This clears the transaction record state when the transaction finishes
with a `:committed` status.

Considering the following example where `name` is a required attribute.
Before we had `new_record?` returning `true` for a persisted record:

```ruby
  author = Author.create! name: 'foo'
  author.name = nil
  author.save        # => false
  author.new_record? # => true
```

As per the docs, `mark_for_destruction` should do nothing if `autosave`
is not set to true. We normally persist associations on a record no
matter what if the record is a new record, but we were always skipping
records which were `marked_for_destruction?`.

Fixes #20882

Also removes a false positive test that depends on the fixed bug:

At this time, counter_cache does not work with polymorphic relationships
(which is a bug). The test was added to make sure that no
StaleObjectError is raised when the car is destroyed. No such error is
currently raised because the lock version is not incremented by
appending a wheel to the car.

Furthermore, `assert_difference` succeeds because `car.wheels.count`
does not check the counter cache, but the collection size. The test will
fail if it is replaced with `car.wheels_count || 0`.

This code is so fucked. Things that cause this bug not to replicate:

- Defining the validation before the association (we end up calling
  `uniq!` on the errors in the autosave validation)
- Adding `accepts_nested_attributes_for` (I have no clue why. The only
  thing it does that should affect this is adds `autosave: true` to the
  inverse reflection, and doing that manually doesn't fix this).

This solution is a hack, and I'm almost certain there's a better way to
go about it, but this shouldn't cause a huge hit on validation times,
and is the simplest way to get it done.

Fixes #20874.

Since nested hashes are also instances of
`ActionController::Parameters`, and we're explicitly looking to work
with a hash for nested attributes, this caused breakage in several
points.

This is the minimum viable fix for the issue (and one that I'm not
terribly fond of). I can't think of a better place to handle this at the
moment. I'd prefer to use some sort of solution that doesn't special
case AC::Parameters, but we can't use something like `to_h` or `to_a`
since `Enumerable` adds both.

While I've added a trivial test case for verifying this fix in
isolation, we really need better integration coverage to prevent
regressions like this in the future. We don't actually have a lot of
great places for integration coverage at the moment, so I'm deferring it
for now.

Fixes #20922.

This is to simplify the association API, as you can call `reload` on the
association proxy or the parent object to get the same result.

For collection association, you can call `#reload` on association proxy
to force a reload:

    @user.posts.reload   # Instead of @user.posts(true)

For singular association, you can call `#reload` on the parent object to
clear its association cache then call the association method:

    @user.reload.profile   # Instead of @user.profile(true)

Passing a truthy argument to force association to reload will be removed
in Rails 5.1.

The concurrent-ruby gem is a toolset containing many concurrency
utilities. Many of these utilities include runtime-specific
optimizations when possible. Rather than clutter the Rails codebase with
concurrency utilities separate from the core task, such tools can be
superseded by similar tools in the more specialized gem. This commit
replaces `ActiveSupport::Concurrency::Latch` with
`Concurrent::CountDownLatch`, which is functionally equivalent.

The changes introduced to through associations in c80487eb were quite
interesting. Changing `relation.merge!(scope)` to `relation =
relation.merge(scope)` should in theory never cause any changes in
behavior. The subtle breakage led to a surprising conclusion.

The old code wasn't doing anything! Since `merge!` calls
`instance_exec` when given a proc, and most scopes will look something
like `has_many :foos, -> { where(foo: :bar) }`, if we're not capturing
the return value, it's a no-op. However, removing the `merge` causes
`unscope` to break.

While we're merging in the rest of the chain elsewhere, we were never
merging in `unscope` values, causing a breakage on associations where a
default scope was being unscoped in an association scope (yuk!). This is
subtly related to #20722, since it appears we were previously relying on
this mutability.

Fixes #20721.
Fixes #20727.

Closes #20743.

The task `db:_dump` now only dumps the schema if
`ActiveRecord::Base.dump_schema_after_migration` is true. This has
effects:

- `db:migrate:up`
- `db:migrate:down`
- `db:forward`
- `db:rollback`

Passing `:from` and `:to` to `change_column_default` makes this command
reversible as user has defined its previous state.

So, instead of having the migration command as:

    change_column_default(:posts, :state, "draft")

They can write it as:

    change_column_default(:posts, :state, from: nil, to: "draft")

thrown ActiveRecord::AssociationTypeMismatch when assigning a wrong value for a namespaced association

fixes #20541

Associated objects that were marked for destruction are considered absent.

After discussing, we've decided it makes more sense to include it. We're
already forwarding every conflicting method to `to_a`, and there's no
conflation of concerns. `Enumerable` has no mutating methods, and it
just allows us to simplify the code. No existing methods will have a
change in behavior. Un-overridden Enumerable methods will simply
delegate to `each`.

[Sean Griffin & bogdan]

This matches our behavior in other cases where useful enumerable methods
might have a different definition in `Relation`. Wanting to actually
enumerate over the records in this case is completely reasonable, and
wanting `.sum` is reasonable for the same reason it is on `Enumerable`
in the first place.

This used to raise a `IrreversibleMigration` error (since #10437).
However since `remove_index :table, :column` is probably the most basic
use-case we should make it reversible again.

Conflicts:
	activerecord/CHANGELOG.md

If a getter has side effects on the DB, `changes_applied` will be called
twice. The second time will try and remove the changed attributes cache,
and will crash because it's already been unset. This also demonstrates
that we shouldn't assume that calling getters won't change the value of
`changed_attributes`, and we need to clear the cache if an attribute is
modified.

Fixes #20531.

Fixes #17511 and #17415

[Sean Griffin & jmondo]

If the subtype provides custom schema dumping behavior, we need to defer
to it. We purposely choose not to handle any values other than an array
(which technically should only ever be `nil`, but I'd rather code
defensively here).

Fixes #20515.

This introduces a deprecation cycle to change the behavior of the
default point type in the PostgreSQL adapter. The old behavior will
continue to be available for the immediate future as `:legacy_point`.

The current behavior of returning an `Array` causes several problems,
the most significant of which is that we cannot differentiate between an
array of points, and a point itself in the case of a column with the
`point[]` type.

The attributes API gives us a reasonable way to have a proper
deprecation cycle for this change, so let's take advantage of it. If we
like this change, we can also add proper support for the other geometric
types (line, lseg, box, path, polygon, and circle), all of which are
just aliases for string today.

Fixes #20441

Our general contract in Active Record is that strings are assumed to be
SQL literals, and symbols are assumed to reference a column. If a from
clause is given, we shouldn't include the table name, but we should
still quote the value as if it were a column.

Upon fixing this, the tests were still failing on SQLite. This was
because the column name being returned by the query was `"\"join\""`
instead of `"join"`. This is actually a bug in SQLite that was fixed a
long time ago, but I was using the version of SQLite included by OS X
which has this bug. Since I'm guessing this will be a common case for
contributors, I also added an explicit check with a more helpful error
message.

Fixes #20360

Currently, values for columns backing Active Record enums must be
specified as integers in test fixtures:

    awdr:
      title: "Agile Web Development with Rails"
      status: 2

    rfr:
      title: "Ruby for Rails"
      status: <%= Book.statuses[:proposed] %>

This is potentially confusing, since enum values are typically
specified as symbols or strings in application code. To resolve the
confusion, this change permits the use of symbols or strings to specify
enum values:

    awdr:
      status: :published

It is compatible with fixtures that specify enum values as integers.

See #9683 for the reasons we switched to `distinct`.

Here is the discussion that triggered the actual deprecation #20198.

`uniq`, `uniq!` and `uniq_value` are still around.
They will be removed in the next minor release after Rails 5.

Add full set of MySQL CLI options to support SSL authentication when using db:structure dump and load

If your STI class looks like this:

```ruby
class Company < ActiveRecord::Base
  self.store_full_sti_class = false

  class GoodCo < Company
  end

  class BadCo < Company
  end
end
```

The expectation (which is valid) is that the `type` in the database is saved as
`GoodCo` or `BadCo`. However, another expectation should be that setting `type`
to `GoodCo` would correctly instantiate the object as a `Company::GoodCo`. That
second expectation is what this should fix.

Example:

    create_table :foos do |t|
      t.string :string_en, collation: 'en_US.UTF-8'
      t.text   :text_ja,   collation: 'ja_JP.UTF-8'
    end

If the adapter supports indexes in create table, generated SQL is
slightly more efficient.