Add game-simulation package with multi-run balance testing, fix stalled-pipeline trap

Adds a full simulation harness (game-simulation package) with greedy/random strategies,
36-metric diagnostics, multi-run orchestration via child processes, and a statistical
interpreter. Includes 2.3x engine performance optimizations (research bonus caching,
per-DC dirty tracking, reduced allocations in tick pipeline, single-pass loops).

Fixes a critical balance bug where training pipelines stalled on insufficient VRAM would
permanently block training slots — the engine never re-checked stalled pipelines, and the
greedy strategy didn't pre-check VRAM requirements. This caused 20-25% of seeds to get
stuck in Scale-up era. All three fixes (engine un-stalling, strategy VRAM pre-check,
stalled pipeline cancellation) bring pass rate from 75% to 100% across 20 random seeds.

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

apps/web/src/pages/ModelsPage.tsx

@@ -12,6 +12,7 @@ import {
   SIZE_TIER_MAP,
   SIZE_TIER_LABELS,
   SFT_SPECIALIZATION_BONUSES,
+  PRETRAINING_BASE_TICKS,
 } from '@ai-tycoon/shared';
 import type {
   ModelArchitecture, DataMixAllocation, SFTSpecialization, AlignmentMethod,
@@ -90,7 +91,7 @@ export function ModelsPage() {
   const [safetyWeight, setSafetyWeight] = useState(0.5);
 
   const trainingFlops = totalFlops * trainingAlloc;
-  const estimatedTicks = trainingFlops > 0 ? Math.max(30, Math.ceil(180 / (1 + trainingFlops * 0.1))) : Infinity;
+  const estimatedTicks = trainingFlops > 0 ? Math.max(30, Math.ceil(PRETRAINING_BASE_TICKS / (1 + trainingFlops * 0.1))) : Infinity;
   const estimatedCapability = Math.min(95, Math.sqrt(trainingFlops) * 5 + Math.log10(1 + totalData / 1e8) * 10);
 
   const activePipelines = pipelines.filter(p => p.status === 'active' || p.status === 'stalled');