fix(data): adversarial hardening of trained DataClass classifiers (self-review)

[mdheller]

Adversarial + functional review of #264 (which Copilot didn't review)

Since #264 merged without Copilot's second pass, I ran an adversarial review myself and found three real rigor gaps (plus a latent validator bug the fix surfaced):

1. The monotone constraint was vacuous. The fixture held decimal_places constant, so its weight trained to 0 and the monotone teeth passed trivially. New fixture makes both monotone features (token_magnitude, decimal_places) the primary label signal → both weights bind strongly positive, accuracy realistic (<1.0), monotonicity genuinely exercised.
2. Train/test leakage — the standardiser was fit on full data → now train-split only.
3. Overclaimed manifest — said wide-and-deep/lattice but a linear monotone logistic was trained → renamed monotone-logistic (an honest realisation of the declared tf-lattice-wide-and-deep contract, not a full calibrated lattice).
4. Latent validator bug (surfaced by the now-non-1.0 accuracy): it compared full-precision recompute against the 4-decimal rounded stored value with a 1e-6 tolerance → per-class 0.7857 spuriously failed. Now compares at stored precision.

Validator strengthened: checks both per-class heads monotone AND asserts the constraint binds (non-vacuous). 8 checks; tamper teeth still refuses corrupted weights.

This is the adversarial pass #264 was owed.