https://buildkite.com/rails/rails/builds/61362#99165d42-172d-4ad5-bf72-b29d8cd44f3e/995-1006

https://buildkite.com/rails/rails/builds/61358#a78ee50e-30b5-48a2-858f-63eba287d919/1290-1298

And no longer need to except SCHEMA SQLs manually since 0810c076.

This reverts commit a1ee4a9f.

Even if a1ee4a9f is applied, CI is still flakiness.

https://buildkite.com/rails/rails/builds/61252#2c090afa-aa84-4a2b-8b81-9f09219222c6/994-1005
https://buildkite.com/rails/rails/builds/61252#2e55bf83-1bde-44a2-a4f1-b5c3f6820fb4/929-938

Failing tests by whether schema cache is filled or not, it actually
means that whether SCHEMA SQLs are executed or not is not target for the
tests.

So I've reverted commit a1ee4a9f which filling schema cache before
`assert_no_queries`, and replace `assert_no_queries` to
`assert_queries(0)`.

Unfortunately, a11a8ff7 had no effect as long as using bind param, and
was not tested.

This ensures making the intent of a11a8ff7, which fall back to type
casting from the connection adapter.

Fixes #35205.

```
% ARCONN=postgresql bundle exec ruby -w -Itest test/cases/relation/where_test.rb -n test_type_casting_nested_joins
Using postgresql
Run options: -n test_type_casting_nested_joins --seed 55730

# Running:

E

Error:
ActiveRecord::WhereTest#test_type_casting_nested_joins:
ActiveRecord::StatementInvalid: PG::InvalidTextRepresentation: ERROR:  invalid input syntax for integer: "2-foo"

rails test test/cases/relation/where_test.rb:30

Finished in 0.245778s, 4.0687 runs/s, 0.0000 assertions/s.
1 runs, 0 assertions, 0 failures, 1 errors, 0 skips
```

Testing the result of `capture_sql` is fragile, it is due to whether
SCHEMA SQLs are executed or not depends on whether schema cache is
filled or not.

https://buildkite.com/rails/rails/builds/61248#a5b9dc59-ff0c-40c0-b56e-0895662fbc4c/993-1004
https://buildkite.com/rails/rails/builds/61248#1157b389-f2c7-4554-b6e5-a37624a0e74a/996-1005

I've confirmed all `capture_sql` use cases in our code base, all cases
won't expect SCHEMA SQLs are included.

```
% git grep -n capture_sql
test/cases/associations/belongs_to_associations_test.rb:202:    sql = capture_sql { comment.post }
test/cases/associations/belongs_to_associations_test.rb:204:    assert_not_equal sql, capture_sql { comment.post }
test/cases/associations/has_many_associations_test.rb:169:    sql = capture_sql { post.comments.to_a }
test/cases/associations/has_many_associations_test.rb:171:    assert_not_equal sql, capture_sql { post.comments.to_a }
test/cases/associations/has_many_associations_test.rb:276:    expected_sql = capture_sql { author.thinking_posts.delete_all }
test/cases/associations/has_many_associations_test.rb:281:    loaded_sql = capture_sql { author.thinking_posts.delete_all }
test/cases/associations/has_many_associations_test.rb:289:    expected_sql = capture_sql { author.posts.delete_all }
test/cases/associations/has_many_associations_test.rb:294:    loaded_sql = capture_sql { author.posts.delete_all }
test/cases/associations/left_outer_join_association_test.rb:22:      queries = capture_sql do
test/cases/associations/left_outer_join_association_test.rb:49:    queries = capture_sql { Author.left_outer_joins(:posts).to_a }
test/cases/associations/left_outer_join_association_test.rb:54:    queries = capture_sql { Author.joins(:posts).left_outer_joins(:posts).to_a }
test/cases/associations/left_outer_join_association_test.rb:60:    queries = capture_sql { Author.left_outer_joins({}).to_a }
test/cases/associations/left_outer_join_association_test.rb:65:    queries = capture_sql { Author.left_outer_joins([]).to_a }
test/cases/associations/left_outer_join_association_test.rb:78:    queries = capture_sql { Author.left_outer_joins(:essays).to_a }
test/cases/associations_test.rb:384:    log = capture_sql do
test/cases/associations_test.rb:399:    log = capture_sql do
test/cases/associations_test.rb:414:    log = capture_sql do
test/cases/associations_test.rb:429:    log = capture_sql do
test/cases/associations_test.rb:444:    log = capture_sql do
test/cases/associations_test.rb:459:    log = capture_sql do
test/cases/reflection_test.rb:307:    expected_sql = capture_sql { hotel.recipes.to_a }
test/cases/reflection_test.rb:312:    loaded_sql = capture_sql { hotel.recipes.to_a }
test/cases/relation_test.rb:212:      queries = capture_sql { Author.joins(:posts).merge(Post.joins(:comments)).to_a }
test/cases/relation_test.rb:232:      queries = capture_sql { Post.joins(:author, :categorizations).merge(Author.select(:id)).merge(categorizations_with_authors).to_a }
test/cases/relation_test.rb:347:      log = capture_sql do
test/cases/scoping/relation_scoping_test.rb:146:      log = capture_sql do
test/cases/scoping/relation_scoping_test.rb:159:    log = capture_sql do
test/cases/test_case.rb:33:    def capture_sql
test/cases/test_case.rb:41:      capture_sql { yield }
```

#36293 was an issue for through association with `joins` for a long
time, but since #35864 through association with `left_joins` would also
be affected by the issue.

Implicit through table joins should be appeared before user supplied
joins, otherwise loading through association with joins will cause a
statement invalid error.

Fixes #36293.

```
% ARCONN=postgresql bundle exec ruby -w -Itest test/cases/associations/has_many_through_associations_test
.rb -n test_through_association_with_joins
Using postgresql
Run options: -n test_through_association_with_joins --seed 7116

# Running:

E

Error:
HasManyThroughAssociationsTest#test_through_association_with_joins:
ActiveRecord::StatementInvalid: PG::UndefinedTable: ERROR:  missing FROM-clause entry for table "posts"
LINE 1: ... "comments_posts" ON "comments_posts"."post_id" = "posts"."i...
                                                             ^
: SELECT "comments".* FROM "comments" INNER JOIN "comments" "comments_posts" ON "comments_posts"."post_id" = "posts"."id" INNER JOIN "posts" ON "comments"."post_id" = "posts"."id" WHERE "posts"."author_id" = $1

rails test test/cases/associations/has_many_through_associations_test.rb:61

Finished in 0.388657s, 2.5730 runs/s, 0.0000 assertions/s.
1 runs, 0 assertions, 0 failures, 1 errors, 0 skips
```

Fixes #34456.

This partly reverts the effect of d1107f4d.

d1107f4d makes `touch` tracks the mutation whether the `touch` is
occurred by explicit or not.

Existing apps expects that the previous changes tracks only the changes
which is explicit action by users.

I'd revert the implicit `touch` mutation tracking since I'd not like to
break existing apps.

Fixes #36219.

This commit adds "TRANSACTION" to savepoint and commit, rollback statements
because none of savepoint statements were removed by #36153 since they are not "SCHEMA" statements.

Although, only savepoint statements can be labeled as "TRANSACTION"
I think all of transaction related method should add this label.

Follow up #36153

Currently, `committed!`/`rolledback!` will only be attempted for the
first enrolled record in the transaction, that will cause some
problematic behaviors.

The first one problem, `clear_transaction_record_state` won't be called
even if the transaction is finalized except the first enrolled record.
This means that de-duplicated records in the transaction won't refer
latest state (e.g. won't happen rolling back record state).

The second one problem, the enrolled order is not always the same as the
order in which the actions actually happened, the first enrolled record
may succeed no actions (e.g. `destroy` has already succeeded on another
record during `before_destroy`), it will lose to fire any transactional
callbacks.

To avoid both problems, we should attempt `committed!`/`rolledback!` to
all enrolled records in the transaction.

with `ActiveRecord::BindParameterTest#test_too_many_binds`

sqlite adapter has its own `bind_params_length`, `ActiveRecord::BindParameterTest#test_too_many_binds` respects it.

* Modified `ActiveRecord::BindParameterTest#test_too_many_binds` to show `bind_params_length` value

```
$ git diff
diff --git a/activerecord/test/cases/bind_parameter_test.rb b/activerecord/test/cases/bind_parameter_test.rb
index 85685d1d00..83cd07f1d7 100644
--- a/activerecord/test/cases/bind_parameter_test.rb
+++ b/activerecord/test/cases/bind_parameter_test.rb
@@ -108,6 +108,7 @@ def test_statement_cache_with_sql_string_literal

       def test_too_many_binds
         bind_params_length = @connection.send(:bind_params_length)
+        p bind_params_length

         topics = Topic.where(id: (1 .. bind_params_length).to_a << 2**63)
         assert_equal Topic.count, topics.count
$
```

* Executed modified `ActiveRecord::BindParameterTest#test_too_many_binds`

```
$ bin/test test/cases/bind_parameter_test.rb -n test_too_many_binds
Using sqlite3
Run options: -n test_too_many_binds --seed 47321

999
.

Finished in 0.075249s, 13.2892 runs/s, 26.5784 assertions/s.
1 runs, 2 assertions, 0 failures, 0 errors, 0 skips
$
```

* Make scope arity check consistent

* Add test for arity change

[Rob Trame + Rafael Mendonça França]

Every database executes different type of sql statement to get metadata then `ActiveRecord::TestCase` ignores these database specific sql statements to make `assert_queries` or `assert_no_queries` work consistently.

Connection adapter already labels these statement by setting "SCHEMA" argument, this pull request makes use of "SCHEMA" argument to ignore metadata queries.

Here are the details of these changes:

* PostgresqlConnectionTest

Each of PostgresqlConnectionTest modified just executes corresponding methods

https://github.com/rails/rails/blob/fef174f5c524edacbcad846d68400e7fe114a15a/activerecord/lib/active_record/connection_adapters/postgresql/schema_statements.rb#L182-L195

```ruby
        # Returns the current database encoding format.
        def encoding
          query_value("SELECT pg_encoding_to_char(encoding) FROM pg_database WHERE datname = current_database()", "SCHEMA")
        end

        # Returns the current database collation.
        def collation
          query_value("SELECT datcollate FROM pg_database WHERE datname = current_database()", "SCHEMA")
        end

        # Returns the current database ctype.
        def ctype
          query_value("SELECT datctype FROM pg_database WHERE datname = current_database()", "SCHEMA")
        end
```

* BulkAlterTableMigrationsTest

mysql2 adapter executes `SHOW KEYS FROM ...` to see if there is an index already created as below. I think the main concerns of these tests are how each database adapter creates or drops indexes then ignoring  `SHOW KEYS FROM` statement makes sense.

https://github.com/rails/rails/blob/fef174f5c524edacbcad846d68400e7fe114a15a/activerecord/lib/active_record/connection_adapters/mysql/schema_statements.rb#L11

```ruby
          execute_and_free("SHOW KEYS FROM #{quote_table_name(table_name)}", "SCHEMA") do |result|
```

* Temporary change not included in this commit to show which statements executed

```diff
$ git diff
diff --git a/activerecord/test/cases/migration_test.rb b/activerecord/test/cases/migration_test.rb
index 8e8ed494d9..df05f9bd16 100644
--- a/activerecord/test/cases/migration_test.rb
+++ b/activerecord/test/cases/migration_test.rb
@@ -854,7 +854,7 @@ def test_adding_indexes

       classname = ActiveRecord::Base.connection.class.name[/[^:]*$/]
       expected_query_count = {
-        "Mysql2Adapter"     => 3, # Adding an index fires a query every time to check if an index already exists or not
+        "Mysql2Adapter"     => 1, # Adding an index fires a query every time to check if an index already exists or not
         "PostgreSQLAdapter" => 2,
       }.fetch(classname) {
         raise "need an expected query count for #{classname}"
@@ -886,7 +886,7 @@ def test_removing_index

       classname = ActiveRecord::Base.connection.class.name[/[^:]*$/]
       expected_query_count = {
-        "Mysql2Adapter"     => 3, # Adding an index fires a query every time to check if an index already exists or not
+        "Mysql2Adapter"     => 1, # Adding an index fires a query every time to check if an index already exists or not
         "PostgreSQLAdapter" => 2,
       }.fetch(classname) {
         raise "need an expected query count for #{classname}"
$
```

* Executed these modified tests

```ruby
$ ARCONN=mysql2 bin/test test/cases/migration_test.rb -n /index/
Using mysql2
Run options: -n /index/ --seed 8462

F

Failure:
BulkAlterTableMigrationsTest#test_adding_indexes [/home/yahonda/git/rails/activerecord/test/cases/migration_test.rb:863]:
3 instead of 1 queries were executed.
Queries:
SHOW KEYS FROM `delete_me`
SHOW KEYS FROM `delete_me`
ALTER TABLE `delete_me` ADD UNIQUE INDEX `awesome_username_index`  (`username`), ADD  INDEX `index_delete_me_on_name_and_age`  (`name`, `age`).
Expected: 1
  Actual: 3

bin/test test/cases/migration_test.rb:848

F

Failure:
BulkAlterTableMigrationsTest#test_removing_index [/home/yahonda/git/rails/activerecord/test/cases/migration_test.rb:895]:
3 instead of 1 queries were executed.
Queries:
SHOW KEYS FROM `delete_me`
SHOW KEYS FROM `delete_me`
ALTER TABLE `delete_me` DROP INDEX `index_delete_me_on_name`, ADD UNIQUE INDEX `new_name_index`  (`name`).
Expected: 1
  Actual: 3

bin/test test/cases/migration_test.rb:879

..

Finished in 0.379245s, 10.5473 runs/s, 7.9105 assertions/s.
4 runs, 3 assertions, 2 failures, 0 errors, 0 skips
$
```

* ActiveRecord::ConnectionAdapters::Savepoints

Left `self.ignored_sql` to ignore savepoint related statements because these SQL statements are not related "SCHEMA"

```
self.ignored_sql = [/^SAVEPOINT/, /^ROLLBACK TO SAVEPOINT/, /^RELEASE SAVEPOINT/]
```

https://github.com/rails/rails/blob/fef174f5c524edacbcad846d68400e7fe114a15a/activerecord/lib/active_record/connection_adapters/abstract/savepoints.rb#L10-L20

```ruby
      def create_savepoint(name = current_savepoint_name)
        execute("SAVEPOINT #{name}")
      end

      def exec_rollback_to_savepoint(name = current_savepoint_name)
        execute("ROLLBACK TO SAVEPOINT #{name}")
      end

      def release_savepoint(name = current_savepoint_name)
        execute("RELEASE SAVEPOINT #{name}")
      end
```

If the same id's records are saved and/or destroyed in the transaction,
commit callbackes will only run for the first enrolled record.

https://github.com/rails/rails/blob/a023e2180093ebc517a642aaf21f3c7241c67657/activerecord/lib/active_record/connection_adapters/abstract/transaction.rb#L115-L119

The regression #36132 is caused due to #35920 changed the enrollment
order that the first action's record will be enrolled to last in the
transaction.

We could not change the the enrollment order as long as someone depends
on the enrollment order.

Fixes #36132.

This fixes a regression for #35864.

Usually, stashed joins (mainly eager loading) are performed as LEFT
JOINs.
But the case of merging joins/left_joins of different class, that
(stashed) joins are performed as the same `join_type` as the parent
context for now.
Since #35864, both (joins/left_joins) stashed joins might be contained
in `joins_values`, so each stashed joins should maintain its own
`join_type` context.

Fixes #36103.

Follow up of #35838.

And also this refactors `in_clause_length` handling is entirely
integrated in Arel visitor.

I've realized that `user.id` is 20% slower than `user.name` in the
benchmark (https://github.com/rails/rails/pull/35987#issuecomment-483882480).

The reason that performance difference is that `self.class.primary_key`
method call is a bit slow.

Avoiding that method call will make almost attribute access faster and
`user.id` will be completely the same performance with `user.name`.

Before (02b5b8cb):

```
Warming up --------------------------------------
             user.id   140.535k i/100ms
          user['id']    96.549k i/100ms
           user.name   158.110k i/100ms
        user['name']    94.507k i/100ms
       user.changed?    19.003k i/100ms
 user.saved_changes?    25.404k i/100ms
Calculating -------------------------------------
             user.id      2.231M (± 0.9%) i/s -     11.243M in   5.040066s
          user['id']      1.310M (± 1.3%) i/s -      6.565M in   5.012607s
           user.name      2.683M (± 1.2%) i/s -     13.439M in   5.009392s
        user['name']      1.322M (± 0.9%) i/s -      6.615M in   5.003239s
       user.changed?    201.999k (±10.9%) i/s -      1.007M in   5.091195s
 user.saved_changes?    258.214k (±17.1%) i/s -      1.245M in   5.007421s
```

After (this change):

```
Warming up --------------------------------------
             user.id   158.364k i/100ms
          user['id']   106.412k i/100ms
           user.name   158.644k i/100ms
        user['name']   107.518k i/100ms
       user.changed?    19.082k i/100ms
 user.saved_changes?    24.886k i/100ms
Calculating -------------------------------------
             user.id      2.768M (± 1.1%) i/s -     13.936M in   5.034957s
          user['id']      1.507M (± 2.1%) i/s -      7.555M in   5.017211s
           user.name      2.727M (± 1.5%) i/s -     13.643M in   5.004766s
        user['name']      1.521M (± 1.3%) i/s -      7.634M in   5.018321s
       user.changed?    200.865k (±11.1%) i/s -    992.264k in   5.044868s
 user.saved_changes?    269.652k (±10.5%) i/s -      1.344M in   5.077972s
```

This reverts commit f656bb30.

Reason: Test in Action View expects the `collection_cache_key` working...

https://github.com/rails/rails/blob/ff6b713f5e729859995f204093ad3f8e08f39ea8/actionview/test/activerecord/relation_cache_test.rb#L21
https://github.com/rails/rails/blob/ff6b713f5e729859995f204093ad3f8e08f39ea8/actionview/test/fixtures/project.rb#L6

https://buildkite.com/rails/rails/builds/60609#d19181fb-fe80-4d1e-891c-1109b540fb4b/981-1009

The `collection_cache_key` is private API for a long time, but I've
maintained it in #35848 since it is mentioned in the doc
(https://github.com/rails/rails/pull/35848#discussion_r272011475).

The doc has removed at 1da9a7e4, so there is no longer a reason to
maintain that private API.

https://buildkite.com/rails/rails/builds/60601#81daf2ee-9259-42e5-983d-c57d66499b76/928-939

This follows #35595.

`where.not` with polymorphic association is partly fixed incidentally at
213796fb (refer #33493, #26207, #17010, #16983, #14161), and I've added
test case e9ba12f7 to avoid lose that fix accidentally in the future.

In Rails 5.2, `where.not(polymorphic: object)` works as expected as
NAND, but `where.not(polymorphic_type: object.class.polymorphic_name,
polymorphic_id: object.id)` still unexpectedly works as NOR.

To will make `where.not` working desiredly as NAND in Rails 6.1, this
deprecates `where.not` working as NOR. If people want to continue NOR
conditions, we'd encourage to them to `where.not` each conditions
manually.

```ruby
all = [treasures(:diamond), treasures(:sapphire), cars(:honda), treasures(:sapphire)]
assert_equal all, PriceEstimate.all.map(&:estimate_of)
```

In Rails 6.0:

```ruby
sapphire = treasures(:sapphire)

nor = all.reject { |e|
  e.estimate_of_type == sapphire.class.polymorphic_name
}.reject { |e|
  e.estimate_of_id == sapphire.id
}
assert_equal [cars(:honda)], nor

without_sapphire = PriceEstimate.where.not(
  estimate_of_type: sapphire.class.polymorphic_name, estimate_of_id: sapphire.id
)
assert_equal nor, without_sapphire.map(&:estimate_of)
```

In Rails 6.1:

```ruby
sapphire = treasures(:sapphire)

nand = all - [sapphire]
assert_equal [treasures(:diamond), cars(:honda)], nand

without_sapphire = PriceEstimate.where.not(
  estimate_of_type: sapphire.class.polymorphic_name, estimate_of_id: sapphire.id
)
assert_equal nand, without_sapphire.map(&:estimate_of)
```

Resolves #31209.

Cache versioning enables the same cache key to be reused when the object
being cached changes by moving the volatile part of the cache key out of
the cache key and into a version that is embedded in the cache entry.

This is already occurring when the object being cached is an
`ActiveRecord::Base`, but when caching an `ActiveRecord::Relation`
we are currently still putting the volatile information (max updated at
and count) as part of the cache key.

This PR moves the volatile part of the relations `cache_key` into the
`cache_version` to support recycling cache keys for
`ActiveRecord::Relation`s.

Performance cops will be extracted from RuboCop to RuboCop Performance
when next RuboCop 0.68 will be released.
https://github.com/rubocop-hq/rubocop/issues/5977

RuboCop 0.67 is its transition period.

Since rails/rails repository uses Performance cops, This PR added
rubocop-performance gem to Gemfile.

And this PR fixes some offenses using the following auto-correct.

```console
% bundle exec rubocop -a

Offenses:

activerecord/test/cases/connection_adapters/connection_handlers_multi_db_test.rb:212:26:
C: [Corrected] Layout/SpaceAroundOperators: Operator =
> should be surrounded by a single space.
              "primary"  => { adapter: "sqlite3", database: "db/primary.sqlite3" }
                         ^^
activerecord/test/cases/connection_adapters/connection_handlers_multi_db_test.rb:239:26:
C: [Corrected] Layout/SpaceAroundOperators: Operator => should be
surrounded by a single space.
              "primary"  => { adapter: "sqlite3", database: "db/primary.sqlite3" }
                         ^^
actionview/test/template/resolver_shared_tests.rb:1:1: C: [Corrected]
Style/FrozenStringLiteralComment: Missing magic comment #
frozen_string_literal: true.
module ResolverSharedTests
^
actionview/test/template/resolver_shared_tests.rb:10:33: C: [Corrected]
Layout/SpaceAroundEqualsInParameterDefault: Surrounding space missing in
default value assignment.
  def with_file(filename, source="File at #{filename}")
                                ^
actionview/test/template/resolver_shared_tests.rb:106:5: C: [Corrected]
Rails/RefuteMethods: Prefer assert_not_same over refute_same.
    refute_same a, b
    ^^^^^^^^^^^

2760 files inspected, 5 offenses detected, 5 offenses corrected
```

Currently the rollback only restores primary key value, `new_record?`,
`destroyed?`, and `frozen?`. Since the `save` clears current dirty
attribute states, retrying save after rollback will causes no change
saved if partial writes is enabled (by default).

This makes `remember_transaction_record_state` remembers original values
then restores dirty attribute states after rollback.

Fixes #15018.
Fixes #30167.
Fixes #33868.
Fixes #33443.
Closes #33444.
Closes #34504.

Before this fix, `touch` only clears dirty tracking for touched
attributes, doesn't track saved (touched) changes.

This fixes that tracks saved changes and carry over remaining changes.

Fixes #33429.
Closes #34306.

We can revert migrations using `change_column_comment` or
`change_table_comment` at current master.
However, results are not what we expect: comments are remained in new
status.
This change tells previous comment to these methods in a way like
`change_column_default`.

`add_to_transaction` was added at da840d13, but it should not be called
by except internal, since `remember_transaction_record_state` should be
called only once before saving.

And also, currently `add_to_transaction` doesn't always add the record
to transaction since da8de911, that is the reason hard to use that even
in internal.

Even if `add_to_transaction` ensure to add the record to transaction,
that is an internal concern, people don't need to explicitly call
`add_to_transaction`.

These two tests currently both define acceptance validators on the same
class, Topic. This means that in either one test or the other, there are
not one but *two* instances of the LazilyDefineAttributes module
builder in the class' ancestors, which can result in unpredictable
results.

Subclassing Topic in each test avoids conflicts.

Running this migration on mysql at current master fails
because `add_references_for_alter` is missing.

```
change_table :users, bulk: true do |t|
  t.references :article
end
```

This is also true for postgresql adapter,
but its `bulk_alter_table` implementation can fallback in such case.
postgresql's implementation is desirable to prevent unknown failure like this.

Somehow Code Climate is not working as expected for now?

This fix is necessary to test precision's default correctly.

`Truncate Tables posts` will also reset the `AUTOINCREMENT` this causes
a situation where if a test suite that uses the `taggings` fixtures runs
and subsequently the `test_truncate_tables` run, newly created posts
would reference the `tagging` in the database. This commit resest the
state of the posts table after the `connection.truncate` call in the
`test_truncate_tables`, as well as all other tests that call `trucate`
This ensures the associations and db state remain consistent after the
tests.

Fixes: https://github.com/rails/rails/issues/35941

If an association has a scope, `automatic_inverse_of` is to be disabled.
But extension block is obviously not a scope. It should not be regarded
as a scope.

Fixes #28806.

Regardless of a record isn't saved (e.g. validation is failed),
`after_commit` / `after_rollback` callbacks are invoked for now.

To fix the issue, this adds a record to the current transaction only
when a record is actually saved.

Fixes #29747.
Closes #29833.