TypeScript has rapidly become a popular language for modern web development, yet its effect on software faults remains poorly understood. This paper presents the first large-scale empirical study of bugs in real-world TypeScript projects. We analyze 633 bug reports from 16 popular open-source repositories to construct a taxonomy of fault types, quantify their prevalence, and relate them to project characteristics such as size, domain, and dependency composition. Our results reveal a fault landscape dominated not by logic or syntax errors but by tooling and configuration faults, API misuses, and asynchronous error-handling issues. We show that these categories correlate strongly with build complexity and dependency heterogeneity, indicating that modern failures often arise at integration and orchestration boundaries rather than within algorithmic logic. A longitudinal comparison with JavaScript studies shows that while static typing in TypeScript has reduced traditional runtime and type errors, it has shifted fragility toward build systems and toolchains. These findings offer new insight into how language design and ecosystem evolution reshape the fault profiles of large-scale software systems.

Given a failing test case, test case reduction yields a smaller test case that reproduces the failure. This process can be time consuming due to repeated trial and error with smaller test cases. Current techniques speed up reduction by only exploring syntactically valid candidates, but they still spend significant effort on semantically invalid candidates. In this paper, we propose a model-guided approach to speed up test case reduction. The approach trains a model of semantic properties driven by syntactic test case properties. By using this model, we can skip testing even syntactically valid test case candidates that are unlikely to succeed. We evaluate this model-guided reduction on a suite of 14 large fuzzer-generated C test cases from the bug repositories of two well-known C compilers, GCC and Clang. Our results show that with an average precision of 77%, we can decrease the number of removal trials by 14% to 61%. We observe a 30% geomean improvement in reduction time over the state of the art technique while preserving similar reduction power.