`Ecto.Type.cast/2` drops non-UTC DateTime offsets when casting to naive types

[lukaszsamson]

Elixir version

any

Database and Version

any

Ecto Versions

3.14.1

Database Adapter and Versions (postgrex, myxql, etc)

any

Current behavior

Ecto.Type.cast/2 has no explicit %DateTime{} handling for :naive_datetime, :naive_datetime_usec, or :date. These values fall through to the generic map handling which reconstructs a value from its wall-clock fields and implicitly discard utc_offset, std_offset, and time_zone.

Wall time preservation is consistent with DateTime.to_naive/1 but here it's undocumented and seems accidental. It is inconsistent with other ecto paths that either normalize or reject non-UTC.

Repro:

berlin = %DateTime{
  calendar: Calendar.ISO,
  year: 2020,
  month: 6,
  day: 1,
  hour: 0,
  minute: 30,
  second: 7,
  microsecond: {8000, 6},
  utc_offset: 3600,
  std_offset: 3600,
  time_zone: "Europe/Berlin",
  zone_abbr: "CEST"
}

Ecto.Type.cast(:naive_datetime, berlin)
# => {:ok, ~N[2020-06-01 00:30:07]}

Ecto.Type.cast(:naive_datetime_usec, berlin)
# => {:ok, ~N[2020-06-01 00:30:07.008000]}

Ecto.Type.cast(:date, berlin)
# => {:ok, ~D[2020-06-01]}

Ecto.Type.cast(:utc_datetime_usec, berlin)
# => {:ok, ~U[2020-05-31 22:30:07.008000Z]}

Other Ecto datetime paths establish stronger UTC expectations:

• Casting a non-UTC DateTime to :utc_datetime normalizes it to UTC.
• Loading a DateTime as a naive datetime requires its timezone to be "Etc/UTC".
• Dumping a DateTime directly as a naive datetime is rejected.

Expected behavior

Three solutions possible:

1. Normalize to UTC before removing timezone information
2. Reject non-UTC DateTime values, consistent with the stricter load/dump paths
3. Preserve local wall time intentionally, but add explicit %DateTime{} clauses, documentation, and tests establishing that contract

[lukaszsamson]

I did some digging. It seems the code never worked correctly.
Casting introduced in 12d6791 but omitted %DateTime{}
#1792 added generic map handling - the current bug. Non-UTC cases were nether tested nor discussed at the time
#2054 addressed offsets being silently discarded for :utc_datetime
#2798 added strict UTC checking to load/dump paths.

It seems the offset discarding was thought to be a bad idea at least #2054 (comment) and the conclusion was to either normalize to UTC or reject yet the code paths from this report was never actually addressed nor covered by tests