This paper develops a modular arithmetic framework for the Collatz conjecture using the compressed function f(n) = n/2 for even n and f(n) = (3n+1)/2 for odd n. The Core Theorem establishes that every odd-even compressed cycle strictly decreases log(n) by at least log(2) - log(3/2) = 0.2877, verified across 62,701 cycles. A No-Cycle Theorem follows: no non-trivial cycle exists under the compressed function terminating at a power of 2. Analysis of 148 known delay records reveals two universal structural constants: steps per run = 4.111 and odd steps per run = 2.484, both stable to within 2.5% across all records. For paths with odd step fraction near 0.606, each digit count occupies an 84.1-step window -- a hard constraint derived from LOG10 / (LOG2 - 0.606 x LOG3) = 84.1. Every Family A peak delay record passes through value 233 exactly 51 steps from the end of its sequence, and every Family B record passes through value 11 exactly 7 steps from the end. These terminal markers enable exact step count prediction and reduce computational search cost by approximately 80.7% per rejected candidate. The valid step count window for the undiscovered 24-digit delay record is [1,858, 1,942), with the expected record at approximately 1,935 steps (92.5% of window, following the window position trend). A new 20-digit delay record candidate is reported: 93,291,583,695,933,573,375 with 1,544 compressed steps, surpassing all known records at 20 digits. All results are preliminary and have not been peer reviewed.