Document compatibility with WOE-transformed features

[orgoca]

Problem

Most credit-risk feature pipelines emit weight-of-evidence transformed variables. CompileML says nothing about WOE anywhere — not that it is supported, not that it is discouraged, nothing. To a risk audience that silence reads as unfamiliarity with the field.

WOE inputs should work fine (they are just numerics), but there are real interactions worth documenting:

• WOE bins are already coarse — how does that interact with band construction and within-band AUC?
• Reason codes on a WOE feature explain the transformed variable; the dictionary text needs to describe the underlying attribute
• Monotonicity is often the point of WOE binning, which connects to the monotone-constraints issue
• Missing usually gets its own WOE bin, which connects to the informative-missingness issue

Acceptance criteria

• A section in docs/howto/tuning.md (or its own how-to) on using WOE-transformed inputs
• An example showing the reason dictionary written against the underlying attribute rather than the transform
• Honest note on any interaction we have not yet tested

Files

docs/howto/tuning.md or a new docs/howto/woe.md

Scope

Small, docs only — but it needs someone who has actually built WOE scorecards to write it credibly.