[jkschneider]

RemoveInitMocksIfRunnersSpecified can remove initialization of a helper or another object that the enclosing Mockito runner/extension does not initialize. Restrict removal to an attributed literal this belonging to a managed test instance. Preserve unmanaged nested/anonymous owners, aliases, casts, null targets, and shared or escaping handles.

Keep declarations and cleanup consistent with retained initialization. Scan resolved field references across sources, including inherited reads and explicit static imports. Preserve formal parameters and resources consumed by other resource initializers. Track only field types that could receive the declared AutoCloseable result, retaining Object, incomplete, and generic types conservatively; apply that admission rule equally to declarations, references, and imports.

Validation:

• Normally attributed regressions reproduce unsafe cleanup on the original recipe. Actual original helper fixtures compile and pass, while old emitted sources leave mocks uninitialized and fail at runtime. Controls cover managed/unmanaged owners, formal parameters, dependent resources, cross-source/inherited reads, imports, Object/generic fields, and both source orders.
• All 50 focused tests pass without skips, including replay of the native compiled tests against the actual uploaded candidate JAR. The clean exact-head native module check reports 2,349 tests, zero failures/errors, and nine existing skips on ed2cf94. Independent review approves this source.
• Both uploaded artifacts use identical ordered byte hashes for all 71 runtime dependencies. All nine candidate production classes match the tested native bytes and load from the uploaded JAR; 709 other main class files are identical. Credential scans are clear.
• Twelve original/emitted fixture pairs compile, pass real JUnit/Mockito runtime checks on the captured JDK 21, and are idempotent, with zero candidate recipe errors. This uses the actual uploaded JAR with no production class-directory overrides; no JDK 8 runtime or full consumer application build is claimed.
• A bounded synthetic comparison preserves identical compiling/idempotent output across three input sizes. With 2,048 unrelated fields, recorded references decrease from 2,068 to 14. Single-observation scan time improves; requested-GC measurements are mixed and do not establish peak heap savings or the cause of the earlier remote heap failure.
• Complete paired comparison covers all 1,283 repository identities, with equal source-file counts for each. Changed-file rows decrease from 297 to 288, retaining nine whole files; every candidate changed-file key is a subset of baseline. Both runs have zero source errors. The final candidate completes without the earlier heap failure in this execution; its cause and peak-memory benefit remain unestablished. Final-head CGP CI passes (20261026212830-69sQ1), and the combined native integration check passes all 2,392 tests with zero failures/errors and nine existing skips.

This deliberately retains uncertain cases; it does not infer alias ownership or redesign unrelated lifecycle timing and integration discovery.

Addresses item F7 of #1211. The aggregate issue remains open.

Co-authored-by: Moderne Factory factory@moderne.io