*
   * <ul>
   *   <li>All NaNs are fuzzily equal.
   *   <li>If {@code a == b}, then {@code a} and {@code b} are always fuzzily equal.
   *   <li>Positive and negative zero are always fuzzily equal.
   *   <li>If {@code tolerance} is zero, and neither {@code a} nor {@code b} is NaN, then {@code a}
   *       and {@code b} are fuzzily equal if and only if {@code a == b}.

 * map from keys to <i>nonempty</i> collections of values:
 *
 * <ul>
 *   <li>a → 1, 2
 *   <li>b → 3
 * </ul>
 *
 * ... or as a single "flattened" collection of key-value pairs:
 *
 * <ul>
 *   <li>a → 1
 *   <li>a → 2
 *   <li>b → 3
 * </ul>
 *
 * <p><b>Important:</b> although the first interpretation resembles how most multimaps are

 *   <li>{@code expireAfterWrite=[duration]}: sets {@link CacheBuilder#expireAfterWrite}.
 *   <li>{@code refreshAfterWrite=[duration]}: sets {@link CacheBuilder#refreshAfterWrite}.
 *   <li>{@code weakKeys}: sets {@link CacheBuilder#weakKeys}.
 *   <li>{@code softValues}: sets {@link CacheBuilder#softValues}.
 *   <li>{@code weakValues}: sets {@link CacheBuilder#weakValues}.
 *   <li>{@code recordStats}: sets {@link CacheBuilder#recordStats}.
 * </ul>

   *   <li>If there is any non-private constructor or non-private static factory method declared by
   *       {@code cls}:
   *       <ul>
   *         <li>Test will fail if default value for a parameter cannot be determined.
   *         <li>Test will fail if the factory method returns null so testing instance methods is
   *             impossible.

     * <ul>
     *   <li>If it starts with either "pom." or "project.", the expression is evaluated against the model.</li>
     *   <li>If the value is null, get the value from the context.</li>
     *   <li>If the value is null, but the context contains the expression, don't replace the expression string
     *       with the value, and continue to find other expressions.</li>

 * terms</a>):
 *
 * <ul>
 *   <li>directed graphs
 *   <li>undirected graphs
 *   <li>graphs that do/don't allow parallel edges
 *   <li>graphs that do/don't allow self-loops
 *   <li>graphs whose nodes/edges are insertion-ordered, sorted, or unordered
 *   <li>graphs whose edges are unique objects
 * </ul>
 *
 * <h3>Building a {@code Network}</h3>
 *

 *
 * <p>Unpacking this definition:
 *
 * <ul>
 *   <li><b>block of data:</b> the input for a hash function is always, in concept, an ordered byte
 *       array. This hashing API accepts an arbitrary sequence of byte and multibyte values (via
 *       {@link Hasher}), but this is merely a convenience; these are always translated into raw
 *       byte sequences under the covers.

 * href="https://github.com/google/guava/wiki/GraphsExplained#definitions">definitions of
 * terms</a>):
 *
 * <ul>
 *   <li>directed graphs
 *   <li>undirected graphs
 *   <li>graphs that do/don't allow self-loops
 *   <li>graphs whose nodes/edges are insertion-ordered, sorted, or unordered
 * </ul>
 *
 * <p>{@code Graph} explicitly does not support parallel edges, and forbids implementations or

     * <ul>
     * <li>collect all the vertices for the projects that we want to build.</li>
     * <li>iterate through the deps of each project and if that dep is within
     * the set of projects we want to build then add an edge, otherwise throw
     * the edge away because that dependency is not within the set of projects
     * we are trying to build. we assume a closed set.</li>
     * <li>do a topo sort on the graph that remains.</li>

 *
 * <p>Advantages compared to {@code int[]}:
 *
 * <ul>
 *   <li>All the many well-known advantages of immutability (read <i>Effective Java</i>, third
 *       edition, Item 17).
 *   <li>Has the value-based (not identity-based) {@link #equals}, {@link #hashCode}, and {@link
 *       #toString} behavior you expect
 *   <li>Offers useful operations beyond just {@code get} and {@code length}, so you don't have to