0138 — Lodestar.Preprocessing takes an edge on Lodestar.Stats for the normal quantile

Status: accepted · Date: 2026-09-16 · Applies: 0012, 0076, 0095, 0098, 0132

Context

0132 wrote the remaining scalers, and #763 put RobustScaler in scope with the options the reference carries. One of them does not fit in a package that depends on nothing.

RobustScaler(unit_variance=True) divides the interpercentile range by

Φ⁻¹(q_max/100) − Φ⁻¹(q_min/100)

— 1.3489795 at the quartiles — so that a normally distributed column comes out with a standard deviation of 1 rather than an interquartile range of 1. Φ⁻¹ is the normal quantile, and this repository already publishes it: Lodestar.Stats' Distributions.NormalQuantile, which 0098 published under 0095's rule.

Lodestar.Preprocessing was one of three packages with no inter-package edge at all. Adding one is not a detail: it changes what every consumer of a scaler restores, and tools/check_nuspec_dependencies.py asserts the graph rather than reviewing it.

Decision

Lodestar.Preprocessing takes an edge on Lodestar.Stats, and unit_variance is written against Distributions.NormalQuantile.

The edge is declared the way 0012 requires and Lodestar.Stats.TimeSeries already does: a PackageReference on the published floor, with a ProjectReference behind LodestarUseProjectRefs for the developer loop. The floor is Lodestar.Stats 0.4.0, which is published and has carried the member since 0.2.0 — the depended-on package ships before the consumer's next release, never after.

This is the fourteenth edge to become the fifteenth; tools/check_nuspec_dependencies.py's EXPECTED is the authority and CLAUDE.md's table follows it.

Alternatives rejected

• Write a second normal quantile inside Lodestar.Preprocessing. No edge, and a second implementation of a member this repository already publishes and tests against scipy. Two copies drift, and the reason to keep them apart would have to be written down here — there is none. 0076 bars an external dependency from a core package; it says nothing against depending on a sibling, which Lodestar.Fuzzy, Lodestar.Survival, Lodestar.Stats.Regression and Lodestar.Stats.TimeSeries all do.
• Ship #763 without unit_variance, with an equivalence row recording the gap and its own issue. Honest, and it leaves a published scaler one option short of its reference for as long as nobody takes the edge — which was the spec's recommendation before this decision, and is what it replaces.
• Take the adjustment as a caller-supplied double. It moves the quantile to the caller rather than removing the need for it, and no reference spells the option that way.

Consequences

• Anyone restoring Lodestar.Preprocessing now restores Lodestar.Stats, which is itself dependency-free — one package, no external dependency, and the same floor Lodestar.Survival and Lodestar.Stats.Regression already pull.
• Lodestar.Preprocessing's description no longer claims "no dependencies".
• The precedent this sets is the one to apply next time: a member another Lodestar package publishes is depended on, not copied. Where that would create a cycle, the member moves to Lodestar.Abstractions instead, as CsrMatrix did under 0071.
• unit_variance refuses a percentile of 0 or 100, which have no finite quantile; the reference divides by an infinity and reports a scale of zero. docs/equivalence.md carries that divergence.