[patch] Skip conversion methods that cannot produce the type, so a working Parse is used

[matt-edmondson]

Fixes #104

What changed

IsUsableConversionMethod accepted two kinds of method that can never work, and because FromString is tried before Parse, either one hid a usable Parse(string):

1. A method returning a different closed type. The return type only had to share T's generic type definition, so Box<string> FromString(string) was accepted for Box<int>. Every read then threw InvalidCastException.
2. An open generic method. It was accepted without being closed and was only closed later, in the converter's static initializer, with MakeGenericMethod(typeof(T)). When T broke a constraint, that threw TypeInitializationException and broke RoundTripStringJsonConverter<T> for the rest of the process.

FindStringConversionMethod now does the following:

• It closes each generic candidate over the type while searching, in the new CloseOverType helper, and drops it if closing throws. That happens when a constraint is violated or the method has more than one type parameter.
• IsUsableConversionMethod now requires the return type to be exactly the type.
• FindStringConversionMethod returns the closed method, so the converter uses it directly and CreateStringConversionMethod is gone.
• The catch (AmbiguousMatchException) block is removed. It could not be reached, and it carried a second copy of the loose return-type check. A stale comment in EdgeCaseTests that referred to it is updated.

The existing TestGenericClass<int>.FromString<TSelf> case still resolves, because closing it over the type yields a method that returns exactly that type.

Tests

Two tests in ConversionMethodPriorityTests cover the two shapes from the issue:

• BoxWithMismatchedFromString<int>: FromString returns the <string> closure
• CodeWithConstrainedGenericFromString: FromString<TEncoding> has a struct, ITestEncoding constraint

Both now deserialize through Parse. With the library change reverted, both tests fail, with InvalidCastException and TypeInitializationException respectively. With the change, the full suite passes (121/121, net10.0), and the library builds for all its target frameworks.

This PR is independent of #113 (stack-trace preservation). The two edit different parts of RoundTripStringJsonConverter.cs.

🤖 Generated with Claude Code

https://claude.ai/code/session_013JBDCsjuRez5zdzBcBJ7Y7

---

Generated by Claude Code

[matt-edmondson]

CI is blocked by runner availability, not by this change. In run 37363469055, ci / Classify repository (ubuntu-latest) waited 15 minutes for a runner and was cancelled before any step ran. That happened on the re-run (attempt 2). On the first attempt the job was ci / .NET / Discover Test Projects, and it was cancelled the same way. Either way, the .NET build and test jobs never ran.

#113 here and ktsu-dev/Sorting#65 show the same cancellation at the same time, so it looks like an org-wide Actions capacity or limit issue. Locally the full suite passes (121/121) on this branch. Once runners are available, CI needs to be re-run.

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud