+        RunBatch(count);
+    }
+
+    // Single Enqueue calls pick the least loaded worker by its task count;
+    // a count left wrong by a batch would starve or flood one of them.
+    std::atomic<std::uint32_t> done = 0;
+    for (std::uint32_t i = 0; i < 200; ++i) {
+        ThreadPool::Enqueue([&done] { done++; });
+    }
+    Check(WaitFor(done, 200), std::format("200 single tasks after the batches all run ({} done)", done.load()));
+
+    ThreadPool::Stop();
+    std::println("{}", failures == 0 ? "all checks passed" : "check(s) failed");
+    return failures == 0 ? EXIT_SUCCESS : EXIT_FAILURE;
+}

tests/StopJoinPool/main.cpp

@@ -0,0 +1,43 @@
+//SPDX-License-Identifier: LGPL-3.0-only
+//SPDX-FileCopyrightText: Copyright (C) 2026 Catcrafts®
+
+// ThreadPool::Stop ends the pool for everyone in it, the thread that joined
+// it with JoinPool too. Stop woke and joined the worker threads only, so a
+// JoinPool caller slept on, never told. And Stop called again joined threads
+// already joined, which throws.
+//
+// Here another thread stops the pool this one has joined, then stops it
+// again. A watchdog fails the test if JoinPool doesn't return.
+
+#include <cstdlib>
+#include <unistd.h>
+
+import Crafter.Thread;
+import std;
+
+using namespace Crafter;
+using namespace std::chrono_literals;
+
+int main() {
+    std::thread watchdog([] {
+        std::this_thread::sleep_for(10s);
+        std::println("FAIL  JoinPool didn't return within 10 s of Stop");
+        _exit(EXIT_FAILURE);
+    });
+    watchdog.detach();
+
+    ThreadPool::Start(4);
+    std::atomic<int> ran = 0;
+    for (int i = 0; i < 100; ++i) ThreadPool::Enqueue([&ran] { ++ran; });
+    std::thread stopper([] {
+        std::this_thread::sleep_for(100ms);
+        ThreadPool::Stop();
+    });
+    ThreadPool::JoinPool();
+    stopper.join();
+    std::println("{}  Stop from another thread ends the JoinPool of the thread that joined, after the queued tasks ({} of 100)",
+                 ran == 100 ? "PASS" : "FAIL", ran.load());
+    ThreadPool::Stop();
+    std::println("PASS  and stopping it again does nothing");
+    return ran == 100 ? EXIT_SUCCESS : EXIT_FAILURE;
+}