Frozen Inner Looping Is Brittle for Code Generation with BitNet b1.58

We show that frozen inner looping, repeating a selected decoder-block range T times inside a frozen model, is not a reliable accuracy knob for native 1.58-bit code generation. We test BitNet b1.58 2B4T with a source-preserving schedule patch, using ordinary one-pass inference (T=1) as a paired baseline on 742 tasks from HumanEval+, MBPP+, BigCodeBench-Instruct, and DS-1000. We compare two simple reinjection rules: a running average of repeated block outputs and a damped residual blend. Running-average reinjection is flat or worse on the main EvalPlus suites. Damped residual reinjection at T=2 changes 30 failures to passes and 19 passes to failures, increasing raw pass count from 207/742 to 218/742, but the aggregate effect is not statistically confirmed (paired exact p=0.152, 95% bootstrap CI [-0.27,+3.37] percentage points). At T=8, performance collapses to 66/742. Output comparisons show large divergence from T=1 and asymmetric pass-to-fail transitions at deeper repetition, suggesting perturbation rather than productive refinement.