feat(AF-645): approval prediction — logistic regression trainer + evaluator

[ashish-babu-03]

Closes #648

Implements the mathematical core for the approval-outcome prediction
epic (#645): a pure-Java logistic regression trainer and model evaluator.
No Spring, no database, no external ML libraries — hand-rolled, matching
the precedent set by StatisticalAnomalyDetector.

What's included

• TrainedApprovalModel.java — immutable record holding intercept,
  weights, feature means/stddevs, and a predict() method that
  standardizes + applies sigmoid. JSON round-trip via Jackson 3
  (tools.jackson.databind).
• LogisticRegressionTrainer.java — stateless batch gradient descent
  trainer with L2 regularization, zero-initialized weights, and z-score
  feature standardization. Deterministic by design — no randomness
  anywhere.
• ModelEvaluator.java — stateless evaluator computing AUC via the
  rank-sum method (with average-rank tie handling) and accuracy at a
  configurable threshold.

Testing

All three classes have dedicated test classes per the coverage-parity
rule:

• LogisticRegressionTrainerTest — verifies >95% accuracy on a
  linearly separable synthetic dataset, exact bit-identical determinism
  across repeated training runs, L2 shrinkage effect, and standardization
  correctness (including constant-column edge case).
• TrainedApprovalModelTest — JSON round-trip fidelity, sigmoid bounds.
• ModelEvaluatorTest — known-answer AUC cases (perfect/inverted/tied
  rankings) and hand-computed accuracy.

mvn verify -Pcoverage: 10/10 new tests pass. ApplicationModulesTest
passes — all three classes live in ai/internal, no cross-module
dependency violations.

Note: ran without Docker, so Testcontainers-based integration tests
were skipped locally. Saw 9 pre-existing teardown errors unrelated to
this change (DefaultDriverCatalogServiceTest,
DefaultQueryEngineCatalogTest — JUnit extension context lifecycle
issue), flagging in case it's already a known issue.

[babltiga]

A few robustness/convention items to address before merge — details inline. The math itself checks out (verified the standardization, gradient, and rank-sum AUC against the spec in #648).

[ashish-babu-03]

@babltiga Thanks for the detailed review! I've addressed all the robustness and edge-case items.

Here is what was updated:

TrainedApprovalModel Serialization: Replaced the generic catch (Exception e) with catch (JacksonException e) and wrapped it in a new, public ApprovalModelSerializationException so the retraining jobs can properly catch corrupt DB rows as a sentinel.
Fail-fast map lookups: Updated predict() to strictly use .get() rather than .getOrDefault(). Added validation to the compact constructor to reject duplicate feature names and ensure weights, means, and stddevs are fully populated.
Array shape guards: Added explicit IllegalArgumentException checks to both LogisticRegressionTrainer.train and ModelEvaluator (calculateAuc / calculateAccuracy) to fail fast on length mismatches and ragged rows before entering any loops.
Test Parity: Added 9 new unit tests covering all the new exception branches, including the existing empty-input edge cases.
Everything is pushed and CI should be fully green. Ready for another look!