Kronumos: Cost-Bounded Automated Program Repair via Context Surgery and POSIX Diff Re-Anchoring on SWE-bench Verified

Repository-level automated program repair (APR) is usually addressed with general-purpose agents that spend hundreds of thousands of tokens per issue on multi-turn tool use. We study the opposite end of the design space: a vertical, cost-bounded system built only for bug remediation. Kronumos couples a LoRA fine-tuned Qwen2.5-Coder-7B-Instruct (4-bit NF4) with the Tokenectomy Rust runtime, which provides Tree- sitter traceback pruning, linear-time automata-based secret redaction, and deterministic unified-diff re-anchoring with structural validation. Inference follows a blind single-turn protocol: no tests or program code are executed, and the model receives a file-localized context derived from the issue text. On all 500 SWE-bench Verified instances, Kronumos synthesized 475 candidate patches and emitted an empty patch for 25 (5.0%) through the structural validation gate (Zero Dirty Diff). Because of upstream Docker image availability failures (HTTP 404) in the official testbed harness, 80 of the 475 candidates completed container evaluation; the other 395 are scored as unresolved. Kronumos officially resolves 12 instances, establishing a conservative full-benchmark lower bound of 2.4% (12/500; 95% Wilson CI [1.4%, 4.1%]), and 15.0% precision on evaluated candidates (12/80; 95% Wilson CI [8.8%, 24.4%]). The average prompt is 3,361.0 tokens (a 91.3% reduction from the 38,412-token raw context), achieved at $0.00 marginal API cost on Kaggle Cloud GPUs. We present these results as an efficiency and safety baseline for vertical APR under a file-localized setting, and outline Kronumos v2, which adds bounded test feedback with anti- overfitting sentinels.