string_length is inconsistent between atomic and non-atomic updates, for decomposed strings

[matt-beanland]

Code of Conduct

• I agree to follow this project's Code of Conduct

AI Policy

• I agree to follow this project's AI Policy, or I agree that AI was not used while creating this issue.

Versions

Elixir 1.20.2
Erlang/OTP 29 (erts-17.0.3)
ash 3.33.2
ash_sql 0.7.3
ash_postgres 2.13.1

Operating system

macOS 26.5.2 (Darwin 25.5.0) arm64

Current Behavior

With config :ash, :default_string_length_count, :codepoints, a max_length constraint is enforced by two different counts:

• non-atomic: Ash core evaluates in Elixir: codepoints_length/2, raw codepoints
• atomic: Ash.Type.String.apply_atomic_constraints/2 builds string_length(expr, :codepoints), which ash_sql renders as length(normalize(...)), so Postgres counts NFC-normalized codepoints

These may differ depending on the string.

Reproduction

repro - included below

Expected Behavior

Regardless of the string, or whether evaluated in the database or memory, the answer should be the raw codepoint of the value as stored.

[matt-beanland]

Here is repro after all.

In iex:

Mix.install([{:ash_postgres, "== 2.13.1"}, {:ash_sql, "== 0.7.3", override: true}],
  config: [ash: [default_string_length_count: :codepoints]])

Application.put_env(:repro, Repro.Repo, database: "repro", hostname: "localhost")

defmodule Repro.Repo do
  use AshPostgres.Repo, otp_app: :repro
  def installed_extensions, do: []
  def min_pg_version, do: %Version{major: 14, minor: 0, patch: 0}
end

defmodule Repro.Post do
  use Ash.Resource,
    domain: Repro.Domain,
    data_layer: AshPostgres.DataLayer,
    validate_domain_inclusion?: false

  postgres do
    table("ash_length_repro")
    repo(Repro.Repo)
  end

  attributes do
    uuid_primary_key(:id)
    attribute(:title, :string, public?: true)
  end

  actions do
    default_accept(:*)
    defaults([:read, :create])
  end
end

defmodule Repro.Domain do
  use Ash.Domain, validate_config_inclusion?: false
  resources(do: resource(Repro.Post))
end

Repro.Repo.__adapter__().storage_up(Repro.Repo.config())
Repro.Repo.start_link()
Repro.Repo.query!("DROP TABLE IF EXISTS ash_length_repro")
Repro.Repo.query!("CREATE TABLE ash_length_repro (id uuid PRIMARY KEY, title text)")

require Ash.Query
import Ash.Expr

# 7 graphemes either way. Stored decomposed it is 11 codepoints; recomposed
# to NFC it would be 7. Decomposing splits each accented character in two.

nfd = :unicode.characters_to_nfd_binary("ünïcödé")

Repro.Post |> Ash.Changeset.for_create(:create, %{title: nfd}) |> Ash.create!()

Ash.Query.Function.StringLength.string_length(nfd, :codepoints)
|> IO.inspect(label: "Ash core  -- want 11, got")

Repro.Post
|> Ash.Query.calculate(:len, :integer, expr(string_length(title, :codepoints)))
|> Ash.read_one!()
|> Map.get(:calculations)
|> IO.inspect(label: "via SQL   -- want %{len: 11}, got")

So will do a test in ash_postgres which will be red until fix in ash_sql is released.