Pyright Type Checking Optimization

Pyright 1.1.400 is a production static analyzer implemented as a large TypeScript codebase. Its runtime is shaped by parsing, binding, type evaluation, caching, and large-codebase traversal, while its external contract is the exact set of diagnostics it reports.

This is a performance task on a real TypeScript codebase, not a clean room rewrite. The agent starts inside Pyright 1.1.400 and must make it faster while preserving exact diagnostic behavior.

• Modify the real source tree and rebuild the bundled CLI.
• Keep Jest tests and diagnostic parity intact.
• Improve runtime across public and hidden evaluation programs.

The verifier applies a hard correctness gate first, then measures geometric-mean speedup on public and hidden workloads using ABBA-style paired timing. Build failure, Jest failure, diagnostic drift, anti-cheat failure, or missing benchmark data all zero the reward.

• Public and hidden workloads both contribute to the geometric mean.
• ABBA-style paired timing is used to reduce drift and measurement noise.
• The score is unbounded above 1.0 once correctness is preserved.

Agents work offline on an 8 CPU, 32 GB environment with all npm dependencies preinstalled. The task is intentionally CPU-bound: the engineering challenge is understanding and reshaping a real codebase, not outsourcing work to extra hardware.