bench: heaptrack integration & baseline - #7

Summary

This PR adds a heaptrack-based benchmarking harness and CI workflow to capture performance and memory artifacts and compare them against a committed baseline.

What changed

• tools/bench/run_bench.sh: builds non-sanitized preset, runs test binaries under heaptrack and /usr/bin/time, produces artifacts/ (heap-.zst, heap-.txt, time-*.txt) and artifacts/current_results.json.
• tools/bench/compare_results.py: compares artifacts/current_results.json against benchmarks/baseline.json and exits non-zero when regressions exceed thresholds (defaults: 5% perf, 5% mem).
• .github/workflows/benchmarks.yml: CI job that installs heaptrack, runs the bench script, uploads artifacts, and publishes current_results.json.
• benchmarks/baseline.json: initial baseline committed from a local run (relative paths).
• .github/copilot-instructions.md: repository guidance for Copilot sessions (added alongside this work).

Notes & usage

• To run locally: chmod +x tools/bench/run_bench.sh && ./tools/bench/run_bench.sh (requires heaptrack and heaptrack_print).
• Compare locally: python3 tools/bench/compare_results.py [perf_pct_threshold] [mem_pct_threshold] (defaults: 5 5).
• Baseline is stored in benchmarks/baseline.json. Update baseline by running the bench script on a known-good machine and committing the new file.

Caveats

• CI runs the benchmark job on ubuntu-latest; absolute performance numbers vary between runners. Prefer comparing percentage changes and set thresholds conservatively.
• Artifacts/ contains raw heaptrack outputs (zst) and summaries; they are committed in this branch to provide an initial baseline and speed review. Consider moving large artifacts to CI-only storage or LFS if undesired in main.

Checklist

• Run bench locally and verify outputs
• Review thresholds in tools/bench/compare_results.py
• Decide baseline storage policy (in-repo vs artifact store)
• Consider gating PRs on perf/memory regression once stable

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com