Add objdump parser benchmark with aggregated JSON output

[partouf]

What

Adds a benchmark for the ObjDumpParser pipeline (fromStream + outputJson) plus tooling to run it repeatedly and aggregate run-to-run statistics into a JSON artifact, intended as the basis for a performance baseline.

How it works

• src/test/objdump_benchmark.cpp — a Catch2 BENCHMARK over four pre-captured objdump resource files of increasing size (example.asm, example_intel.asm, gcc12_sort_object_reloc.asm, gcc12_bin_fmt_O2_flto.asm). Each input is pre-loaded into memory so file I/O isn't measured, and re-parsed with a fresh parser per iteration. Tagged [!benchmark] so it stays out of the default ./asm-parser-test run — existing CI is unaffected.
• src/test/CMakeLists.txt — compiles the benchmark and enables CATCH_CONFIG_ENABLE_BENCHMARKING (Catch2 v2.13's built-in micro-benchmarking; no new dependencies).
• scripts/run-benchmarks.sh — runs the benchmark N times (default 5), writing one Catch2 XML per run.
• scripts/aggregate_bench.py — combines the per-run XMLs into a summary JSON: mean of per-run means, run-to-run stddev, CV%, min/max, and the full per-run list, with commit/ref/label/generated_utc metadata.
• .github/workflows/benchmark.yml — manual (workflow_dispatch) workflow mirroring build.yml's Conan/gcc-10/Release setup. Builds, runs the benchmarks, prints the summary, and uploads bench-summary.json + raw run XMLs as an artifact.
• .gitignore — ignores build-release/ and bench-results/.

Usage

cmake -B build-release -G Ninja -S . -DCMAKE_BUILD_TYPE=Release && cmake --build build-release
scripts/run-benchmarks.sh          # -> bench-results/bench-summary.json
# or: BENCH_RUNS=10 scripts/run-benchmarks.sh

Raw per-run data is also available directly as Catch2 XML:

./asm-parser-test "[objdump]" -r xml -o run.xml

Stability notes

Run-to-run variance on the large inputs is small (~2-3%), while example.asm (~150us) is noise-dominated (CV ~15%). The summary's cv_percent field surfaces this per benchmark — the large files are the trustworthy signal.

Next steps (not in this PR)

Once this runs on GitHub, the bench-summary artifact can be promoted to a committed baseline (e.g. bench-results/baseline.json). A follow-up can add a compare_bench.py + PR gate that fails on regressions beyond a threshold. Note GitHub-hosted runners are noisier than a dedicated box, so any gate should use a generous threshold and likely only the large inputs.

🤖 Generated with Claude Code