feat: scenario inflator for 400K token eval scenarios

[BYK]

Summary

Adds a scenario inflator that pads existing eval scenarios to ~400K tokens with realistic filler content. This forces all baselines (tail-window, compaction, Lore) to actually compress, making the comparison fair.

Problem

Current eval scenarios are ~5-20K tokens — well within the 80K tail-window budget. Tail-window sees the full conversation while Lore has distilled it, making the comparison unfair. Real coding sessions are 200-400K tokens.

Solution

inflateScenario(scenario, targetTokens) injects realistic coding conversation filler between existing key turns:

• 15 filler templates across 4 categories: feature implementation, test output, refactor, debug
• Each template produces 2-4K tokens of realistic code content (full TypeScript files, test output, stack traces, etc.)
• Keyword exclusion prevents filler from contaminating test questions
• Proportional distribution across sessions based on current size
• Deterministic seeded PRNG for reproducible inflation
• Token-accurate targeting (400K ± 1%)

Example

import { inflateScenario } from './inflate';
const inflated = inflateScenario(msrScenario, 400_000, 42);
// Original: 5,928 tokens, 63 turns
// Inflated: 400,494 tokens, 2,083 turns
// Questions: all 12 preserved

Files

• packages/core/eval/inflate.ts (NEW, ~2000 lines — mostly template content)