Architecture-specific vector programming remains a difficult frontier for language models: an implementation can satisfy surface syntax yet fail because of a hallucinated intrinsic, an invalid predicate shape, a sizeless-type misuse, or a subtle tail-handling error. This work studies NL2SVE, a function-level code generation problem that maps natural-language specifications and fixed C/C++ signatures to ARM Scalable Vector Extension (SVE) implementations. We introduce SVE-Loom, a model-agnostic generation-and-repair framework that operates within a bounded compile-and-run feedback session. SVE-Loom combines ACLE schema repair, execution feedback, and reference-assisted Regional Semantic Bootstrap when scalar semantics are available. From an initial prompt under a bounded feedback budget, wrapped models using SVE-Loom achieve higher success rates, producing more legal and correct SVE implementations across evaluation benchmarks. We further mitigate the scarcity of public SVE supervision by constructing a NL2SVE corpus from GitHub-mined SVE code, template-generated kernels, and SVE-Loom-guided API-model candidates. This third layer makes the corpus extensible: scalar references or capability gaps can be converted into NL-to-SVE examples, while candidate SVE implementations are retained only after schema checks, compilation, execution testing, speedup measurement, and human audit.