When single-lineage measurements fail to determine treatment order: sister dependence, branching extinction, and decision-focused assays

Measurements sufficient to describe a retained cell lineage need not determine which intervention minimises the survival of the whole family. We construct a finite-type branching source in which complete retained-branch histories have identical laws across a family of sister couplings, while two interventions applied for exactly the same durations have opposite preferred orders. An exact finite-duration formula identifies the decision boundary and shows that it is strictly positive, not at zero covariance. Two new ingredients then control recurring descendant division: the extinction probability is monotone in the descendant division rate, and a sensitivity reserve replaces the crude occupation bound. Together they certify the reversal at $h=\log\frac87$ for descendant division rates up to $\varepsilon=1/30$ instead of $1/500$, and they certify it for the entire range $0\le\varepsilon\le1$ at pulses of length $\log\frac{50}{49}$ rather than $10^{-6}$ — four orders of magnitude of pulse length — so that the phenomenon survives descendants dividing exactly as fast as the founder. A sister-agreement measurement repairs the missing information: forty independent division records give $95\%$ correct choice on a separated constant-treatment class, now throughout $\varepsilon\le1/4$, and exact binomial inversion gives full-class rules with an explicit unresolved output. We distinguish calibration needed for a shifted decision threshold from calibration-free sign information, price correlated read errors as an additive covariance reserve, and express residual uncertainty in excess-survival-risk units. The source is synthetic: its rate ratios, instantaneous readout and independence assumptions are stated explicitly, and the conclusion is a statement about an observation protocol, not about a clinically effective regimen.