Structure-Survival Tool-Space Compression: Causal Repair of Hard Dependency Injection in Tool Retrieval

This work investigates structure-survival tool-space compression for tool retrieval and analyzes a failure mode caused by hard dependency injection. Across 5,114 held-out TaskBench tasks from three domains and four retrieval budgets, hard dependency insertion systematically degraded complete tool recall by displacing relevance-ranked candidates. Replacing hard insertion with soft dependency prioritization improved performance over the hard-dependency parent in all 12 evaluated domain-budget conditions, including rescue-disabled comparisons. Paired bootstrap confidence intervals and McNemar tests support that this repair effect is not attributable solely to selective rescue. Selective rescue provides an additional, separately measurable benefit. However, the incremental value of dependency information itself is conditional: soft dependency prioritization does not consistently outperform a no-dependency baseline across all settings. The study distinguishes 17,331 raw records, 17,044 accepted records, and 5,114 held-out test tasks. A dependency-overlap audit found 99.82–100% train-test edge overlap in the evaluated IID setting. Therefore, out-of-distribution dependency-graph generalization is not claimed. The accompanying reproducibility package contains the submission manuscript, frozen v1.4 validation runner, complete evidence archive, README, and SHA-256 manifest. Important scope boundaries: - The DTDR comparison uses a paper-based reimplementation/proxy and is not an author-exact reproduction. - Downstream FSA/SR is not verified. - Generalization to unseen dependency graphs remains unverified. - The results support a mechanism-level correction rather than a claim of universal superiority of dependency-aware retrieval.