Since 07e5301e, `Relation#first` will order by primary key if no order is
defined.

By making the Rails minitest behave like a standard minitest plugin
we're much more likely to not break when people use other minitest
plugins. Like minitest-focus and pride.

To do this, we need to behave like minitest: require files up front
and then perform the plugin behavior via the at_exit hook.
This also saves us a fair bit of wrangling with test file loading.

Finally, since the environment and warnings options have to be applied
as early as possible, and since minitest loads plugins at_exit, they
have to be moved to the test command.

* Don't expect the root method.

It's likely this worked because we eagerly loaded the Rails minitest plugin
and that somehow defined a root method on `Rails`.

* Assign a backtrace to failed exceptions.

Otherwise Minitest pukes when attempting to filter the backtrace (which
Rails' backtrace cleaner then removes).

Means the exception message test has to be revised too.

This is likely caused by the rails minitest plugin now being loaded for
these tests and assigning a default backtrace cleaner.

The uuid validation regex was allowing uuids to have a single leading
curly brace or single trailing curly brace. Saving with such a uuid
would cause Postgres to generate an exception even though the record
seemed valid. With this change, the regex requires both a leading *and*
a trailing curly brace or neither to be valid.

Since MariaDB 10.2, `CURRENT_TIMESTAMP` is shown as a function
(`current_timestamp()`). Fix matching column default to address that
case.

Fixes #29698.

Since MariaDB 10.2.5, `information_schema` supports Virtual Columns.

Fixes #29670.

This name more accurately describes what the method does, and also
disambiguates it from `_write_attribute`, which ignores aliases.

We can also make the method private, since it's not public API and only
called from one place - `update_columns` - without an explicit receiver.

Currently `version` method always returns `5.5.5` because the
`full_version` is `5.5.5-10.x.y-MariaDB...` since MariaDB 10.x.
It should be ignored if the prefix is `5.5.5-`.

Whoops.

The `find_each`, `find_in_batches` and `in_batches` APIs usually operate
on large numbers of records, where it's preferable not to load them all
into memory at once.

If the query cache is enabled, it will hold onto the query results until
the end of the execution context (request/job), which means the memory
used is still proportional to the total number of records. These queries
are typically not repeated, so the query cache isn't desirable here.

The AbstractAdapter will translate all StandardErrors generated during the course of a query into ActiveRecord::StatementInvalids. Unfortunately, it'll also mangle non-database-related errors generated in ActiveSupport::Notification callbacks after the query has successfully completed. This should prevent it from translating errors from ActiveSupport::Notifications.

`ActiveModel::Type::DateTime#serialize` should return a `Time` object
so that finding by a datetime column works correctly.

We already have database agnostic `Type::Json` since #29220.

`attributes` (`values` in `LazyAttributeHash` in `AttributeSet`) has
string keys, not symbols.

Fixes #29665.

Fixes #28324.

These are no longer used since b98668de.

It is covered by following assertion.

Because the deprecation message is not yet released.

Let's say you have a nested transaction and both records are saved.
Before the outer transaction closes, a rollback is performed. Previously
the record in the outer transaction would get marked as not persisted
but the inner transaction would get persisted.

```ruby
Post.transaction do
  post_one.save # will get rolled back

  Post.transaction(requires_new: true) do
    post_two.save # incorrectly remains marked as persisted
  end

  raise ActiveRecord::Rollback
end
```

To fix this the PR changes transaction handling to have the child
transaction ask the parent how the records should be marked. When
there are child transactions, it will always be a SavpointTransaction
because the stack isn't empty. From there we pass the parent_transaction
to the child SavepointTransaction where we add the children to the parent
so the parent can mark the inner transaction as rolledback and thus mark
the record as not persisted.

`update_attributes_from_transaction_state` uses the `completed?` check to
correctly mark all the transactions as rolledback and the inner record as
not persisted.

```ruby
Post.transaction do
  post_one.save # will get rolled back

  Post.transaction(requires_new: true) do
    post_two.save # with new behavior, correctly marked as not persisted
    on rollback
  end

  raise ActiveRecord::Rollback
end
```

Fixes #29320

`set_state` was directly setting the transaction state instance
variable. It's better to set the state via specific methods (`rollback!`
and `commit!` respectively.

While undocumented and untested, it's possible someone is using
`set_state` in their app or gem so I've added a deprecation notice to
it.

No where in the app do we use `nullify!` but I wanted to keep existing
behavior while replacing the method with a better pattern.

Since #29540, `# frozen_string_literal: true` included original
migration files.

This reverts commit 3420a145, reversing
changes made to afb66a5a.

`test_middleware_caches` is sometimes failed since #29454.
The failure is due to schema statements are affected by query caching.
Bypassing query caching for schema statements to avoid the issue.

Fixes #29627.

Active Record doesn't rely delegating to `arel` in the internal since
425f2cac. The delegation is a lower priority than delegating to `klass`,
so it is pretty unclear which method is delegated to `arel`.

For example, `bind_values` method was removed at b06f64c3 (a series of
changes https://github.com/rails/rails/compare/79f71d3...b06f64c). But a
relation still could respond to the method because `arel` also have the
same named method (#28976).

Removing the delegation will achieve predictable behavior.

`test_cache_is_available_when_using_a_not_connected_connection` is
always failed if running only the test since #29609.

```
% ARCONN=mysql2 be ruby -w -Itest test/cases/query_cache_test.rb -n test_cache_is_available_when_using_a_not_connected_connection
Using mysql2
Run options: -n test_cache_is_available_when_using_a_not_connected_connection --seed 15043

F

Finished in 0.070519s, 14.1806 runs/s, 28.3612 assertions/s.

  1) Failure:
QueryCacheTest#test_cache_is_available_when_using_a_not_connected_connection [test/cases/query_cache_test.rb:336]:
2 instead of 1 queries were executed.
Queries:
SELECT  `tasks`.* FROM `tasks` WHERE `tasks`.`id` = ? LIMIT ?
SET NAMES utf8 COLLATE utf8_unicode_ci,  @@SESSION.sql_mode = CONCAT(CONCAT(@@sql_mode, ',STRICT_ALL_TABLES'), ',NO_AUTO_VALUE_ON_ZERO'),  @@SESSION.sql_auto_is_null = 0, @@SESSION.wait_timeout = 2147483.
Expected: 1
  Actual: 2

1 runs, 2 assertions, 1 failures, 0 errors, 0 skips
```

This failure is due to `LogSubscriber` will use not connected
`ActiveRecord::Base.connection` even if `Task.connection` is connected.
I fixed to always pass `type_casted_binds` to log subscriber to avoid
the issue.

This regression was caused by #23004.

If STI subclass is using scoping in parent class scoping,
`current_scope` in subclass is never restored.
I fixed to restore `current_scope` to previous value correctly.

`aliased_table_name` in `Association` was added at a3502c41.
`aliased_table_name` in `JoinDependency` (added at 55854c41) is used, but
it looks like that added one in `Association` is never used from the
beginning.