fix: scope field extraction to target class to prevent cross-class injection

[mashraf-222]

Problem

When the LLM generates optimization code containing inner/anonymous classes with fields, those fields are incorrectly injected into the target method's enclosing class. This produces uncompilable code — e.g., type parameters like Iterator<? extends F> injected where F is not in scope.

Observed in Guava PR #12 (Iterables.mergeSorted) where inner class fields were injected into the outer Iterables class.

Root Cause

_parse_optimization_source() in replacement.py (line 72) called analyzer.find_fields(new_source) without a class_name filter, extracting ALL fields from ALL classes in the generated code — including inner classes, static nested classes, and anonymous classes.

Fix

Pass class_name=target_method.class_name to find_fields() so only fields belonging to the target method's class are extracted. The find_fields() API already supported this filter via its _walk_tree_for_fields() implementation — it just wasn't being used.

The field extraction was also moved after the target method lookup (previously it ran before we knew which class the target was in).

Validation

• Bug reproduced: test creates an inner class with badField → it was injected into outer class
• Fix confirmed: badField no longer injected, only OFFSET (same class as target) is added
• All 33 replacement tests pass (32 existing + 1 new)

Test Coverage

• tests/test_languages/test_java/test_replacement.py — TestFieldInjectionClassFiltering::test_inner_class_fields_not_injected_into_outer

Closes CF-1087

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.