zeitwerk 2.6.8 → 2.6.18

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 418f16f6224359c716a990f72ae916c980a3f388b09ab018c0972b8558530620
-  data.tar.gz: '04847c6c8f8f894fe3c2b3c39d175b5854b583d3a02ef5cca27a53cb1b8dafd4'
+  metadata.gz: a1fa60f08c282d8471eaf321deba97207ce56d2cba32619cff78f09e65f2f258
+  data.tar.gz: 1c803846848fc8830913acfd8fbf04bfa59f26c3c8806d4fa5bed9c751f48c5a
 SHA512:
-  metadata.gz: f86272e84721cf9b8b1e07182b1b60cfe732ef2c266dbe076225d32af52fc9a30c278a14651ddbb3d435636646dd23c948d3fe0f00ed9b8393abdeb34dd40028
-  data.tar.gz: 562738b53779310e899c2065c7046844d27daabb5f354972e4c9b65bec3e5c0cdade92d19ed815fe7b0dc01538603daa3e4dd2f0049fe9a7a2e74313c0f040f6
+  metadata.gz: abb7134976f1ae00cafd77dcaa39fe2239632639c136d866ac75e55c8ec15f1c157ff0e434969dade286ee36e6422856799acd3e86dcd61e4ef39d1fd23909a4
+  data.tar.gz: 42e2309bd69f0bba3bfc74acecf2fcf3b3206a6cd26583afdc1012d34a063431bc03ec4ec42e1cf452f41bfa0de43a4d171cb2e99c7ca3e64b61412fbef83819

data/README.md

@@ -3,7 +3,8 @@
 
 
 [![Gem Version](https://img.shields.io/gem/v/zeitwerk.svg?style=for-the-badge)](https://rubygems.org/gems/zeitwerk)
-[![Build Status](https://img.shields.io/github/actions/workflow/status/fxn/zeitwerk/ci.yml?branch=main&event=push&style=for-the-badge)](https://github.com/fxn/zeitwerk/actions/workflows/ci.yml?query=branch%3main)
+[![Build Status](https://img.shields.io/github/actions/workflow/status/fxn/zeitwerk/ci.yml?branch=main&event=push&style=for-the-badge)](https://github.com/fxn/zeitwerk/actions/workflows/ci.yml?query=branch%3Amain)
+
 
 <!-- TOC -->
 
@@ -39,6 +40,7 @@
   - [Inflection](#inflection)
     - [Zeitwerk::Inflector](#zeitwerkinflector)
     - [Zeitwerk::GemInflector](#zeitwerkgeminflector)
+    - [Zeitwerk::NullInflector](#zeitwerknullinflector)
     - [Custom inflector](#custom-inflector)
   - [Callbacks](#callbacks)
     - [The on_setup callback](#the-on_setup-callback)
@@ -56,6 +58,9 @@
   - [Beware of circular dependencies](#beware-of-circular-dependencies)
   - [Reopening third-party namespaces](#reopening-third-party-namespaces)
   - [Introspection](#introspection)
+    - [`Zeitwerk::Loader#dirs`](#zeitwerkloaderdirs)
+    - [`Zeitwerk::Loader#cpath_expected_at`](#zeitwerkloadercpath_expected_at)
+    - [`Zeitwerk::Loader#all_expected_cpaths`](#zeitwerkloaderall_expected_cpaths)
   - [Encodings](#encodings)
   - [Rules of thumb](#rules-of-thumb)
   - [Debuggers](#debuggers)
@@ -76,15 +81,15 @@
 
 Zeitwerk is an efficient and thread-safe code loader for Ruby.
 
-Given a [conventional file structure](#file-structure), Zeitwerk is able to load your project's classes and modules on demand (autoloading), or upfront (eager loading). You don't need to write `require` calls for your own files, rather, you can streamline your programming knowing that your classes and modules are available everywhere. This feature is efficient, thread-safe, and matches Ruby's semantics for constants.
+Given a [conventional file structure](#file-structure), Zeitwerk is capable of loading your project's classes and modules on demand (autoloading) or upfront (eager loading). You don't need to write `require` calls for your own files; instead, you can streamline your programming by knowing that your classes and modules are available everywhere. This feature is efficient, thread-safe, and aligns with Ruby's semantics for constants.
 
-Zeitwerk is also able to reload code, which may be handy while developing web applications. Coordination is needed to reload in a thread-safe manner. The documentation below explains how to do this.
+Zeitwerk also supports code reloading, which can be useful during web application development. However, coordination is required to reload in a thread-safe manner. The documentation below explains how to achieve this.
 
-The gem is designed so that any project, gem dependency, application, etc. can have their own independent loader, coexisting in the same process, managing their own project trees, and independent of each other. Each loader has its own configuration, inflector, and optional logger.
+The gem is designed to allow any project, gem dependency, or application to have its own independent loader. Multiple loaders can coexist in the same process, each managing its own project tree and operating independently of each other. Each loader has its own configuration, inflector, and optional logger.
 
-Internally, Zeitwerk issues `require` calls exclusively using absolute file names, so there are no costly file system lookups in `$LOAD_PATH`. Technically, the directories managed by Zeitwerk do not even need to be in `$LOAD_PATH`.
+Internally, Zeitwerk exclusively uses absolute file names when issuing `require` calls, eliminating the need for costly file system lookups in `$LOAD_PATH`. Technically, the directories managed by Zeitwerk don't even need to be in `$LOAD_PATH`.
 
-Furthermore, Zeitwerk does at most one single scan of the project tree, and it descends into subdirectories lazily, only if their namespaces are used.
+Furthermore, Zeitwerk performs a single scan of the project tree at most, lazily descending into subdirectories only when their namespaces are used.
 
 <a id="markdown-synopsis" name="synopsis"></a>
 ## Synopsis
@@ -144,7 +149,7 @@ Zeitwerk::Loader.eager_load_all
 <a id="markdown-the-idea-file-paths-match-constant-paths" name="the-idea-file-paths-match-constant-paths"></a>
 ### The idea: File paths match constant paths
 
-To have a file structure Zeitwerk can work with, just name files and directories after the name of the classes and modules they define:
+For Zeitwerk to work with your file structure, simply name files and directories after the classes and modules they define:
 
 ```
 lib/my_gem.rb         -> MyGem
@@ -153,7 +158,7 @@ lib/my_gem/bar_baz.rb -> MyGem::BarBaz
 lib/my_gem/woo/zoo.rb -> MyGem::Woo::Zoo
 ```
 
-You can tune that a bit by [collapsing directories](#collapsing-directories), or by [ignoring parts of the project](#ignoring-parts-of-the-project), but that is the main idea.
+You can fine-tune this behavior by [collapsing directories](#collapsing-directories) or [ignoring specific parts of the project](#ignoring-parts-of-the-project), but that is the main idea.
 
 <a id="markdown-inner-simple-constants" name="inner-simple-constants"></a>
 ### Inner simple constants
@@ -211,7 +216,7 @@ serializers/user_serializer.rb -> UserSerializer
 <a id="markdown-custom-root-namespaces" name="custom-root-namespaces"></a>
 #### Custom root namespaces
 
-While `Object` is by far the most common root namespace, you can associate a different one to a particular root directory. The method `push_dir` accepts a non-anonymous class or module object in the optional `namespace` keyword argument.
+Although `Object` is the most common root namespace, you have the flexibility to associate a different one with a specific root directory. The `push_dir` method accepts a non-anonymous class or module object as the optional `namespace` keyword argument.
 
 For example, given:
 
@@ -227,14 +232,14 @@ a file defining `ActiveJob::QueueAdapters::MyQueueAdapter` does not need the con
 adapters/my_queue_adapter.rb -> ActiveJob::QueueAdapters::MyQueueAdapter
 ```
 
-Please, note that the given root namespace must be non-reloadable, though autoloaded constants in that namespace can be. That is, if you associate `app/api` with an existing `Api` module, that module should not be reloadable. However, if the project defines and autoloads the class `Api::Deliveries`, that one can be reloaded.
+Please note that the provided root namespace must be non-reloadable, while allowing autoloaded constants within that namespace to be reloadable. This means that if you associate the `app/api` directory with an existing `Api` module, the module itself should not be reloadable. However, if the project defines and autoloads the `Api::Deliveries` class, that class can be reloaded.
 
 <a id="markdown-nested-root-directories" name="nested-root-directories"></a>
 #### Nested root directories
 
-Root directories should not be ideally nested, but Zeitwerk supports them because in Rails, for example, both `app/models` and `app/models/concerns` belong to the autoload paths.
+Root directories are recommended not to be nested; however, Zeitwerk provides support for nested root directories since in frameworks like Rails, both `app/models` and `app/models/concerns` belong to the autoload paths.
 
-Zeitwerk detects nested root directories, and treats them as roots only. In the example above, `concerns` is not considered to be a namespace below `app/models`. For example, the file:
+Zeitwerk identifies nested root directories and treats them as independent roots. In the given example, `concerns` is not considered a namespace within `app/models`. For instance, consider the following file:
 
 ```
 app/models/concerns/geolocatable.rb
@@ -245,9 +250,9 @@ should define `Geolocatable`, not `Concerns::Geolocatable`.
 <a id="markdown-implicit-namespaces" name="implicit-namespaces"></a>
 ### Implicit namespaces
 
-If a namespace is just a simple module with no code, you do not need to define it in a file: Directories without a matching Ruby file get modules created automatically on your behalf.
+If a namespace consists only of a simple module without any code, there is no need to explicitly define it in a separate file. Zeitwerk automatically creates modules on your behalf for directories without a corresponding Ruby file.
 
-For example, if a project has an `admin` directory:
+For instance, suppose a project includes an `admin` directory:
 
 ```
 app/controllers/admin/users_controller.rb -> Admin::UsersController
@@ -255,7 +260,7 @@ app/controllers/admin/users_controller.rb -> Admin::UsersController
 
 and does not have a file called `admin.rb`, Zeitwerk automatically creates an `Admin` module on your behalf the first time `Admin` is used.
 
-For this to happen, the directory has to contain non-ignored Ruby files with extension `.rb`, directly or recursively, otherwise it is ignored. This condition is evaluated again on reloads.
+To trigger this behavior, the directory must contain non-ignored Ruby files with the ".rb" extension, either directly or recursively. Otherwise, the directory is ignored. This condition is reevaluated during reloads.
 
 <a id="markdown-explicit-namespaces" name="explicit-namespaces"></a>
 ### Explicit namespaces
@@ -370,7 +375,7 @@ require "zeitwerk"
 loader = Zeitwerk::Loader.new
 loader.tag = File.basename(__FILE__, ".rb")
 loader.inflector = Zeitwerk::GemInflector.new(__FILE__)
-loader.push_dir(__dir__)
+loader.push_dir(File.dirname(__FILE__))
 ```
 
 If the main module references project constants at the top-level, Zeitwerk has to be ready to load them. Their definitions, in turn, may reference other project constants. And this is recursive. Therefore, it is important that the `setup` call happens above the main module definition:
@@ -576,7 +581,7 @@ root_dir2/my_app/routes
 root_dir3/my_app/routes
 ```
 
-where `root_directory{1,2,3}` are root directories, eager loading `MyApp::Routes` will eager load the contents of the three corresponding directories.
+where `root_dir{1,2,3}` are root directories, eager loading `MyApp::Routes` will eager load the contents of the three corresponding directories.
 
 There might exist external source trees implementing part of the namespace. This happens routinely, because top-level constants are stored in the globally shared `Object`. It happens also when deliberately [reopening third-party namespaces](#reopening-third-party-namespaces). Such external code is not eager loaded, the implementation is carefully scoped to what the receiver manages to avoid side-effects elsewhere.
 
@@ -733,9 +738,34 @@ loader.inflector.inflect "html_parser" => "HTMLParser"
 loader.inflector.inflect "mysql_adapter" => "MySQLAdapter"
 ```
 
+Overrides have to match exactly directory or file (without extension) _basenames_. For example, if you configure
+
+```ruby
+loader.inflector.inflect("xml" => "XML")
+```
+
+then the following constants are expected:
+
+```
+xml.rb         -> XML
+foo/xml        -> Foo::XML
+foo/bar/xml.rb -> Foo::Bar::XML
+```
+
+As you see, any directory whose basename is exactly `xml`, and any file whose basename is exactly `xml.rb` are expected to define the constant `XML` in the corresponding namespace. On the other hand, partial matches are ignored. For example, `xml_parser.rb` would be inflected as `XmlParser` because `xml_parser` is not equal to `xml`. You'd need an additional override:
+
+```ruby
+loader.inflector.inflect(
+  "xml"        => "XML",
+  "xml_parser" => "XMLParser"
+)
+```
+
+If you need more flexibility, you can define a custom inflector, as explained down below.
+
 Overrides need to be configured before calling `setup`.
 
- The inflectors of different loaders are independent of each other. There are no global inflection rules or global configuration that can affect this inflector. It is deterministic.
+The inflectors of different loaders are independent of each other. There are no global inflection rules or global configuration that can affect this inflector. It is deterministic.
 
 <a id="markdown-zeitwerkgeminflector" name="zeitwerkgeminflector"></a>
 #### Zeitwerk::GemInflector
@@ -746,6 +776,31 @@ This inflector is like the basic one, except it expects `lib/my_gem/version.rb`
 
 The inflectors of different loaders are independent of each other. There are no global inflection rules or global configuration that can affect this inflector. It is deterministic.
 
+<a id="markdown-zeitwerknullinflector" name="zeitwerknullinflector"></a>
+#### Zeitwerk::NullInflector
+
+This is an experimental inflector that simply returns its input unchanged.
+
+```ruby
+loader.inflector = Zeitwerk::NullInflector.new
+```
+
+In a project using this inflector, the names of files and directories are equal to the constants they define:
+
+```
+User.rb       -> User
+HTMLParser.rb -> HTMLParser
+Admin/Role.rb -> Admin::Role
+```
+
+Point is, you think less. Names that typically need custom configuration like acronyms no longer require your attention. What you see is what you get, simple.
+
+This inflector is experimental since Ruby usually goes for snake case in files and directories. But hey, if you fancy giving it a whirl, go for it!
+
+The null inflector cannot be used in Rails applications because the `main` autoloader also manages engines. However, you could subclass the default inflector and override `camelize` to return the basename untouched if it starts with an uppercase letter. Generators would not create the expected file names, but you could still experiment to see how far this approach takes you.
+
+In case-insensitive file systems, this inflector works as long as directory listings return the expected strings. Zeitwerk lists directories using Ruby APIs like `Dir.children` or `Dir.entries`.
+
 <a id="markdown-custom-inflector" name="custom-inflector"></a>
 #### Custom inflector
 
@@ -978,7 +1033,7 @@ Zeitwerk::Loader.default_logger = method(:puts)
 
 If there is a logger configured, you'll see traces when autoloads are set, files loaded, and modules autovivified. While reloading, removed autoloads and unloaded objects are also traced.
 
-As a curiosity, if your project has namespaces you'll notice in the traces Zeitwerk sets autoloads for _directories_. That's a technique used to be able to descend into subdirectories on demand, avoiding that way unnecessary tree walks.
+As a curiosity, if your project has namespaces you'll notice in the traces Zeitwerk sets autoloads for _directories_. This allows descending into subdirectories on demand, thus avoiding unnecessary tree walks.
 
 <a id="markdown-loader-tag" name="loader-tag"></a>
 #### Loader tag
@@ -1004,13 +1059,13 @@ Zeitwerk@my_gem: constant MyGem::Foo loaded from ...
 <a id="markdown-ignoring-parts-of-the-project" name="ignoring-parts-of-the-project"></a>
 ### Ignoring parts of the project
 
-Zeitwerk ignores automatically any file or directory whose name starts with a dot, and any files that do not have extension ".rb".
+Zeitwerk ignores automatically any file or directory whose name starts with a dot, and any files that do not have the extension ".rb".
 
 However, sometimes it might still be convenient to tell Zeitwerk to completely ignore some particular Ruby file or directory. That is possible with `ignore`, which accepts an arbitrary number of strings or `Pathname` objects, and also an array of them.
 
 You can ignore file names, directory names, and glob patterns. Glob patterns are expanded when they are added and again on each reload.
 
-There is an edge case related to nested root directories. Conceptually, root directories are independent source trees. If you ignore a parent of a nested root directory, the nested root directory is not affected. You need to ignore it explictly if you want it ignored too.
+There is an edge case related to nested root directories. Conceptually, root directories are independent source trees. If you ignore a parent of a nested root directory, the nested root directory is not affected. You need to ignore it explicitly if you want it ignored too.
 
 Let's see some use cases.
 
@@ -1214,6 +1269,9 @@ With that, when Zeitwerk scans the file system and reaches the gem directories `
 <a id="markdown-introspection" name="introspection"></a>
 ### Introspection
 
+<a id="markdown-zeitwerkloaderdirs" name="zeitwerkloaderdirs"></a>
+#### `Zeitwerk::Loader#dirs`
+
 The method `Zeitwerk::Loader#dirs` returns an array with the absolute paths of the root directories as strings:
 
 ```ruby
@@ -1235,6 +1293,92 @@ By default, ignored root directories are filtered out. If you want them included
 
 These collections are read-only. Please add to them with `Zeitwerk::Loader#push_dir`.
 
+<a id="markdown-zeitwerkloadercpath_expected_at" name="zeitwerkloadercpath_expected_at"></a>
+#### `Zeitwerk::Loader#cpath_expected_at`
+
+Given a path as a string or `Pathname` object, `Zeitwerk::Loader#cpath_expected_at` returns a string with the corresponding expected constant path.
+
+Some examples, assuming that `app/models` is a root directory:
+
+```ruby
+loader.cpath_expected_at("app/models")                  # => "Object"
+loader.cpath_expected_at("app/models/user.rb")          # => "User"
+loader.cpath_expected_at("app/models/hotel")            # => "Hotel"
+loader.cpath_expected_at("app/models/hotel/billing.rb") # => "Hotel::Billing"
+```
+
+If `collapsed` is a collapsed directory:
+
+```ruby
+loader.cpath_expected_at("a/b/collapsed/c") # => "A::B::C"
+loader.cpath_expected_at("a/b/collapsed")   # => "A::B", edge case
+loader.cpath_expected_at("a/b")             # => "A::B"
+```
+
+If the argument corresponds to an [ignored file or directory](#ignoring-parts-of-the-project), the method returns `nil`. Same if the argument is not managed by the loader.
+
+`Zeitwerk::Error` is raised if the given path does not exist:
+
+```ruby
+loader.cpath_expected_at("non_existing_file.rb") # => Zeitwerk::Error
+```
+
+`Zeitwerk::NameError` is raised if a constant path cannot be derived from it:
+
+```ruby
+loader.cpath_expected_at("8.rb") # => Zeitwerk::NameError
+```
+
+This method does not parse file contents and does not guarantee files define the returned constant path. It just says which is the _expected_ one.
+
+`Zeitwerk::Loader#cpath_expected_at` is designed to be used with individual paths. If you want to know all the expected constant paths in the project, please use `Zeitwerk::Loader#all_expected_cpaths`, documented next.
+
+<a id="markdown-zeitwerkloaderall_expected_cpaths" name="zeitwerkloaderall_expected_cpaths"></a>
+#### `Zeitwerk::Loader#all_expected_cpaths`
+
+The method `Zeitwerk::Loader#all_expected_cpaths` returns a hash that maps the absolute paths of the files and directories managed by the receiver to their expected constant paths.
+
+Ignored files, hidden files, and files whose extension is not ".rb" are not included in the result. Same for directories, hidden or ignored directories are not included in the result. Additionally, directories that contain no files with extension ".rb" (recursively) are also excluded, since those are not considered to represent Ruby namespaces.
+
+For example, if `lib` is the root directory of a gem with the following contents:
+
+```
+lib/.DS_Store
+lib/my_gem.rb
+lib/my_gem/version.rb
+lib/my_gem/ignored.rb
+lib/my_gem/drivers/unix.rb
+lib/my_gem/drivers/windows.rb
+lib/my_gem/collapsed/foo.rb
+lib/tasks/my_gem.rake
+```
+
+`Zeitwerk::Loader#all_expected_cpaths` would return (maybe in a different order):
+
+```ruby
+{
+  "/.../lib"                           => "Object",
+  "/.../lib/my_gem.rb"                 => "MyGem",
+  "/.../lib/my_gem"                    => "MyGem",
+  "/.../lib/my_gem/version.rb"         => "MyGem::VERSION",
+  "/.../lib/my_gem/drivers"            => "MyGem::Drivers",
+  "/.../lib/my_gem/drivers/unix.rb"    => "MyGem::Drivers::Unix",
+  "/.../lib/my_gem/drivers/windows.rb" => "MyGem::Drivers::Windows",
+  "/.../lib/my_gem/collapsed"          => "MyGem"
+  "/.../lib/my_gem/collapsed/foo.rb"   => "MyGem::Foo"
+}
+```
+
+In the previous example we assume `lib/my_gem/ignored.rb` is ignored, and therefore it is not present in the returned hash. Also, `lib/my_gem/collapsed` is a collapsed directory, so the expected namespace at that level is still `MyGem` (this is an edge case).
+
+The file `lib/.DS_Store` is hidden, hence excluded. The directory `lib/tasks` is also not present because it contains no files with extension ".rb".
+
+Directory paths do not have trailing slashes.
+
+The order of the hash entries is undefined.
+
+This method does not parse or execute file contents and does not guarantee files define the corresponding constant paths. It just says which are the _expected_ ones.
+
 <a id="markdown-encodings" name="encodings"></a>
 ### Encodings
 
@@ -1256,7 +1400,7 @@ The test suite passes on Windows with codepage `Windows-1252` if all the involve
 
 3. In that line, if two loaders manage files that translate to the same constant in the same namespace, the first one wins, the rest are ignored. Similar to what happens with `require` and `$LOAD_PATH`, only the first occurrence matters.
 
-4. Projects that reopen a namespace defined by some dependency have to ensure said namespace is loaded before setup. That is, the project has to make sure it reopens, rather than define. This is often accomplished just loading the dependency.
+4. Projects that reopen a namespace defined by some dependency have to ensure said namespace is loaded before setup. That is, the project has to make sure it reopens, rather than defines, the namespace. This is often accomplished by loading (e.g., `require`-ing) the dependency.
 
 5. Objects stored in reloadable constants should not be cached in places that are not reloaded. For example, non-reloadable classes should not subclass a reloadable class, or mixin a reloadable module. Otherwise, after reloading, those classes or module objects would become stale. Referring to constants in dynamic places like method calls or lambdas is fine.
 
@@ -1265,9 +1409,11 @@ The test suite passes on Windows with codepage `Windows-1252` if all the involve
 <a id="markdown-debuggers" name="debuggers"></a>
 ### Debuggers
 
-Zeitwerk works fine with [debug.rb](https://github.com/ruby/debug) and [Break](https://github.com/gsamokovarov/break).
+Zeitwerk and [debug.rb](https://github.com/ruby/debug) are fully compatible if CRuby is ≥ 3.1 (see [ruby/debug#558](https://github.com/ruby/debug/pull/558)).
+
+[Byebug](https://github.com/deivid-rodriguez/byebug) is compatible except for an edge case explained in [deivid-rodriguez/byebug#564](https://github.com/deivid-rodriguez/byebug/issues/564). Prior to CRuby 3.1, `debug.rb` has a similar edge incompatibility.
 
-[Byebug](https://github.com/deivid-rodriguez/byebug) is compatible except for an edge case explained in [deivid-rodriguez/byebug#564](https://github.com/deivid-rodriguez/byebug/issues/564).
+[Break](https://github.com/gsamokovarov/break) is fully compatible.
 
 <a id="markdown-pronunciation" name="pronunciation"></a>
 ## Pronunciation

data/lib/zeitwerk/cref.rb

@@ -0,0 +1,99 @@
+# frozen_string_literal: true
+
+# This private class encapsulates pairs (mod, cname).
+#
+# Objects represent the constant cname in the class or module object mod, and
+# have API to manage them that encapsulates the constants API. Examples:
+#
+#   cref.path
+#   cref.set(value)
+#   cref.get
+#
+# The constant may or may not exist in mod.
+class Zeitwerk::Cref
+  include Zeitwerk::RealModName
+
+  # @sig Symbol
+  attr_reader :cname
+
+  # The type of the first argument is Module because Class < Module, class
+  # objects are also valid.
+  #
+  # @sig (Module, Symbol) -> void
+  def initialize(mod, cname)
+    @mod   = mod
+    @cname = cname
+    @path  = nil
+  end
+
+  if Symbol.method_defined?(:name)
+    # Symbol#name was introduced in Ruby 3.0. It returns always the same
+    # frozen object, so we may save a few string allocations.
+    #
+    # @sig () -> String
+    def path
+      @path ||= Object.equal?(@mod) ? @cname.name : "#{real_mod_name(@mod)}::#{@cname.name}"
+    end
+  else
+    # @sig () -> String
+    def path
+      @path ||= Object.equal?(@mod) ? @cname.to_s : "#{real_mod_name(@mod)}::#{@cname}"
+    end
+  end
+
+  # The autoload? predicate takes into account the ancestor chain of the
+  # receiver, like const_defined? and other methods in the constants API do.
+  #
+  # For example, given
+  #
+  #   class A
+  #     autoload :X, "x.rb"
+  #   end
+  #
+  #   class B < A
+  #   end
+  #
+  # B.autoload?(:X) returns "x.rb".
+  #
+  # We need a way to retrieve it ignoring ancestors.
+  #
+  # @sig () -> String?
+  if method(:autoload?).arity == 1
+    # @sig () -> String?
+    def autoload?
+      @mod.autoload?(@cname) if self.defined?
+    end
+  else
+    # @sig () -> String?
+    def autoload?
+      @mod.autoload?(@cname, false)
+    end
+  end
+
+  # @sig (String) -> bool
+  def autoload(abspath)
+    @mod.autoload(@cname, abspath)
+  end
+
+  # @sig () -> bool
+  def defined?
+    @mod.const_defined?(@cname, false)
+  end
+
+  # @sig (Object) -> Object
+  def set(value)
+    @mod.const_set(@cname, value)
+  end
+
+  # @raise [NameError]
+  # @sig () -> Object
+  def get
+    @mod.const_get(@cname, false)
+  end
+
+  # @raise [NameError]
+  # @sig () -> void
+  def remove
+    @mod.__send__(:remove_const, @cname)
+  end
+end

data/lib/zeitwerk/gem_loader.rb

@@ -42,13 +42,12 @@ module Zeitwerk
     def warn_on_extra_files
       expected_namespace_dir = @root_file.delete_suffix(".rb")
 
-      ls(@root_dir) do |basename, abspath|
+      ls(@root_dir) do |basename, abspath, ftype|
         next if abspath == @root_file
         next if abspath == expected_namespace_dir
 
         basename_without_ext = basename.delete_suffix(".rb")
         cname = inflector.camelize(basename_without_ext, abspath).to_sym
-        ftype = dir?(abspath) ? "directory" : "file"
 
         warn(<<~EOS)
           WARNING: Zeitwerk defines the constant #{cname} after the #{ftype}

data/lib/zeitwerk/kernel.rb

@@ -14,10 +14,6 @@ module Kernel
   # should not require anything. But if someone has legacy require calls around,
   # they will work as expected, and in a compatible way. This feature is by now
   # EXPERIMENTAL and UNDOCUMENTED.
-  #
-  # We cannot decorate with prepend + super because Kernel has already been
-  # included in Object, and changes in ancestors don't get propagated into
-  # already existing ancestor chains on Ruby < 3.0.
   alias_method :zeitwerk_original_require, :require
   class << self
     alias_method :zeitwerk_original_require, :require
@@ -28,10 +24,10 @@ module Kernel
     if loader = Zeitwerk::Registry.loader_for(path)
       if path.end_with?(".rb")
         required = zeitwerk_original_require(path)
-        loader.on_file_autoloaded(path) if required
+        loader.__on_file_autoloaded(path) if required
         required
       else
-        loader.on_dir_autoloaded(path)
+        loader.__on_dir_autoloaded(path)
         true
       end
     else
@@ -39,7 +35,7 @@ module Kernel
       if required
         abspath = $LOADED_FEATURES.last
         if loader = Zeitwerk::Registry.loader_for(abspath)
-          loader.on_file_autoloaded(abspath)
+          loader.__on_file_autoloaded(abspath)
         end
       end
       required

data/lib/zeitwerk/loader/callbacks.rb

@@ -2,38 +2,45 @@
 
 module Zeitwerk::Loader::Callbacks
   include Zeitwerk::RealModName
+  extend Zeitwerk::Internal
 
   # Invoked from our decorated Kernel#require when a managed file is autoloaded.
   #
-  # @private
   # @sig (String) -> void
-  def on_file_autoloaded(file)
-    cref  = autoloads.delete(file)
-    cpath = cpath(*cref)
+  internal def on_file_autoloaded(file)
+    cref = autoloads.delete(file)
 
     Zeitwerk::Registry.unregister_autoload(file)
 
-    if cdef?(*cref)
-      log("constant #{cpath} loaded from file #{file}") if logger
-      to_unload[cpath] = [file, cref] if reloading_enabled?
-      run_on_load_callbacks(cpath, cget(*cref), file) unless on_load_callbacks.empty?
+    if cref.defined?
+      log("constant #{cref.path} loaded from file #{file}") if logger
+      to_unload[cref.path] = [file, cref] if reloading_enabled?
+      run_on_load_callbacks(cref.path, cref.get, file) unless on_load_callbacks.empty?
     else
-      msg = "expected file #{file} to define constant #{cpath}, but didn't"
+      msg = "expected file #{file} to define constant #{cref.path}, but didn't"
       log(msg) if logger
-      crem(*cref)
-      to_unload[cpath] = [file, cref] if reloading_enabled?
-      raise Zeitwerk::NameError.new(msg, cref.last)
+
+      # Ruby still keeps the autoload defined, but we remove it because the
+      # contract in Zeitwerk is more strict.
+      cref.remove
+
+      # Since the expected constant was not defined, there is nothing to unload.
+      # However, if the exception is rescued and reloading is enabled, we still
+      # need to deleted the file from $LOADED_FEATURES.
+      to_unload[cref.path] = [file, cref] if reloading_enabled?
+
+      raise Zeitwerk::NameError.new(msg, cref.cname)
     end
   end
 
   # Invoked from our decorated Kernel#require when a managed directory is
   # autoloaded.
   #
-  # @private
   # @sig (String) -> void
-  def on_dir_autoloaded(dir)
-    # Module#autoload does not serialize concurrent requires, and we handle
-    # directories ourselves, so the callback needs to account for concurrency.
+  internal def on_dir_autoloaded(dir)
+    # Module#autoload does not serialize concurrent requires in CRuby < 3.2, and
+    # we handle directories ourselves without going through Kernel#require, so
+    # the callback needs to account for concurrency.
     #
     # Multi-threading would introduce a race condition here in which thread t1
     # autovivifies the module, and while autoloads for its children are being
@@ -43,10 +50,10 @@ module Zeitwerk::Loader::Callbacks
     # That not only would reassign the constant (undesirable per se) but, worse,
     # the module object created by t2 wouldn't have any of the autoloads for its
     # children, since t1 would have correctly deleted its namespace_dirs entry.
-    mutex2.synchronize do
+    dirs_autoload_monitor.synchronize do
       if cref = autoloads.delete(dir)
-        autovivified_module = cref[0].const_set(cref[1], Module.new)
-        cpath = autovivified_module.name
+        implicit_namespace = cref.set(Module.new)
+        cpath = implicit_namespace.name
         log("module #{cpath} autovivified from directory #{dir}") if logger
 
         to_unload[cpath] = [dir, cref] if reloading_enabled?
@@ -57,9 +64,9 @@ module Zeitwerk::Loader::Callbacks
         # these to be able to unregister later if eager loading.
         autoloaded_dirs << dir
 
-        on_namespace_loaded(autovivified_module)
+        on_namespace_loaded(implicit_namespace)
 
-        run_on_load_callbacks(cpath, autovivified_module, dir) unless on_load_callbacks.empty?
+        run_on_load_callbacks(cpath, implicit_namespace, dir) unless on_load_callbacks.empty?
       end
     end
   end
@@ -73,7 +80,7 @@ module Zeitwerk::Loader::Callbacks
   def on_namespace_loaded(namespace)
     if dirs = namespace_dirs.delete(real_mod_name(namespace))
       dirs.each do |dir|
-        set_autoloads_in_dir(dir, namespace)
+        define_autoloads_for_dir(dir, namespace)
       end
     end
   end