Label-free dust families converge; warmth fails high-recall retrieval

Can label-free nonlinear spectral families support useful dust-region candidate retrieval beyond raw warmth?

Eight masked Gaussian families per region, excluding test and calibration regions with buffers; fixed physical warmth prototypes, no class/70um input; calibration-only thresholds and fixed review budgets.

Eight analytic controls. Independent native/inverse-Gaussian replay checks473020 feature states,279379 training likelihoods,2575156 score/posterior scalars,189208 records and567606 curve groups; all94604 test clouds and20485 candidates.140-real-row streaming batch/metadata controls plus synthetic missing control.

All five fits converge100/178/140/112/116updates.20485 research candidates include2609 unknown stages;5%review4730 is contained in the threshold shortlist. Primary pooled88.91%precision/45.17%recall versus raw warmth88.77%/38.61% and temperature88.49%/44.55%. At5%review temperature98.48%/12.72% beats learned94.09%/11.73%. Every joint spatial gate fails; no downloads.

Next: Retain raw temperature benchmark; target cold/embedded misses using compatible native source-centered optional mid-IR/astrometric/morphology views learned without class loss. Freeze before fresh broad-survey or independently vetted evidence; use stellar/YSO/cloud bundle for bounded spectral/image/temporal refinement. Full robust goal stays active.

Actually label-free representation/rank, but labelled operating calibration and consumed,related catalogue-stage development references. No physical probability,global optimum,measurement-error/OOD guarantee,ownership/membership or birth. Distinct fold coordinate systems. Original checker name error and interactive plot abort preserved; registered correction/Agg figure do not alter science.