feat: add @hyperframes/renderer turbo renderer package

[miguel-heygen]

Summary

• New @hyperframes/renderer package with high-performance turbo renderer
• Uses pipelined CDP with pipe transport for fast frame capture
• Streams frames directly to FFmpeg for concurrent capture + encoding
• Auto-detects GPU for hardware-accelerated rendering and encoding
• Includes all 6 experiment benchmarks and results documentation

Architecture

Standard Engine:
  Node.js → Puppeteer (WebSocket) → Chrome → sequential seek → beginFrame → disk → FFmpeg
  ~10ms/frame (6 workers), 8540ms total

Turbo Renderer:
  Node.js → Puppeteer (pipe) → Chrome → pipelined seek+beginFrame → FFmpeg stdin
  ~26ms/frame (1 worker, CPU), concurrent encode

Experiment Results (6 experiments)

Experiment	Speedup	Key Finding
CDP Pipelined	1.47x	Overlap seek IPC with beginFrame
Streaming Encode	1.05x	Concurrent capture+encode
Xvfb Framebuffer	0.28x	Dead end — no frame sync
Multi-Tab (4 tabs)	2.38x	Shared GPU context
Frame Skipping	1.22x	40% of frames are static
Half Resolution	2.09x	4x fewer pixels

Critical finding: Bottleneck is SwiftShader CPU rasterization. Production runs on m6a.12xlarge (CPU-only). Switching to g5.4xlarge (NVIDIA A10G, $1.62/hr — cheaper than current $1.94/hr) would give 10-30x speedup.

Files

• packages/renderer/src/turbo-renderer.ts — Core renderer with pipelined CDP
• packages/renderer/src/node-bridge.ts — Public API exports
• packages/renderer/src/check-renderer.ts — Smoke test script
• packages/engine/src/experiments/ — All 6 experiment benchmarks
• docs/superpowers/specs/ — Design docs and results

Test plan

• Run turbo renderer on CPU machine — verify produces valid MP4
• Test on GPU instance (g5.xlarge) with useGpuRendering: true
• Compare output quality (PSNR) against standard engine
• Benchmark on GPU instance to validate 10-30x speedup estimate

🤖 Generated with Claude Code