[CaffeinatedCoder]

Closes #485

Stacked on #492 (#483). It relies on #483's guarantee that a string set column compares bytes, and it touches the same guard file.

Problem

On StringSet and StringSet<T>, PostgreSQL can't sort a set operation's result as memory does: it sorts text by the database collation (#479). So since #478, every set type but text got canonical Union/Add/Intersect/Except in SQL, while on text:

• Intersect, Except and Add ran in memory only;
• Count, == and != over a server-side Union threw.

#479 can change the order only in 11.0.0, so on 10.x this would stay for the line's whole life.

Change

• Translator: text Add/Intersect/Except now emit ValueSetSqlGuard.UnguardedAdd/UnguardedIntersect/UnguardedExcept (inside the NULL-over-NULL CASE), as Union already emitted UnguardedUnion.
• Walker: a value derived from any of the four may now feed cardinality, =/<> and NULLIF, besides the multiset uses it already allowed. GROUP BY, DISTINCT, ORDER BY, deduplicating set operations, IN and UPDATE setters stay refused, with the messages reworded for all four operations.
• Expander:
  • It builds the four unsorted: array_cat, array_cat with a one-element array, or a filter of the first array.
  • It rewrites cardinality(derived) as (SELECT count(DISTINCT x)::int FROM unnest(…)), NULL for a NULL set.
  • It rewrites =/<>/NULLIF with a derived side as mutual containment, a @> b AND b @> a. A stored-column = keeps = and its index.
  • It recognizes a derived value structurally, plus the walker's derived subquery columns. UnionNamer folds into it, because a separate renaming pass would rebuild the tree before the expander sees it.
• ValueSetsMemberTranslator no longer declines Count over a visible union; MayBeConcatenation is gone.
• No COLLATE "C": measured on Text value set operators follow a nondeterministic column collation: on a case-insensitive column, Contains("a") is true for {A} and Remove("a") deletes A #483, it costs a GIN index its use. Text value set operators follow a nondeterministic column collation: on a case-insensitive column, Contains("a") is true for {A} and Remove("a") deletes A #483 makes the column's equality byte equality instead.
• Docs: in the EF guide, Combine sets with Union is rewritten, plus both what-runs-where tables, Good to know, Rows written by other tools and the value set section. Also value-sets.md, getting-started.md, both READMEs, AGENTS.md (the intro and the text rule), architecture.md and testing.md. Changelogs (root, EF Core) under 11.0.0-preview.1 Added.
• Oracles:
  • translation-support.json: text Intersect/Except/Add are now server.
  • Claims E17–E20, E35 and E82 follow the new sentences.
  • New seeds text-set-count-counts-repeats and text-set-equality-compares-arrays.
  • union-equality-not-refused is retired: its refusal no longer exists, and the second new seed holds the rewrite.

Tests

• CanonicalSetOperationParityTests.Text_AnswersAsInMemory_WhereOrderAndRepeatsDontCount (new sweep).
  • Data: StringSet and StringSet<T>, every pair of subsets of a, 😀 and ｡ (the pair UTF-16 and code points order apart), stored canonical, out of order and twice, and with a NULL element.
  • Asked: Count on every row; ==/!=, the projection, containment and IsEmpty on the rows without a NULL element.
  • Refused: storing a result through ExecuteUpdate.
• UnionCanonicalityTests: the equality and Count cases now run and compare with memory, including through a UNION ALL column, a conditional and a coalesce. The ordering, grouping, DISTINCT, set-operation, IN and ExecuteUpdate refusals stay. SetIntegrationTests and EfCore11RetargetTests likewise.
• No-database: 13 tests that asserted a refusal now assert the SQL; DeclaredNullabilityTests treats the ARRAY(…) constructor as non-strict.
• Proved by reverting (tools/oracles.cs seeds):
  • leaving Count as cardinality fails the text sweep and 2 UnionCanonicalityTests;
  • rewriting equality only when both sides are derived fails the sweep and 4 UnionCanonicalityTests.
• Full suite green: unit, conventions, ASP.NET Core and EF no-database projects, and both integration projects on PostgreSQL 16 and 18.

🤖 Generated with Claude Code

[CaffeinatedCoder]

Merged into main as 40c81da ("Merge pull request #493 …", parents 3776504 and this PR's head 304c3d7). GitHub's merge call failed with an internal GraphQL error after it had written the merge commit, so the PR was never marked merged. Fast-forwarding the branch to main let GitHub close it, but as closed rather than merged, and its Closes #485 didn't fire. #485 is closed by hand with the same pointer. The 10.x port is #497 (PR #501).