This case study demonstrates a large-scale architectural refactoring performed by an AI coding agent using a "specification-first" protocol on a 717,725-line TypeScript production application. The objective was to dismantle a core lifetime invariant—specifically, ensuring that streaming AI generations persist even after their associated UI panels are closed—without the use of a human code reviewer or a pre-existing test oracle. The agent followed a rigorous workflow consisting of formal specification, 14 refinement cycles to audit the specification against the source code, atomic implementation via a compile/test feedback loop, and 17 verification cycles to audit the implementation against the frozen specification. Throughout 31 total audit passes, the agent identified and corrected 201 defects. The refactoring impacted 189 files (with 31 new files added) and involved a total of 288 files across two commits, totaling 34,770 insertions and 16,422 deletions. Convergence was achieved when two consecutive verification passes yielded zero findings. The entire process was completed in three days at a cost of USD 2,430, resulting in a system that behaved exactly as specified with no observed bugs during testing.