Move the benchmarks to scaling 0.9, and compare with the previous release

[droundy]

Summary

scaling 0.9 has no bench / bench_gen_env, so benches/bench.rs (the only user of scaling) is rewritten, as four binaries:

• versions: this release against the previous published one (internment_old, 0.8.6) at 1% accuracy, with old as the baseline of every group.
• controls: each release against itself, so there is a noise floor to judge versions by.
• libraries: Intern, NewIntern, ArcIntern, Intern32, arc-interner and plain Strings, to within 10%.
• contended: the old 100000-value workloads, one group per workload and contention level, one candidate per interner.

Run with cargo bench --features benchmarks --bench versions (README has a new section).

Design

• scaling is an optional dependency, not a dev-dependency. It needs Rust 1.81 and cargo test builds dev-dependencies, so a dev-dependency fails cargo +1.71 test --no-run. internment_old is optional too (a dev-dependency cannot be).
• Two features. bench is unchanged and stays a leaf (it only exposes benchmarking_only_clear_interns); the new benchmarks adds scaling, internment_old and arc-interner. This matters for later: once this version is "the previous release", the next one enables its bench, and that must not pull in scaling and the release before it, and so on down. The split has to ship with the first release that has these benchmarks.
• Each release then bumps the internment_old pin; plans/scaling-0.9.md describes that cycle and the reasoning.
• Fixes the old bench, which did not compile on master (bench did not enable arc-interner).

Behaviour changes in what is measured

• compare/hash is now lookup: the old two sets had nothing in common, so its removals never found anything. lookup does the same hashing and comparing read-only, which lets the sets be built once, untimed.
• sort sorts references (so a String baseline is not charged for copying) and starts from a fixed shuffled order; the old one got a random order from HashSet iteration.
• Intern32 is a new candidate wherever the others run.

Also

• src/intern32.rs: one #[cfg(feature = "deepsize")] on an impl whose bound named deepsize unconditionally, so --features intern32 compiles.
• CI: a job that runs cargo check --benches --features benchmarks on stable.

Not done / known

• The CI msrv job's cargo test very likely fails today from a fresh resolve (criterion 0.4 pulls in half 2.7.1, needing 1.81, and textwrap 0.16.4, needing 1.88), independent of this change; I checked with scratch copies but could not see recent CI runs.
• cargo check --no-default-features --features intern and --features intern32 already fail on master.
• controls separate the benchmark code from the library but do not build a second copy of the previous release as its own package.
• Numbers are not committed (machine-dependent). A short run of versions already shows differences from 0.8.6, e.g. Intern<String> sort about +12% and Intern<i64>::new random about -18%, which were not measured at full accuracy.

Test plan

• cargo check --benches --features benchmarks (stable)
• cargo test --all-features -- --test-threads=1 (stable)
• cargo +1.71 check, with and without --features arc, from a fresh lockfile
• each of versions, controls, libraries and contended (the latter filtered) run briefly and produce results
• a full contended run on a machine with reserved or unpinned CPUs

🤖 Generated with Claude Code

[codecov]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered ☂️