rspack-perf-valgrind-goal

Use the run_local_valgrind.sh helper to build a pinned Docker image, prepare fixtures once with pnpm bench:prepare, then run the measure command with the benchmark target, exact filter, and output directory. It executes the benchmark binary under Callgrind and writes instruction counts to run-*.instructions files.

Provide the exact benchmark filter, threshold, and scope, then follow the per-round loop: make a scoped change, validate correctness, review the diff, commit, measure locally under Valgrind, and retain or revert based on the instruction count delta. Repeat until the target is met or the round limit is reached.

No. All performance decisions come exclusively from local Callgrind instruction counts in a pinned Docker environment. The --cfg codspeed flag is only a compile-time source switch exposing Valgrind client-request boundaries; no CodSpeed service, token, upload, or CI status participates.

No. Comparisons are only valid within the same native Docker platform, image ID, fixture directory, benchmark target, filter, and environment. Absolute instruction counts differ between linux/arm64 and linux/amd64, so base and candidate must be measured on identical frozen inputs.

The optimized benchmark build can receive SIGKILL with Docker Desktop's default 8 GiB memory allocation. Increase Docker Desktop memory to about 16 GiB, restart the engine, and resume; cached build volumes are preserved.

A neutral or slower round is reverted with a new follow-up commit rather than amending history. Correctness-only fixes may be retained even if performance is neutral, and their measured value becomes the comparison point for the next round.