Notes. ^(a)Not directly available from Gaia DR3 variability tables. Computing rise_fraction requires phase-folded light curve analysis, which is not possible from catalog-level statistics alone. In our Gaia catalog pipeline, this concept is set to NaN for all sources and subsequently filled via per-class median imputation on training data; because no valid values exist for any source, per-class medians are also NaN and a subsequent nan_to_num step replaces every entry with 0.0, yielding a zero-variance column. An earlier version of our feature-extraction pipeline had instead assigned a 0.5–0.1 tanh(skewness) approximation to rise_fraction; on inspection this approximation carried no independent information (it is anticorrelated with skewness at r = −0.89) and all affected experiments have been rerun on the canonical NaN→0 dataset (Sect. 6.6). When full light curves are available (e.g., the OGLE cross-survey pipeline), rise_fraction is computed directly from the binned, smoothed phase-folded curve. ^(b) When full light curves are available, period_snr is computed as – log₁₀(FAP), where FAP denotes the false-alarm probability of a Lomb–Scargle periodogram peak (Lomb 1976; Scargle 1982) (higher values indicate greater period significance). For the Gaia catalog pipeline, where FAP values are not directly available, we introduced the proxy 2 log₁₀(period/σ_P), where σ_P is the catalog period uncertainty; this proxy is specific to the present work and is not a standard astronomical quantity. See Sect. 6.6 for a discussion of the implications. For classes without type-specific Gaia tables (DSCT/GDOR/SXPhe, ECL, MIRA/SR), pf_error is unavailable and period_snr defaults to a constant value of 5.0. ^(c)The NaN% column reflects values after per-class median imputation. Before imputation, R₂₁, R₃₁, and φ₂₁ are NaN for the three classes lacking Gaia type-specific Fourier tables (DSCT/GDOR/SXPhe, ECL, MIRA/SR; 9 000/18 000 = 50% of samples); after imputation these become 0.0%. For rise_fraction, all Gaia sources are NaN before imputation. Because no valid values exist in any class, per-class median imputation cannot fill them, so 100% NaN remains; a subsequent nan_to_num step replaces these with 0.0 before standardization. See Sect. 6.6 for a discussion of the implications.