Phase 3: Claude acted as a log-sifter to extract targeted leaked identifiers, uncovering a concurrency anomaly where a DISCARD path and an asynchronous UPDATE path race, causing objects to persist in native memory as 'Zombie Subjects.'

Tone and scope: The article outlines a concrete, actionable method for debugging JNI memory leaks and presents a general blueprint applicable to other systems with native components and cross-language interactions.

Phase 2: Due to performance costs of Valgrind, non-destructive core dumps (gcore) were captured at 30-minute intervals and analyzed to identify a native cache that accumulated entries.

Phase 5: Introduces 'Agentic Time-Slice Differential Analysis' as a repeatable debugging paradigm, combining differential snapshots, log analysis, and repository cross-referencing to automate detection of memory creep, deadlocks, and CPU issues.

Phase 1: Passive inspection revealed minor leaks during static inspection and limited runtime monitoring; intrusive monitoring crashed a high-throughput server.

Specifics: The leak involved a native cache in the C server holding pivot results for a 500,000-row grid, the problematic Subject_A lifecycle, and the exact race sequence that created Zombie Subjects.

Intro summary: An experienced QA engineer used Claude Code to identify and resolve a complex JNI memory leak causing linear native memory growth in a high-throughput system with a Java-C JNI bridge.

Key practical takeaway: The approach shows how to investigate native leaks without heavy instrumentation in production, using AI-assisted analysis, differential data, and targeted log correlation to detect race conditions across language boundaries.

Phase 4: Root cause discovery linked the race to a timing issue where a DISCARD operation clears a native cache while an UPDATE simultaneously instantiates a new Subject_A, leaving the original context invalid and the new object orphaned in native memory.

Architecture and symptom: The system comprises a Java module, a C server, and a JNI bridge; Java heap GC remains normal while native RSS grows, signaling a native leak at the JNI boundary.