Matching Decompilation as a Verifier-Guided Task for Human-Centered Coding Agents

Matching decompilation projects attempt to recover C or C++ source that recompiles into the exact same binary object as an original program. This creates a hard-mode but unusually crisp setting for coding agents: compilers, pinned build environments, and object-diff tools provide a repeatable local verifier, while human maintainers still enforce style, legality, and review boundaries. We present a verifier-guided harness that turns existing matching-decompilation repositories into durable agent goals with setup diagnostics, target ranking, prompt generation, run history, progress dashboards, and bounded variant search. The system is designed for human-centered supervision rather than full autonomy: it records evidence, exposes last-mile metrics such as matching prefix and first mismatch offset, and gates commits on measured verifier improvement. We argue that matching decompilation is a practical environment for studying verifiable, steerable coding agents and describe a path from local harnesses to community task brokers that let non-expert contributors support existing decompilation projects without increasing maintainer burden.