C# impact: seven resolution gaps found by a 16-repo mutation oracle

[swapnilpaliwal-sd]

Symptom

On real C# libraries, impact <file:line> --tests missed most of the tests that fail when the method breaks, and selected nothing at all for a quarter of breakable methods; path <test file>.cs <method> refused every C# file endpoint.

Measured with a behavioural oracle on 16 open-source C# repositories pinned at tags at least six months old (xUnit, NUnit and MSTest; libraries, DI containers and extensions, parsers, mappers, mocking). For 50 methods per repository, sampled from the source with a Roslyn syntax walk (never from the graph), a guarded throw is inserted at entry, the repository is rebuilt and its suite run; the truth is the set of test files with a test that passed in both clean runs and fails now. 11 repositories were used for tuning; 5 were held out and scored once at the end.

Mechanism, one commit per pattern (each pattern seen in at least two repositories)

1. A name written inside a namespace is read from it. namespace A.Tests; using Results; imports A.Results, and Reflection.Info.Create() inside namespace A names A.Reflection.Info. Both were taken as written, so the using imported nothing and client types read as library types (boundary_lib), or the receiver stayed untyped.
2. path takes a .cs file as an endpoint. The file branch listed .ts/.tsx/.js/.mjs/.cjs/.py/.java only. It now takes every extension in ax_langs.BY_EXT; a new extension falls through to the name lookup when no file matches (Owner.cs can still name a method).
3. A using resource is typed by its initializer. using (var w = new Writer()) / using var w = …; have declarationKind USING, which var_value_initializer did not read, so every call on the resource had an untyped receiver.
4. The implicit base() call is an edge. A constructor without : base(..)/: this(..), or a class with no constructor at its new, runs the base's parameterless constructor. New synthesised kind implicit_base_ctor (bundle schema + ax_edges vocabulary updated).
5. An unstaged receiver keeps its type name. var services = new ServiceCollection(); services.AddX() and Configure(s => s.AddX()) with Action<IServiceCollection> gave call_recv_type_name nothing, so a client AddX(this IServiceCollection) matched neither by name nor on the loose DI rung (csharp: a this IServiceCollection extension called on a ServiceCollection receiver is labelled a boundary call and impact omits it #1472).
6. A static member read through its type name types the chain. Ctx.Current.Factory.Create(): the getter was an edge but its result had no type, because expr_type on a member access needed a typed qualifier and a type name has none.
7. A constructed generic type is a static receiver. Cache<int>.Get(), Registry<A,B>.Create(): the parser emitted the generic_name receiver with no written name. It now carries the identifier (arity stays in typeArgumentCount) and the engine resolves it at that arity; every receiver without type arguments keeps arity "0", as both clauses used before.

Each commit adds a tests/cases/csharp/ case that fails on the base and passes with the fix, with controls that pass on both (a file in an unrelated namespace, a foreach variable, : base(name) / : this(..) constructors, an in-source this type, a property named like a type, a non-generic type of the same name).

Numbers (base = apps/integration-0.1.9 a811ccf)

Tuning set, 11 repos, 550 mutants (443 with a failing test), cumulative:

step	micro recall	macro recall	precision	empty selections	every failing file selected	path found
base	0.462	0.546	0.400	116	181	0.000
+ 1, 2	0.473	0.583	0.403	95	200	0.470
+ 3	0.549	0.607	0.382	97	210	0.497
+ 4	0.553	0.607	0.384	96	211	0.498
+ 5	0.561	0.648	0.387	83	229	0.527
+ 6	0.618	0.674	0.373	80	246	0.557
+ 7	0.624	0.682	0.374	78	249	0.572

tests verb with the break applied: recall 0.494 → 0.641. No tuning repo loses recall at any step. Precision falls with the added reach (tests that call a method without failing when it throws). Selections for mutants no test detects: 849 → 859 cumulative; step 6 alone cut them 900 → 837.

Held-out set, 5 repos, 250 mutants (199 with a failing test), scored once:

	micro recall	macro recall	precision	empty	all selected	path found	context r@5
base	0.573	0.672	0.361	27	107	0.000	0.458
fix	0.604	0.701	0.339	21	114	0.693	0.483

One held-out repo loses one mutant's two files (0.724 → 0.712): the base selected them only through a [by name] lead at a call on a static singleton (Empty.Instance.ToText()), which step 6 now resolves correctly to the singleton's own override; the real route is lost earlier, at an out var receiver, on both.

Cost: forced full index (parse + solve), base vs fix interleaved on 10 repos: within ±1 s on 9, +2 s on one (17 → 19 s); graph.sqlite +0–2 MB; edges to client methods +0.1% to +27%.

Checks

• tests/run.py csharp 241/241, java 333/333, typescript 269/269, javascript 307/307, python 306/306
• graph/test/csharp/run-tests.sh (oracle self-tests, 22 engine cases, staging, remote edges) green, output identical to the base
• tests/hook_languages.py 7/7, tests/fastpath.py --lang csharp 8/8, parser csharp-tests.ts 62/62, gen_decls_all.py --check

Still open (not in this PR)

A library calling client overrides (GetHashCode/Equals as dictionary keys, GetEnumerator through foreach/LINQ), types created by reflection (Register<TMap>(), Activator), expression-tree mappers, an Outer.Inner qualifier resolved through an enclosing type, and path's optimistic closure ignoring boundary_lib sites (a client method mislabelled as a library call reads as "independent" instead of naming the site).