Compilation Contracts and Runtime Guarantees: How Structural Type Enforcement, Trace-Guided Repair, Numeric Format Registries, and Harness Governance Jointly Define a Falsifiable Framework for Software Correctness Infrastructure

This paper advances a candidate reading — explicitly heuristic rather than derivational — that a cluster of recent software engineering and programming language research converges on a shared structural pattern: correctness properties that are enforced *at the wrong layer* of a software stack are systematically bypassable, and the measurable cost of that mislocation is documented across credential leakage, budget overruns, numeric format divergence, decompiler reusability, and agentic harness failures. The thesis is not that these domains share a formal unification, but that each independently arrives at the same engineering prescription: push the enforcement boundary earlier in the compilation or deployment pipeline, make violations structurally inexpressible rather than merely detectable at runtime, and instrument the gap between what the contract says and what execution produces. The corpus spans cs.SE, cs.PL, and cs.AR preprints from May–June 2026. Five primary findings anchor the synthesis: (1) affine type ownership in Rust makes LLM-agent token-budget double-spending a compile-time error rather than a runtime race [corpus:arxiv:2606.04056]; (2) a fixed-point combinator in the Clef compiler carries dimensional and numeric-representation structure through MLIR lowering via categorical functors, making structural violations detectable during compilation [corpus:arxiv:2606.02854]; (3) an 84-format numeric catalog with bit-exact conformance vectors provides a vendor-neutral reference that makes silent divergence diagnosable rather than invisible [corpus:arxiv:2606.09686]; (4) trace-guided harness repair localizes failures to specific harness layers rather than applying broad prompt-level patches [corpus:arxiv:2606.06324]; and (5) SBOM tooling gaps show that component inclusion has no shared definition, making supply-chain security structurally unenforceable with current tools [corpus:arxiv:2606.02442]. Three supporting findings on undefined behavior in C/C++ [corpus:arxiv:2606.12064], governed harness mutation [corpus:arxiv:2605.27328], and semantic entropy for code quality [corpus:arxiv:2606.09800] extend the pattern. The primary falsification path: if enforcement-layer migration (from runtime to compile-time or from ad-hoc to registry-anchored) does not reduce the *rate* of the specific failure class it targets — measured against a regression suite or production incident catalog — the thesis collapses to a taxonomy, not a design principle. ---

Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename.

Cited arXiv preprints: 2605.21337, 2605.27328, 2605.27332, 2605.29490, 2605.31004, 2605.31520, 2606.01490, 2606.02442, 2606.02494, 2606.02854, 2606.04056, 2606.06324, 2606.06492, 2606.07314, 2606.07412, 2606.09686, 2606.09800, 2606.11076, 2606.11117, 2606.12064, 2606.12212