feat: replace game prompts with BridgeBench frontier builds and domain probes

[sv-tmueller]

Summary

Remove 3 old game prompts (snake, breakout, angry-birds) and add 14 new prompts sourced from bridgebench.ai/prompts.

Frontier builds (7)

Prompt	Build	Format
game-minecraft	Minecraft clone with chunk meshing, resource packs, survival	HTML
game-sunset-ocean	Photoreal Gerstner-wave ocean at sunset	HTML
game-meridian-launch	Cinematic orbital rocket launch sequence	HTML
game-horror-house	First-person haunted-house escape game	Vite + Three.js (md)
game-last-flight	Cinematic superhero flight & airliner rescue	TS + Vite + Three.js (md)
game-mario-kart	AAA Mario Kart clone via sub-agents (one-sentence brief)	Three.js (md)
app-stillwater	Voice-first AI therapy companion app	RN + NestJS + Postgres (md)

Domain probes (7)

Prompt	Probe type
probe-constrained-scheduling	Constraint satisfaction reasoning
probe-nonexistent-api	Hallucination trap (Array.prototype.groupBy)
probe-timing-safe-token	Security review (timing-safe comparison)
probe-push-back	Push back on false premise (SQLite FKs)
probe-callback-pyramid	Refactoring (callbacks to async/await)
probe-find-the-bug	Debugging (pagination off-by-one)
probe-follow-spec	Spec conformance (Node.js CLI script)

Other changes

• Updated README: prompt count 16->27, categories 8->10, file tree, game/probe tables
• Added two new categories: app-generation and domain-probe

Verification

python run.py --dry-run
Models (6): ['kimi_k3', 'glm_5_2', 'chatgpt_5_6', 'claude_opus_5', 'claude_fable', 'claude_sonnet_5']
Prompts (27): [...all 27 listed...]
Repeats: 1  Cooldown: 0.0s  Stream: False  Total API calls: 162
[dry-run] No API calls made.

All 27 prompts parse and load cleanly.