Creature Battle Card Game Naive Bayes tests

name: Creature Battle NB Tests overview: Replace the weak card-duel suite with a homemade Creature Battle Card Game rules oracle and several Naive Bayes test classes that train/predict on held-out boards with hard accuracy asserts (no feature leakage, no algo-vs-algo theater). todos:

• id: delete-duel

content: Delete NaiveBayesCardGameDuelTests.cs status: completed

• id: rules-helpers

content: Add CreatureBattleRules + feature encoders under Algorithms/CreatureBattle/ status: completed

• id: knockout-tests

content: Add KnockOut Gaussian + Bernoulli held-out grid test classes status: completed

• id: attack-edge-tests

content: Add AttackChoice + EdgeCase extensive test classes status: completed

• id: verify

content: Build and run Algorithms.Tests filter; fix any flaky floors status: completed isProject: false

Goal

Delete NaiveBayesCardGameDuelTests.cs and add a small Creature Battle Card Game test domain plus multiple extensive TUnit classes exercising Gaussian and Bernoulli Naive Bayes. Keep NaiveBayesTests.cs as the general contract/validation suite.

No third-party TCG IP / names — original types and terms only (Ember, Tide, Verdant, etc. if named at all).

Mini ruleset (test oracle only)

Shared pure helpers under tests/.../Algorithms/CreatureBattle/:

effectiveDamage = baseDamage
  + (hasWeakness ? 20 : 0)
  - (hasResistance ? 10 : 0)

canPay = attachedEnergy >= attackCost
KnockOut  iff canPay && effectiveDamage >= defenderHp
Survive   otherwise

Attack choice (two candidates A/B):
  legal = canPay
  pick highest effectiveDamage among legal; if neither legal → Retreat

Board fields used as inputs: baseDamage, defenderHp, hasWeakness, hasResistance, attachedEnergy, attackCost (plus a second attack’s base/cost for choice tests).

Feature encoding (honest):

Train on a discrete grid with held-out slices (e.g. train even defenderHp, evaluate odd; or hold out one energy band). Assert absolute correctness / accuracy floors — not “who won the duel.”

flowchart LR
  board[BoardState] --> oracle[CreatureBattleRules]
  board --> encG[GaussianFeatures]
  board --> encB[BernoulliFeatures]
  oracle --> label[KnockOut_or_Attack]
  encG --> fitG[GaussianNaiveBayes]
  encB --> fitB[BernoulliNaiveBayes]
  label --> fitG
  label --> fitB
  fitG --> assertG[Exact_or_accuracy_assert]
  fitB --> assertB[Exact_or_accuracy_assert]

Files to add

Under tests/Novolis.MachineLearning.Unit/Algorithms/:

Coverage targets (extensive)

• KnockOut Gaussian: train ~even HP × energy/cost/flags grid; evaluate odd HP; assert every held-out board matches oracle (or ≥98% if float noise — prefer exact with integer-valued doubles).
• KnockOut Bernoulli: same split; assert ≥90% on holdout (threshold features are lossy by design — document floor in test name/comment).
• Attack choice: enumerate pairs of attacks; assert Predict equals oracle on held-out energy values for both trainers where encoding allows; Gaussian should be near-exact with raw numbers.
• Edges: effectiveDamage == defenderHp → KnockOut; energy == cost - 1 → Survive even if damage would KO; weakness alone vs resistance alone vs both.
• API smoke in-domain: PredictScores probabilities sum ≈ 1 on a creature board; FeatureCount matches encoder arity.

Out of scope

• No production library changes under src/ (rules live in tests only).
• No ML.NET ClassicTrainers involvement.
• No algo-vs-algo winner enum / soft Tie asserts.

Verify

dotnet build d:\novolis\novolis-machinelearning\tests\Novolis.MachineLearning.Unit\Novolis.MachineLearning.Unit.csproj -c Release
dotnet run --project d:\novolis\novolis-machinelearning\tests\Novolis.MachineLearning.Unit\Novolis.MachineLearning.Unit.csproj -c Release --no-build -- --treenode-filter "/Novolis.MachineLearning.Unit/Novolis.MachineLearning.Algorithms.Tests/*"

Expect prior NaiveBayesTests + new creature classes green; duel class gone.