GC young-log rule-1 assertion aborts perry-stdlib tests on clean main: a writer of closure.dynamic_props publishes without note()-ing the key #11117

RUST_TEST_THREADS=1 cargo test -p perry-stdlib aborts on clean main with a GC young-log assertion. It is not the parallel-safety hazard CLAUDE.md documents — it is deterministic under single-threaded tests, and it reproduces with no PR applied.

The abort

panicked at crates/perry-runtime/src/gc/young_log.rs:173:13:
young log for closure.dynamic_props does not name 50252359400, which holds a
minor-relevant pointer: a writer of that table publishes without `note`-ing the
key first (see gc/young_log.rs rule 1)
thread caused non-unwinding panic. aborting.

Triggering test: readline::mod_tests::listeners_provider_roots_readable_snapshot_across_array_allocation.

It is pre-existing — three-arm A/B, not an assumption

Identical assertion, identical test, on clean main with none of the PR's code present. Deterministic, so it is not flake and not the shared-global races that per_test_global! is converting (#7672) — those present as a fixture precondition failure ("fixture must start proven…"), which this is not.

Two further failures in the same run, also reproducing on clean main in isolation:
crypto::random::tests::native_dispatch_pbkdf2_value_form_fires_callback and
crypto::random::tests::native_dispatch_random_bytes_value_form_fires_callback.

Why nobody has seen it in CI

cargo-test on #11036 aborted at perry-codegen, the 4th crate of 44, so perry-stdlib never ran there. Whether the sweep/full tiers reach it on Linux is worth checking — the reproduction above is on a macOS/aarch64 host, and the assertion may be host-sensitive.

Why it matters more than a failing test

The assertion is the young log catching rule 1: a writer published a pointer into closure.dynamic_props without note-ing the key first. That is the missing-write-barrier class. The instrument is doing exactly its job here — the alternative to this abort is a minor collection that does not trace the pointer, which surfaces cycles later as TypeError: value is not a function in an unrelated function and is invisible to every runtime GC probe.

So the question is not "why does this test abort" but "which writer of closure.dynamic_props publishes without noting" — and whether that writer is reachable from ordinary JS rather than only from this fixture.

Reproduce

RUST_TEST_THREADS=1 cargo test -p perry-stdlib --lib readline::mod_tests::listeners_provider_roots_readable_snapshot_across_array_allocation

Found while repairing #11036's CI; that PR is unrelated and its own failures are fixed separately.