(3) #[concurrency::test] macro and thread::scope shim

[daniel-noland]

Summary

The capstone PR. Ship the test-time infrastructure that lets a single source file run under every backend the facade (in #1542) routes to.

Three pieces, plus the test suite that exercises them.

concurrency::stress + #[concurrency::test]

• concurrency::stress(body) dispatches to loom::model / shuttle::check_random / _pct / _dfs / direct call based on the active feature. The dispatch has a defensive _ => compile_error! arm: if a new model-checker feature is added to the exclusion list but no matching dispatch arm is wired up, the build fails loudly rather than silently skipping exploration.
• #[concurrency::test] proc-macro expands to #[test] plus stress(|| body). Under model-checker backends it emits a nested mod <fn_name> { mod concurrency_model { #[test] fn <backend>() { ... } } } so test reports / JUnit show the active backend as the leaf name (e.g. snapshot_observes::concurrency_model::loom). Default backend stays flat. #[should_panic] / #[ignore] thread through; async fn and fn(args) are rejected at parse time.
• Crate path is resolved at expansion time via proc-macro-crate, so consumers under any alias work without setup; integration tests inside this crate use extern crate dataplane_concurrency as concurrency; since cargo rejects self-deps.
• stress is re-exported at the crate root for the macro expansion but hidden from rustdoc (#[doc(hidden)]) -- users land on the macro.

thread::scope shim

• concurrency::thread::scope works on every backend. Std and shuttle re-export theirs; loom 0.7 doesn't have scope, so we ship concurrency/src/thread/loom_scope.rs: built on loom::spawn plus an Arc<Mutex<Option<T>>> keepalive that preserves the same drop-affinity guarantee std::thread::scope gives.
• Drop-affinity is enforced by an explicit take_payload on main during teardown, not by an Arc-count assertion. Loom's JoinHandle::join synchronises only on notify, which can fire before the spawned thread's wrapper has finished dropping its captures -- so the count-based check fires spuriously in some schedules. Manual take on main gives the same guarantee unconditionally.
• One mem::transmute to lift the spawned closure's 'scope to 'static (loom 0.7 has no spawn_unchecked); the keepalive trait object keeps its honest 'scope and ScopeInner<'scope> lives in ManuallyDrop so dropck doesn't constrain 'scope past scope()'s body. Per-site SAFETY comments at both the transmute and the ManuallyDrop::drop.

Tests

• tests/quiescent_model.rs replaces the loom-only quiescent_loom.rs. One source, every backend, via #[concurrency::test] + thread::scope.
• tests/thread_scope.rs -- contract tests for thread::scope (single/multi-spawn, drop affinity, nested-scope re-entry).
• tests/scope_property.rs -- bolero x shuttle conservation property over generated spawn/op plans.
• tests/stress_dispatch.rs -- stress invokes body once on default, more than once on model-check backends; plus #[should_panic] / #[ignore] attribute pass-through.
• tests/arc_weak.rs -- Arc / Weak contract (passes under every backend, plus loom-only quirk tests). File-level #![cfg(not(feature = "shuttle_pct"))] -- PCT can't run single-threaded protocol checks.
• justfile nextest filter: empty under loom (the archived binaries are already feature-filtered), shuttle substring under shuttle (matches both quiescent_shuttle and the new concurrency_model::shuttle leaves).

Documentation

• lib.rs gains a "Compiles under loom != exhaustively checked under loom" section listing each documented shim limitation (Weak strong-clone semantics, weak_count panic, RwLock upgradable_read, OnceLock ordering invisibility, Mutex::new const divergence) with the workspace consequence -- most notably calling out NAT's Weak::upgrade().is_none() patterns and why NAT isn't in the loom matrix today.
• sync/mod.rs gains a "Portability footguns the facade does not paper over" section.

PR 3 of 3 in the concurrency facade stack. Depends on #1542 -- sync facade + backends + workspace sweep. After this lands, the stack reaches feature parity with the original PR.

[mvachhar]

One question here, if ::stress enables loom everywhere (it currently doesn't because of Weak), aren't there many cases where the model checker will fail to terminate or consume too much time/memory? Again, no change needed if this is not currently and issue, but it is a concern.

[daniel-noland]

One question here, if ::stress enables loom everywhere (it currently doesn't because of Weak), aren't there many cases where the model checker will fail to terminate or consume too much time/memory? Again, no change needed if this is not currently and issue, but it is a concern.

I really have not seen that so far. The more common issue is just failing to compile under loom. If you give loom an appropriate domain then it finishes very quickly