flaky: the_intrusive_free_list_round_trips_a_whole_cohort asserts a delta on a process-global counter under parallel tests

[proggeramlug]

Flaky: the_intrusive_free_list_round_trips_a_whole_cohort asserts a delta on a process-global counter

crates/perry-runtime/src/box/release_tests.rs:485 fails intermittently:

box::release_tests::the_intrusive_free_list_round_trips_a_whole_cohort
assertion `left == right` failed: second cohort allocates N cells
  left: 515
 right: 512

Observed 3 times tonight across unrelated branches while validating #8392, #8397 and
#8399, always as a single failure in an otherwise green cargo test --release -p perry-runtime --lib (2594 passed / 1 failed), and never reproducible in isolation — the
same suite reruns clean, and the test passes on its own every time.

Cause

The assertions bracket a global counter:

let (a0, r0, _) = box_release_stats();
let second: Vec<*mut Box> = (0..N).map(...).collect();
let (a1, r1, _) = box_release_stats();
assert_eq!(a1 - a0, N as u64, "second cohort allocates N cells");

box_release_stats() is process-wide, but cargo test runs the suite multi-threaded. Any
other test that allocates a box between the two reads inflates the delta. 515 - 512 = 3
stray allocations from a concurrent test, not a free-list defect.

The intent of the test is sound and worth keeping — it pins that all N cells come from the
intrusive free list rather than falling through to std::alloc. Only the measurement is
unsound under parallelism.

Suggested fix

Either serialize it against other box-allocating tests (a shared mutex / serial_test-style
guard), or make the accounting local — e.g. snapshot and compare free-list occupancy for the
cells this test minted rather than a global allocation counter.

Notes

• Not caused by perf(codegen): skip redundant async frame capture tracking #8392, fix(gc): rekey built-in closure metadata after moves #8397 or perf: cache prepared native archives for incremental links #8399; none of them touch box/release_tests.rs.
• Introduced with the test split in fix(runtime): release closure-visible async boxes #8303 (test(gc): split box release tests out of box.rs).
• Because it is a 1-in-~7 failure in a 2595-test suite, it will read as a random red on
  unrelated PRs and cost reviewers time attributing it — which is exactly what happened
  three times tonight.