This study characterises the identifiability limits of individual insulation admittances in an ungrounded network from phase-to-earth voltage data alone. The Millman nodal balance, recast as a homogeneous real-valued regression, is analysed through singular value decomposition (SVD). The Millman ratio is scale-invariant, so the full admittance vector cannot be uniquely recovered. The SVD exposes instead a near-null subspace whose dimension and stability reflect the insulation state. A two-stage gate on zero-sequence excitation and angular diversity selects which observation windows enter the regression, and five spectral indicators are extracted from the admitted set: effective rank, null-space dimension, conditioning, tail-gap separation, and informative-window share. On a single feeder whose phase-A insulation resistance is swept across several decades, the tail gap widens, the admitted-window fraction grows and the near-null subspace progressively captures the true admittance direction. The onset shifts earlier under healthy-pair symmetry and requires at least two non-triplen harmonics. Cable attenuation, voltage symmetrisation and source-side zero-sequence contamination each impose hard limits that longer records alone do not remove. The framework is diagnostic and refuses to claim identifiability from data that lack the requisite structure.