Overhaul model system with multi-stage training, variants, benchmarks, and eval

Replace the single-stage training + flat capability score with a realistic AI
development pipeline: pre-training with Chinchilla scaling laws, SFT with
specializations, alignment with safety/capability tradeoffs (RLHF/DPO/Constitutional),
model families with distillation/fine-tuning/quantization variants, named benchmark
suite with compute-costing eval jobs, and segment-specific market quality.

Phases 1-6 of the model rework plan: new types, engine rewrite, save migration,
training events/risk system, concurrent training, variant creation, benchmark
evaluation with leaderboard, and market integration.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/game-engine/src/systems/competitorSystem.ts

@@ -43,7 +43,7 @@ export function processCompetitors(state: GameState): CompetitorState {
 
   const allCaps = [
     ...rivals.filter(r => r.status === 'active').map(r => r.estimatedCapability),
-    state.models.trainedModels.reduce((best, m) => Math.max(best, m.benchmarkScore), 0),
+    state.models.bestDeployedModelScore,
   ];
   const industryBenchmark = allCaps.length > 0 ? Math.max(...allCaps) : 0;