Review finding 1: generic load/store instrumentation consumed replay prefixes before target atomic events. Resolution: atomic-only compile mode and a focused native replay test.
Review finding 2: full LLVM/QSYM symbolic instrumentation exposes the target threads to a global shadow map without a target-thread concurrency contract; LLVM 17 reproduced heap corruption. Resolution: separate schedule-only concrete probe artifacts from symbolic solving artifacts. The broader QSYM concurrent-shadow problem is not claimed fixed.
Review finding 3: LowerAtomicPass ran after probe insertion and converted target atomics to ordinary accesses. Resolution: schedule-only mode skips LowerAtomicPass; LLVM 17/18 IR tests require load atomic and store atomic.
Review finding 4: a second deferred atomic could silently advance the prefix while a commit was pending. Resolution: emit atomic_pending_conflict without advancing; campaign rejects it and commit mismatch.
Review finding 5: campaign verification trusted stored protocol counters. Resolution: verifier recomputes fallback, prefix fidelity, commit mismatch/conflict, event counters, trace digests, graph analyses, and topology.
Review finding 6: max graph candidates allowed one million combinations. Resolution: public API and campaign cap are 65536.
Review finding 7: the native campaign C fixture had no lit RUN line and caused a full-suite UNRESOLVED. Resolution: convert it into a preload/action/memory smoke; rerun both complete LLVM suites with zero unresolved.