l-diversity is distinct-only; document/optionally add entropy-l or t-closeness (skewness attack)

[cognis-digital]

Summary

deidproof implements distinct l-diversity (the count of distinct sensitive
values per equivalence class). It does not offer entropy l-diversity or
recursive (c,l)-diversity, nor t-closeness. Distinct l-diversity is known
to be vulnerable to a skewness attack: a class can have l distinct values yet
be 99% one value, still leaking the sensitive attribute in practice.

Why it matters

A steward who gates only on distinct-l can ship a class that is technically
"2-diverse" but effectively homogeneous (e.g. 49 Negative + 1 Positive).
The README/docs should either (a) state clearly that only distinct l-diversity
is computed and what it does/does not defend against, or (b) add an entropy-l
and/or t-closeness metric behind a flag.

Suggested direction

• Short term: document the limitation in docs/ADVANCED.md and the l-diversity
  docstring.
• Longer term: add --l-metric {distinct,entropy} to the CLI and a
  l_diversity(..., metric=...) parameter, computing Shannon entropy per class
  and reporting min entropy. Optionally add a t-closeness check comparing
  each class's sensitive-value distribution to the global distribution.

This is a genuine methodological gap, not a bug in the current (correct) distinct
implementation.