Add more Processors

[RBendias]

• DropStypes
• NumericalProcessor
• EncodeHolidays
• EncodeRelativeTime
• OneHotEncode
• TargetDenormalize
• LabelDecode

[RBendias]

Self-contained coverage note: this candidate list is based on cross-model processor needs for near-term SDM targets (TabICLv2 / TabICL and KumoRFM / RFM), plus comparison/future targets (Google TabFM, TabPFN, Kumo-TFM-style workflows, TabDPT, CARTE, SAP-RPT / ConTextTab, pytorch-frame / TabPrep-style supervised encoders).

This list excludes functionality already supported on current main (327c5c2), matching by behavior rather than exact class name.

P0: remaining RFM + TabICL gaps

Name	Where / note
QuantileOutputReduce	RFM quantile head, TabICLv2 regression. Main already sorts 999 quantile outputs and applies target inverse; this gap is selecting/reducing the quantile axis to median/mean/requested quantile levels as an output policy.
RelationalNegativeSampling	RFM link-task workflows. Exact negative-sampling policy still needs pinning before turning it into a public method/API.

P1: additional Google TabFM / TabPFN / Kumo-TFM-style gaps

Name	Where / note
OneHotEncode	TabPFN. Low-cardinality categorical expansion; shape-changing schema-aware processor.
SquashingScaler / SoftClip	TabPFN. Robust squashing + soft outlier clipping; distinct from existing SigmaClip and Clip / QuantileClip.
KDI	TabPFN. Kernel-density-integral distribution transform.
NanToNum / MissingIndicator	TabPFN / TabDPT-style missing handling when exposed outside the model; sentinel-channel details may stay model-internal.
PostHocTemperatureCalibration	TabPFN / Kumo-style output calibration. Fit temperature on validation logits; separate from fixed SoftmaxTemperature.
ProbabilityBalance	TabPFN. Adjust probabilities using class priors/counts.
FeatureSubsample / PCAReduce	TabPFN / Google TabFM. Feature selection or basis reduction for wide tables.
FeaturePad	Google TabFM / TabPFN-style view assembly. Pad to model-required width.
SVDFeatures	Google TabFM / TabPFN. Add low-rank derived features.
PolynomialFeatures / FeatureCrosses	Google TabFM / TabPFN. Add interaction features.
FingerprintFeatures	TabPFN. Add hashed/fingerprint-style features.
RobustScale / RTDLQuantile	Google TabFM. Extra normalization choices beyond existing standard/quantile/power.
EncodeHolidays / EncodeRelativeTime	Google TabFM direction. EncodeDatetime calendar fields exist on main; holiday and relative-time features remain separate gaps.
RareCategoryEncode	Google TabFM. Map rare categories into stable sentinel buckets; distinct from unknown/missing handling already covered by categorical alignment.
CategoricalMask	Google TabFM. Per-view categorical mask metadata.

P2: remaining methods not already covered above

Name	Where / note
HardSigmaClip	TabDPT. Hard clamp at +/- n sigma; separate from TabICL soft sigma clip and RFM quantile clip.
MinMaxScale / Log1pScale	TabDPT normalizer choices.
Retrieval / BalancedSubsample	TabDPT. ICL context construction; row-changing view/driver policy, separate from relational neighbor sampling.
TextEmbed	CARTE, SAP-RPT. FastText / sentence-transformer cell embeddings; likely ingest/model-specific because it is carrier-changing and cache-backed.
StarGraphBuild	CARTE. Table-to-star-graph construction for text/semantic cell models; distinct from existing relational edge/index support.
TargetEncode / MeanEncode	pytorch-frame, TabPrep. Supervised categorical -> numeric encoding; needs fit(X, y, folds=...) / OOF policy to avoid leakage.
MedianImpute	Public sklearn/AutoGluon-style fallback; not strongly evidenced in the local FM docs.
RegressionBinDecode / BarDistributionDecode	TabDPT, TabPFN, SAP-RPT. Distribution/bin logits -> scalar or quantiles; related to but not identical to RFM/TabICLv2 quantile output reduction.

Notes:

• Several remaining items are not ordinary processors: output policies, graph/text ingest, supervised encoders, and row/context construction should probably live beside Recipe rather than inside a simple value-transform list.
• Shape-changing methods (OneHotEncode, SVDFeatures, PolynomialFeatures, etc.) need an output schema/lineage contract, not just Tensor -> Tensor.