GeezerVison

…cess

The upstream pyperformance run_benchmark.py files will be vendored
unmodified under benches/pyperformance/ and executed by a Criterion
harness inside a RustPython interpreter. The one thing they need that
is not stdlib is `pyperf`; this module stands in for it.

Each bench_* method runs the benchmark body exactly once and does no
timing: repetition and measurement belong to Criterion, which runs the
whole script per iteration. parse_args goes through real argparse so
benchmark variants (extra_opts placed in sys.argv) take the upstream
code path. Only the API surface the vendored benchmarks use exists;
anything else raises AttributeError instead of silently measuring
nothing.

Tests run under both CPython and RustPython with unittest.

Assisted-by: Claude Code:claude-fable-5-1

Copies a fixed list of bm_* directories from a pinned pyperformance
release tag into benches/pyperformance/, verbatim. The release archive
is checked against a SHA-256 hash before extraction, unlisted bm_*
directories already in the tree are removed, and a few files that are
listed but unusable here (the asyncio_tcp SSL variant and its
certificates) are dropped. `--check` compares the tree with the archive
without writing and exits non-zero on any difference; CI runs it so the
vendored copy cannot drift from upstream.

Assisted-by: Claude Code:claude-fable-5-1

Generated by `python3 scripts/vendor-pyperformance.py`; nothing in this
commit is hand-written. 43 bm_* directories (65 benchmarks counting
variants) are copied unmodified from the pyperformance 1.14.0 release
archive, including their data files.

Assisted-by: Claude Code:claude-fable-5-1

CodSpeed now measures wall time on a CodSpeed Macro runner instead of
simulated CPU on a hosted runner, and what it measures is the vendored
upstream pyperformance programs, unmodified.

benches/pyperformance.rs is a new Criterion target that discovers the
benchmarks from pyperformance's own [tool.pyperformance] metadata, runs
each run_benchmark.py as __main__ in its own interpreter with the
variant's extra_opts in sys.argv, and does not link pyo3, so the built
executable can be moved between machines. tests/pyperformance.rs covers
discovery and runs two benchmarks end to end; it replaces
tests/benchmark_scripts.rs, which exercised the hand-ported files.

The 13 hand-ported pyperformance workloads in benches/benchmarks/ go
away in favour of the originals. Every one of the 73 upstream
directories is accounted for in benches/pyperformance/README.md:
third-party dependencies, start-up benchmarks, two that fail under
RustPython (decimal_factorial, sqlite_synth), one that needs child
interpreters (concurrent_imap) and three that take 13 to 46 seconds per
iteration (bpe_tokeniser, decimal_pi, asyncio_tcp_ssl) are excluded.

The workflow is split in two: `build` compiles the benchmark executables
on ubuntu-latest with `cargo codspeed build --measurement-mode walltime`
and uploads target/codspeed/walltime/; `run` on the codspeed-macro
runner downloads them and runs `cargo codspeed run`, which in walltime
mode also aggregates the results the compat layer writes. cargo-codspeed
is installed from its released musl executable (v5.0.3, matching
codspeed-criterion-compat 5.x) by .github/scripts/install-cargo-codspeed.sh.
scripts/codspeed-environment.py and its guard steps are removed: the
Macro runner is fixed hardware. The execution and microbenchmarks
targets stay available to `cargo bench` but leave CodSpeed, since they
link libpython through pyo3 for the CPython comparison. The CI lint job
gains a `vendor-pyperformance.py --check` step.

Assisted-by: Claude Code:claude-fable-5-1