Noise-aware flux families recover measurement support but no new candidates

Can signed flux uncertainties restore sparse stellar recovery without losing screening precision?

Eight label-free error-deconvolved families on four native flux/profile coordinates, propagated correlated noise and missing views; compatible E112 union with calibration-only interpretation.

Five analytic controls. Independent Jacobian/Gaussian replay compares 36,676 feature states, 146,704 covariance entries and 265,901 Gaussian scalars; all-source provenance/flags, candidate identities, reference counts and 249 calibration thresholds reviewed.

963 training sources; converges in114 updates. Flux support8,367/9,169; compatible support8,977. Nine of ten prior unknown stellar proxies become supported, zero newly selected. Registered precision floor permits no recovery improvement; disabling threshold preserves3,175 candidates and original development132/147 positive,8/151 other selections. No downloads.

Next: Retain E112 screen; improve sparse discrimination with additional independent context and reference coverage. Preserve failed branch, cloud compatibility and temporal refinement; freeze before further fresh validation.

Measurement support is not candidate recall. First-order transformed Gaussian noise, no profile uncertainty/systematic-error certification. Inherited class/target-conditioned population, supervised calibration interpretation, consumed E113 diagnostics rather than fresh validation. No adopted physical class or birth.