Migrator records or deletes a migration version when the migration fails

[lukaszsamson]

Elixir version

any

Database and Version

any

Ecto Versions

3.14.0

Database Adapter and Versions (postgrex, myxql, etc)

any (Reproduced with `EctoSQL.TestAdapter)

Current behavior

If a migration module does not implement neither the requested callback nor change/0, Ecto.Migrator.up/4 and down/4 raise Ecto.MigrationError but still modify the schema_migrations table. A broken migration applied during tests/development corrupts the state and requires manual intervention.

Repro:
run a broken migration

defmodule InvalidMigration do
  use Ecto.Migration
end

assert_raise Ecto.MigrationError, fn ->
  Ecto.Migrator.up(TestRepo, 13, InvalidMigration, log: false)
end

# Incorrect: version 13 was recorded even though no migration ran.
assert Ecto.Migrator.migrated_versions(TestRepo) == [1, 2, 3, 13]

Reason:
do_up/5 and do_down/5 return {:error, %Ecto.MigrationError{}}. In async_migrate_maybe_in_transaction/7, fun_with_status unconditionally calls SchemaMigration.up/4 or SchemaMigration.down/4.

Secondary issue:

With strict_version_order: true, up/4 currently checks version ordering only after do_up/5 returns. An out-of-order migration is applied and its version is committed before Ecto.MigrationError is raised:

assert_raise Ecto.MigrationError, fn ->
  Ecto.Migrator.up(TestRepo, 0, ValidMigration,
    log: false,
    strict_version_order: true
  )
end

# Incorrect: migration 0 has already run and is recorded.
assert 0 in Ecto.Migrator.migrated_versions(TestRepo)

This contradicts the documented behavior that strict ordering should abort the migration.

Expected behavior

Broken migrations should not affect schema_migrations table. Retrying after fixing the module should execute the migration instead of returning :already_up or :already_down. With strict_version_order: true, version ordering should be validated before executing any migration commands