* Visited Nodes = {N1, N2, N3, N4}
   * </pre>
   */
  private static final class Undirected<N> extends EndpointPairIterator<N> {
    // It's a little weird that we add `null` to this set, but it makes for slightly simpler code.
    private @Nullable Set<@Nullable N> visitedNodes;

    private Undirected(BaseGraph<N> graph) {
      super(graph);
      this.visitedNodes = Sets.newHashSetWithExpectedSize(graph.nodes().size() + 1);
    }

                    + jdk.getHash()
                    + "/"
                    + jdk.getBuild()
                    + "/GPL/openjdk-[revision]_[module]-[classifier]_bin.[ext]";
            } else {
                // simpler legacy pattern from JDK 9 to JDK 12 that we are advocating to Oracle to bring back
                artifactPattern = "java/GA/jdk"
                    + jdk.getMajor()
                    + "/"

This new `examples` field in JSON Schema is **just a `list`** of examples, not a dict with extra metadata as in the other places in OpenAPI (described above).

/// info

Even after OpenAPI 3.1.0 was released with this new simpler integration with JSON Schema, for a while, Swagger UI, the tool that provides the automatic docs, didn't support OpenAPI 3.1.0 (it does since version 5.0.0 🎉).

     * initial thread's.
     */
    Set<Throwable> seenExceptionsLocal = seenExceptionsField;
    if (seenExceptionsLocal == null) {
      // TODO(cpovirk): Should we use a simpler (presumably cheaper) data structure?
      /*
       * Using weak references here could let us release exceptions earlier, but:
       *

When you check the benchmarks, it is common to see several tools of different types compared as equivalent.

Specifically, to see Uvicorn, Starlette and FastAPI compared together (among many other tools).

The simpler the problem solved by the tool, the better performance it will get. And most of the benchmarks don't test the additional features provided by the tool.

The hierarchy is like:

* **Uvicorn**: an ASGI server

   *
   * <p>If all input data to this {@code TopKSelector} is in a single {@code Iterable}, prefer
   * {@link Ordering#leastOf(Iterable, int)}, which provides a simpler API for that use case.
   */
  public void offerAll(Iterable<? extends T> elements) {
    offerAll(elements.iterator());
  }

  /**

   *
   * - https://github.com/jspecify/jdk/commit/71d826792b8c7ef95d492c50a274deab938f2552
   */
  /*
   * TODO(cpovirk): Is the unchecked cast avoidable? Would System.arraycopy be similarly fast (if
   * likewise not type-checked)? Could our single caller do something different?
   */
  @SuppressWarnings({"nullness", "unchecked"})

   *
   * - https://github.com/jspecify/jdk/commit/71d826792b8c7ef95d492c50a274deab938f2552
   */
  /*
   * TODO(cpovirk): Is the unchecked cast avoidable? Would System.arraycopy be similarly fast (if
   * likewise not type-checked)? Could our single caller do something different?
   */
  @SuppressWarnings({"nullness", "unchecked"})

Le `TestClient` est basé sur [HTTPX](https://www.python-httpx.org) et, heureusement, nous pouvons l'utiliser directement pour tester l'API.

## Exemple { #example }

Pour un exemple simple, considérons une structure de fichiers similaire à celle décrite dans [Applications plus grandes](../tutorial/bigger-applications.md) et [Tests](../tutorial/testing.md) :

```
.
├── app
│   ├── __init__.py
│   ├── main.py
│   └── test_main.py
```

  /*
   * TODO: consider adding MultisetFeatures.SUPPORTS_SET_COUNT. Currently we
   * assume that using setCount() to increase the count is permitted iff add()
   * is permitted and similarly for decrease/remove(). We assume that a
   * setCount() no-op is permitted if either add() or remove() is permitted,
   * though we also allow it to "succeed" if neither is permitted.
   */