perf: add direct binary A/B bench runner

[ianm199]

Summary

• add harness/bench/compare_bins.sh, a direct A/B runner for two arbitrary Lua interpreter binaries
• document when to use it: local Rust-vs-Rust packet validation without appending dashboard ledger rows

Why

The intern-retain packet needed a direct Rust A/B check between two built artifacts (/tmp/lua-rs-intern-baseline and /tmp/lua-rs-intern-skip). That was done with an ad hoc shell loop. This runner makes that workflow reusable for future small packets where reference-C timing noise is the wrong comparison.

The runner:

• accepts --a and --b binaries plus labels
• supports --runs N and comma-separated --workloads
• checks byte-identical workload output between binaries
• reports best-of-N wall time and max RSS
• writes ignored TSV/JSON artifacts under harness/bench/results/
• intentionally does not append harness/evidence/ledger.jsonl

Validation

• bash -n harness/bench/compare_bins.sh
• git diff --check
• self-compare smoke:
  • bash harness/bench/compare_bins.sh --a target/release/lua-rs --b target/release/lua-rs --label-a same_a --label-b same_b --runs 2 --workloads gc_pressure,binarytrees
  • output artifact: harness/bench/results/20260602T200046Z-bd1b0a3-bin-ab.tsv
  • JSON artifact parsed with python3 -m json.tool harness/bench/results/20260602T200046Z-bd1b0a3-bin-ab.json

Stacked on #128.