The workspace cost of observation-based control of stochastic memories

SFP-M03 / v0.2. Preprint; not peer-reviewed. An observable resource profile need not supply safe access to the controls that realize it. We quantify this distinction for row-stochastic memories F = AB with a known cubic bipartite reference without four-cycles and a positive maximum-row L1 tolerance at most 0.001. Exact response and stationary pair laws are observed after commands. Continuous stochastic kernels act on routing rows and prototypes; protection holds throughout, with no free unprotected test query. The same interface realizes every continuous factor path when the factors are known. If h is the known-factor minimum peak and K the initial available width, the minimum peak of a policy safe on the whole observational fiber satisfies h <= H_obs <= max{h, K+1}. One protected row with slack measures all prototypes; one idle buffer creates slack when needed. For a positive saturated family, the full pair law determines exactly an atomic permutation fiber. A finite bound on all joint routing and prototype commands excludes a safe first experiment at width N, although exact calibration has known-factor peak N. Thus H_obs = N+1. A second saturated Fano fiber has the same calibration task and no overhead. The sharp worst-case overhead is one component slot; saturation alone does not classify it. Files: the unchanged manuscript PDF, its self-contained LaTeX source, and a public reproducibility archive containing the exact-arithmetic code and recorded results. Finite checks supplement the analytic proofs. AI assistance is disclosed in the manuscript. Related foundation: Temporal observability and resource-preserving reductions of stochastic memories (SFP-M02), https://doi.org/10.5281/zenodo.23466364. This is a separate work, not a new version of SFP-M02.