Edit Order Sensitivity in Language-Driven Executable 3D Worlds: A Benchmark for Cross-Layer Semantic-State Divergence

While natural language interfaces and Large Language Models (LLMs) have substantially simplified 3D scene creation, the temporal dynamics of multi-turn conversational authoring remain largely unexplored. When users compose iterative modifications, they intuitively assume semantic order commutativity: applying edit A then edit B should yield an executable world state equivalent to applying B then A, provided the two edits are designed to be mutually compatible. We study Edit Order Sensitivity (EOS) in one language-driven executable 3D authoring system using a fixed synthetic benchmark. Under naive sequential execution (C0), reversing instruction order yields non-equivalent terminal states in 57.0% of cases (114/200); 43.0% (86/200) are equivalent. Asymmetric instruction-satisfaction regressions occur in 35.5% of cases (71/200). These results concern the evaluated implementation and its internal state and satisfaction evaluators; they do not establish prevalence across other systems or validate visual evaluation metrics. To study this phenomenon, we report EO-Bench v2.5, a fixed, stratified synthetic benchmark of 200 cases across four design strata (S1 Direct, S2 Indirect, S3 Lexical False Positive, and S4 True Independence). The cases use 74 initial worlds and 68 unique edit pairs; 90% are domestic, educational, or workshop scenarios. We evaluate naive execution (C0), deterministic hash canonicalization (C_HASH_ONLY), dependency-aware lookahead execution (C4), and invariant-guided local repair (C5). Our findings establish that forcing convergence does not imply semantic correctness: while hash-sorting trivially achieves 100% state equivalence, it drops instruction satisfaction to 67.5% by executing causally inverted operations. Conversely, our dependency-aware lookahead policy (C4) achieves 92.0% state equivalence and 99.5% satisfaction, while our invariant-guided local repair pipeline (C5) resolves 100% of detected invariant violations (16 -> 0 cases) and restores 100% state equivalence with 99.5% joint instruction satisfaction. We evaluate geometric clearances via a Three.js spatial validator, examine depth and component ablations, and document methodological limitations including textual motif leakage and internal oracle coupling.