This preprint presents TholusOS, a constraint-guided co-authoring architecture for translating LLM-generated creative intent into runtime-validated multisensory experiences. The system separates deterministic structural decisions from bounded AI-authored content through a system-owned experience skeleton and a hierarchical compilation process. A macro compiler defines targets, progression, emotional windows, required surfaces, and readiness evidence, while a micro choreography compiler maps creative moments to semantic anchors, modality lanes, temporal placements, and runtime cues. TholusOS further introduces family-aware progression and observability audits, mode-dependent defect escalation, typed timing contracts, proposal-only repair, human review, revalidation, and audience-specific readiness gates. The architecture is implemented for immersive environments coordinating projection, audio, lighting, participant phones, wind and low-frequency actuators, runtime scheduling, and operator workflows. The paper documents the implemented system architecture, its current maturity boundaries, and a reproducible evaluation plan. It does not claim completed comparative benchmarks, audience studies, operator studies, or hardware-in-the-loop validation. Publication status: Preprint, Version 1.0. This manuscript has not undergone formal peer review.