feat(arena): arena-backed hashconsing behind bumpalo feature (2.16x)

[elefthei]

Measurements

cargo bench --features bumpalo --bench arena_vs_arc, 200k instructions, min of 5 reps, same generated term DAG replayed through both consigns:

phase	Arc (HConsed)	Arena (BHConsed)	speedup
build	30.989 ms	18.200 ms	1.70x
re-intern	22.953 ms	11.956 ms	1.92x
traverse	15.506 ms	6.697 ms	2.32x
teardown	10.282 ms	0.002 ms	5411.73x
total	79.730 ms	36.855 ms	2.16x

The bench is a target in this PR, not a one-off: it is harness = false, required-features = ["bumpalo"], and asserts the wins (total ≥ 2.0x, traverse ≥ 1.5x, teardown ≥ 3.0x), so it exits non-zero if the arena path regresses. Numbers above are from one machine and are not the pass criterion — the exit code is.

Only traverse and teardown carry thresholds, because those wins are structural: traversal pushes/pops/inserts handles, which is several uncontended atomic RMWs per node on the Arc side versus a 16-byte copy on the arena side; teardown frees N Arc allocations plus N owned table keys versus dropping one POD HashMap and a few bump chunks. build and re-intern ratios are printed but not asserted — real, but smaller and more machine-dependent.

Summary

Adds BHConsed/BHConsign, an arena-backed counterpart to HConsed/HConsign, behind a new optional bumpalo feature. Values live in a caller-owned bumpalo::Bump; handles are Copy (&'bump T + u64 uid) instead of refcounted. The existing Arc path is untouched — this is an addition, not a change.

Two wins beyond raw speed:

• the consign keys its table on the interned reference itself, so it needs no T: Clone bound (the Arc consign clones every element to use as its key);
• cloning a term is a register copy, not an atomic increment.

Changes

• src/arena.rs — new module, compiled only under feature = "bumpalo". BHConsed<'bump, T> mirrors HConsed's impls (PartialEq/Eq/Ord/Hash on uid only, Deref, Borrow, Debug, Display) with Clone/Copy hand-written to avoid spurious T: Clone/T: Copy bounds. BHConsign<'bump, T, S> provides new/with_capacity/with_hasher/with_capacity_and_hasher, mk/mk_is_new, contains, len/is_empty/capacity, iter/consed_iter, arena, reserve/shrink_to_fit. Re-exports bumpalo so callers need not pin a matching version.
• Cargo.toml — optional bumpalo = "^3.16" + bumpalo = ["dep:bumpalo"] feature, added to the unstable_docrs aggregate so the module renders on docs.rs. rust-version untouched. New [[bench]] arena_vs_arc.
• src/lib.rs — #[cfg(feature = "bumpalo")] pub mod arena; plus re-exports, and a short crate-doc section. Mentions are plain backticks, not intra-doc links, so cargo doc stays clean with the feature off.
• src/test/arena.rs, tests/send_sync.rs — unit tests and a rayon test.
• benches/arena_vs_arc.rs — the bench above.

Caveats

Destructors never run. Values are bump-allocated, so any heap owned by T (String, Vec<_>, …) is leaked until the arena drops. This is forced rather than incidental: handles are Copy and carry 'bump, so no owner could drop a value earlier without leaving handles dangling. Documented prominently in the module header, with the recommendation to use arena-friendly payloads (&'bump str, &'bump [T] allocated via BHConsign::arena) or to stay on the Arc path when destructors matter.

Other differences from the Arc path, all documented:

• no weak refs, hence no collect/collect_to_fit — memory returns when the Bump drops or resets;
• no consign! analogue: Bump is Send but not Sync, so a lazy_static arena consign is impossible. BHConsign is !Send; individual BHConsed handles are Send + Sync when T: Sync;
• hash_coll::HConSet/HConMap stay HConsed-only. BHConsed::hash writes exactly one u64, so HashSet<Term, hash_coll::hashers::p_hash::Builder> works directly and is already fast. Generalizing those collections over HashConsed is a crate-wide refactor and deliberately out of scope here.

Verification

command	result
cargo test --workspace	pass, no warnings (feature off)
cargo test --workspace --all-features	5 lib + 2 integration + 11 doc, all pass
cargo clippy --workspace --all-features	0 warnings (note: #![deny(warnings)])
cargo doc --workspace --all-features / cargo doc --workspace	both clean
cargo fmt --check	clean
cargo bench --features bumpalo --bench arena_vs_arc	exit 0

Two module doctests double as the user-facing correctness proof: the lambda-calculus one interns Var(3) twice and asserts factory.len() == 3, v2.uid() == v3.uid() and std::ptr::eq(v2.get(), v3.get()), i.e. perfect sharing with each distinct term allocated exactly once. Note its term type deliberately does not derive Clone.

benches/traversals.rs is #![feature(test)] and still nightly-only; the new bench is harness = false and selected explicitly, so nothing here requires nightly and default target selection is unaffected.

[lenianiva]

Is there a reason that htis uses bumpalo rather than some other arenas in https://donsz.nl/blog/arenas/? For example, https://crates.io/crates/concurrent_arena is concurrent and can reuse memory.