Game balance audit: wire research effects, rework capability formula, fix dead systems

- Create researchBonuses utility to aggregate tech tree effects into all game systems
  (infrastructure energy costs, compute efficiency, training speed, model capability, reputation)
- Rework model capability from sqrt(compute) to 4-pillar formula (params + compute + data + research)
- Make context window affect benchmarks and inference speed
- Add MoE tradeoffs: 1.5x VRAM, 0.8x training speed
- Enforce research point costs as a gate for unlocking research
- Add real consequences to data contamination events (reputation hit, legal costs)
- Scale talent costs from $0.03 to $5/tick per headcount
- Scale compliance costs 100x to be meaningful
- Rework competitor acquisition: cheaper but grants headcount, RP, and reputation
- Remove dead code: sfxVolume, autoSaveInterval, notificationsEnabled,
  FAST_FORWARD_BATCH_SIZE, CHINCHILLA_OPTIMAL_RATIO

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

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

@@ -1,5 +1,6 @@
 import type { GameState, ComputeState, InfrastructureState } from '@ai-tycoon/shared';
 import { FLOPS_TO_TOKENS_MULTIPLIER } from '@ai-tycoon/shared';
+import type { ResearchBonuses } from './researchBonuses';
 
 export interface CapacityResult {
   totalFlops: number;
@@ -13,20 +14,19 @@ export interface CapacityResult {
   tokensPerSecondCapacity: number;
 }
 
-export function computeCapacity(state: GameState, infrastructure: InfrastructureState): CapacityResult {
+export function computeCapacity(state: GameState, infrastructure: InfrastructureState, researchBonuses?: ResearchBonuses): CapacityResult {
   const { totalTrainingFlops, totalInferenceFlops, totalVramGB } = infrastructure;
   const trainingAllocation = state.compute.trainingAllocation;
   const inferenceAllocation = 1 - trainingAllocation;
 
-  // Training hardware can do inference at ~50% efficiency
-  // Inference hardware can do training at ~30% efficiency (no NVLink, poor scaling)
   const effectiveTrainingFlops =
     totalTrainingFlops * trainingAllocation +
     totalInferenceFlops * trainingAllocation * 0.3;
 
+  const inferenceBoost = 1 + (researchBonuses?.tokensPerFlopBonus ?? 0) + (researchBonuses?.inferenceEfficiencyBonus ?? 0);
   const effectiveInferenceFlops =
-    totalInferenceFlops * inferenceAllocation +
-    totalTrainingFlops * inferenceAllocation * 0.5;
+    (totalInferenceFlops * inferenceAllocation +
+    totalTrainingFlops * inferenceAllocation * 0.5) * inferenceBoost;
 
   const tokensPerSecondCapacity = effectiveInferenceFlops * FLOPS_TO_TOKENS_MULTIPLIER;