[CaffeinatedCoder]

Closes #261

Problem

LeftJoin and RightJoin (.NET 10), and .NET 11's FullJoin, hand their result selector (or their tuple) the default of the side with no match. Over a value object, Result or [RequireCustomInitialization] struct that is an instance no factory made, and nothing reported it.

Change

• CMTK0009 (DefaultProducingCallAnalyzer) reports every overload of LeftJoin whose inner element, RightJoin whose outer element, and FullJoin where either element is a no-default struct, a warning per side, on Enumerable, Queryable and AsyncEnumerable. That includes .NET 11's tuple-returning overloads. The message names an Option of each element in memory (inner.Select(Option.Some)) and a nullable projection in a query (inner.Select(x => (OrderId?)x)). Warning, no code fix.
• An Option element of a join is not reported. Its default is None, and the advised rewrite produces exactly that element type, so reporting it would flag the way out. Recorded on CMTK0009: report LeftJoin, RightJoin and FullJoin whose unmatched side is a no-default struct #261, in the README and in design/analyzers.md.
• A join that does not bind (a lambda reading a same-project .Value, as before weaving) is read from its candidates: the elements from the sequences passed, and the candidate whose outer sequence the receiver is (Queryable's for a query, which the first test run caught giving in-memory advice).
• LinqJoins (new, internal): resolves the joins by symbol, maps arguments to parameters, and ranks candidates. CMTK0008's join check now uses it in place of its own copy (no behaviour change; its tests pass).
• Docs: the Analyzers README's CMTK0009 rows, design/analyzers.md, a changelog line in 2.4.0's "Changed" (warnings-as-errors builds can newly fail), claims K09-* in test/oracles/claims.json, and KB-059. KB-059 records the .NET 11 join surface, and that EF Core 11.0.0-rc.3.26522.148 translates the nullable projection under FULL JOIN on SQLite and throws without it. That was measured in a scratch net11.0 project outside the solution, built by the 11.0.100-rc.1 SDK and run on the 11.0.0-rc.1 runtime, the only ones on this machine.

Tests

• OuterJoinDefaultTests: a row per method, host and side, with and without a comparer, async key selectors, static and named forms, conditional access, and the unbound form. Controls: the side a join never fills, a class element, a nullable element, an ordinary struct, an Option element, both ways out, Join and GroupJoin. A theory over TypeArgumentKinds (79 kinds) uses NoDefaultKindTests.HasNoDefault as the oracle. FullJoin and the tuple forms compile against the real Microsoft.NETCore.App.Ref 11.0.0-rc.3.26522.148 through the verifier's ReferenceAssemblies, resolved from nuget.org as Net100 is; no stub.
• Earlier, end to end against the RC 1 SDK's own reference assemblies (a scratch net11.0 project with the built analyzer): the same forms reported and both ways out stayed silent.
• Verify the guard: each seed fails its tests and nothing else:
  • rule unregistered: 9 failures (4 positive facts plus the 5 no-default kinds);
  • unbound path dropped: the candidates test;
  • candidate ranking dropped: the candidates test;
  • Option exemption dropped: the silent test;
  • LeftJoin side swapped: 10 failures;
  • conditional access: dropping the name form fails both conditional rows, dropping the receiver fails the unbound one.
• Weaving review: discovery is by symbol, resolved once per compilation, and nothing is registered without TypeKit. It found conditional access (rows?.LeftJoin(…)) silent, now fixed with a row. A join reached as a method group converted to a delegate has no invocation and isn't reported.
• Analyzer tests 4114/4114 and the full solution through tools/BuildVerdict (SQL Server tests on azure-sql-edge, as AGENTS says for an arm64 Mac).

Follow-up: #273 (CMTK0008 doesn't list FullJoin among its joins).

🤖 Generated with Claude Code