[SamdaniSayam]

Summary

RClassField::ConstructValue() ignored failures from TClass::New(). For an abstract class, or a class without a default constructor, TClass::New() first JITs a constructor wrapper that fails to compile and then returns nullptr. RNTuple ignored that return value, so MakeField<T>() returned a non-null std::shared_ptr<T> to unconstructed memory. Destroying that value crashes in RClassDeleter. The failed JIT compilation is also the trigger for the Cling state corruption reported in #13910.

This PR:

• checks kIsAbstract and HasDefaultConstructor(kTRUE) before calling TClass::New(), throws RException with a useful message otherwise, and checks the TClass::New() return value;
• applies the same guard to RSoAField and RStreamerField;
• frees the value buffer in RFieldBase::CreateObjectRawPtr() if construction throws;
• constructs the value in REntry::AddValue() before recording the field name/type, so a failed MakeField() leaves the default entry/model unchanged;
• adds a roottest regression for abstract types, no-default-constructor types, partial-entry corruption, and continued use of the same model/interpreter.

There are no public API changes: the new behavior is a thrown RException where a broken object used to be returned.

Validation

• g++ -fsyntax-only passes for the modified translation units against 6.40.04 headers.
• Before the patch, MakeField<Abstract>() emitted the Cling error, returned a non-null pointer to unconstructed memory, and teardown segfaulted in RClassDeleter.
• After the patch, both the abstract and no-default-constructor cases throw RException; the same model still accepts std::vector<float> and a concrete derived class with an abstract base.
• With the guard but the stock REntry::AddValue, the regression's HasField checks catch the partially added field; with the reordered AddValue, they pass.
• The new roottest macro was run through ACLiC against 6.40.04 with the patched header/guard: all MakeField failure checks passed.

I could not run a full ROOT rebuild in this environment (8 cores, limited memory), so CI will be the first full-build run.

Related to #13910 (this addresses the RNTuple side; the Cling transaction/rollback discussion may still need a separate fix).

[dpiparo]

Did you have a look to https://github.com/root-project/root/blob/master/CONTRIBUTING.md#ai-assisted-coding-disclosure ?
Also, the configuration you mention is perfectly fine for building and testing root: could you please honour this requirement before we inspect the code?

[SamdaniSayam]

AI-assisted coding disclosure (per CONTRIBUTING.md):

• This PR's code was drafted with LLM assistance and has been thoroughly reviewed and understood by me; I take full responsibility for it.
• Validation done locally: g++ -fsyntax-only passes for the modified TUs against 6.40.04 headers; roottest regression run through ACLiC. Full ROOT rebuild was not possible in this environment (8 cores, limited memory) — CI will be the first full build.
• Re: the reviewer's note on configuration: I'll enable dev=ON (warnings-as-errors) and re-run the build/test before the code is inspected.