Good codes can be bad records: when coarse-graining destroys objective redundancy

A record is objective when many observers can each read it from their own small fragment and agree. We ask when that redundancy survives coarse-graining: if registers are repeatedly blocked and the checks carried forward, do bounded-weight readable fragments still exist at every scale? They need not. The canonical casualty is a sparse expander -- an excellent error-correcting code whose record becomes unreadable after two or three blocking steps. Good codes can be bad records. The property that fails, persistence, is logically independent of protection (can the information survive noise?) and of capacity (can enough disjoint fragments coexist?), and we prove a two-sided counting bound for it: readable redundancy dies at a computable depth when the fraction of blocks still admitting a bounded-weight readable star collapses (densification), and provably survives when that fraction and the maximum induced coordination both stay bounded. Between the two regimes packing decides, and we exhibit families on both sides. Stress tests (eight blocking schemes, six recovery rules, ~120 hosts, sizes to 10^5) show the kill mechanism is prescription-independent and size-stable, but the survivor class depends on the reading cap: at a fixed cap a three-dimensional lattice dies like the expander, and a cap sweep turns that observation into a cap-depth-dimension law, with the expander alone cap-proof. All host discrimination is relative to the stated readability rule. The quantum-gravity question that motivated all this is discussed only at the end.v2.3 (2026-07-05): adds the large-volume continuation of the cap sweep (l1b_large3d: cubes to 60^3, N to 216,000, caps to w=64; pipeline validated bit-identically against the published 16^3/22^3 rows). The cap–depth law's linear form survives at every size but its coefficient is size-dependent (w≈2K+6 is the 22^3 value; the bulk-3D cost is nearer w≈3.4K); coordination growth is measured non-saturating on high-trust data (the 22^3 rollover was finite-size); and the bulk-3D retained-density crossing also pins cap-invariantly, so the robust geometric-vs-arithmetic separator is the admissible-fraction cap-law slope plus the coordination growth mode, not the extensivity crossing alone. Theorems unchanged.