ci: cargo-test (required) is red on main since 83a6767ff — 33 perry-runtime typed_feedback/proxy lib tests fail; #6939 merged with the same 33 failing

[proggeramlug]

cargo-test is a required check and it is failing on main's current HEAD 83a6767ff ("GC memory: shape-key typed layouts (#6893) + major-GC pacing; don't link CoreFoundation in every binary (#6939)"). 33 perry-runtime lib tests fail — 32 in typed_feedback::tests plus proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof.

It came in with #6939, which merged red

PR #6939's own cargo-test run failed with exactly this set — 33/33 identical after sorting both lists. It was merged anyway. (Its lint was red too; that side is #6956.)

Reproduced on an unrelated PR that changes zero .rs files (#6954: shell script, TS test file, two text manifests, a workflow edit, docs). git diff origin/main...HEAD -- '*.rs' Cargo.toml Cargo.lock is empty there, so the failures are wholly inherited from main:

test result: FAILED. 1456 passed; 33 failed; 3 ignored; 0 measured; 0 filtered out

Sample assertions:

---- proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof ----
panicked at crates/perry-runtime/src/proxy.rs:2052:9:
assertion `left == right` failed: finite numbers are valid in verified ordinary JSValue typed slots

---- typed_feedback::tests::typed_feedback_class_field_get_guard_requires_raw_f64_layout_when_requested ----
panicked at crates/perry-runtime/src/typed_feedback/tests.rs:1660:5:
assertion `left == right` failed

Full list (sorted): proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof, and typed_feedback::tests:: — typed_feedback_class_field_get_guard_requires_raw_f64_layout_when_requested, typed_feedback_class_field_set_guard_fails_for_frozen_object, typed_feedback_class_field_set_guard_falls_back_for_class_setter, typed_feedback_class_field_set_guard_requires_raw_f64_value_and_layout, typed_feedback_closure_direct_guard_passes_and_rejects_bound_sentinel, typed_feedback_helper_return_guard_failure_returns_original_value, typed_feedback_helper_return_keys_use_shape_facts_not_sample_identity, typed_feedback_invalidation_counters_are_site_attributed, typed_feedback_method_direct_guard_fails_after_megamorphic_site, typed_feedback_method_direct_guard_fails_for_native_receiver, typed_feedback_method_direct_guard_fails_for_own_method_replacement, typed_feedback_method_direct_guard_fails_for_prototype_method_registration, typed_feedback_method_direct_guard_passes_for_exact_registered_method, typed_feedback_non_bounded_array_set_guard_failure_uses_jsvalue_object_fallback, typed_feedback_numeric_array_get_guard_requires_numeric_layout, typed_feedback_numeric_array_guards_reject_registered_class_ref_bits, typed_feedback_numeric_array_push_guard_rejects_mutability_restricted_arrays, typed_feedback_numeric_array_push_guard_requires_room_numeric_value_and_layout, typed_feedback_numeric_array_set_guard_requires_numeric_value_and_layout, typed_feedback_object_set_fast_falls_back_for_uncached_dynamic_key, typed_feedback_object_set_fast_hits_learned_dynamic_key_transition, typed_feedback_packed_i32_loop_guard_rejects_fractional_numeric_layout, typed_feedback_packed_u32_loop_guard_rejects_signed_fractional_and_overflow_layouts, typed_feedback_property_and_method_keys_ignore_receiver_identity, typed_feedback_registers_source_attribution, typed_feedback_roots_rewrite_shape_observations, typed_feedback_state_transitions_to_megamorphic, typed_feedback_trace_dump_honors_env_paths, typed_feedback_trace_json_includes_observed_kinds, typed_feedback_trace_json_reports_counts, typed_feedback_tracks_all_site_categories, typed_feedback_unboxed_numeric_write_falls_back_for_string_values.

Why it will keep hurting

Any PR whose ci_test_scope.py scope includes perry-runtime — which includes PRs that only add a test-files/*.ts gap test — now goes red for reasons unrelated to its diff. Combined with #6956 (lint red on the same commit) that is two of the required checks broken repo-wide.

The failing invocation is the scoped per-PR one, RUST_TEST_THREADS=1 cargo test --lib -p perry-runtime. Worth checking whether the full-workspace invocation (cargo test -p perry-runtime, all targets) still passes: if it does, feature unification across targets is masking these, and the nightly full run structurally cannot see the failure — which is how it would stay broken.

Given 83a6767ff reworked typed-shape/shape-key GC layouts and typed_feedback guards read layout bits, the guard tests asserting on raw_f64 / numeric layout are the plausible blast radius rather than a test-harness artifact.

Related: #6956 (lint red on the same commit), #6893, #6939.

[proggeramlug]

Narrowed it on an arm64 macOS box at 83a6767ff (identical result to CI: 1456 passed; 33 failed, so this is not Linux-specific). The 33 are two different defects, and my original "feature unification" guess in the issue body was wrong:

1. One genuine failure — proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof. Fails standalone:

$ RUST_TEST_THREADS=1 cargo test --lib -p perry-runtime -- --exact \
    proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof
test result: FAILED. 0 passed; 1 failed; ... 1491 filtered out
panicked at crates/perry-runtime/src/proxy.rs:2052:9:
assertion `left == right` failed: finite numbers are valid in verified ordinary JSValue typed slots

This one is a real regression, and the assertion is about typed-slot layout validity — squarely in 83a6767ff's (#6893 shape-key typed layouts) blast radius.

2. The other 32 are a test-isolation defect inside typed_feedback::tests. Each passes alone but the module fails as a group:

invocation	result
single test, --exact (e.g. typed_feedback_registers_source_attribution)	ok, 1 passed
single test, --exact (typed_feedback_state_transitions_to_megamorphic)	ok, 1 passed
cargo test --lib -p perry-runtime typed_feedback:: (49 tests, RUST_TEST_THREADS=1)	FAILED, 17 passed / 32 failed
full lib suite, RUST_TEST_THREADS=1	FAILED, 1456 / 33
full lib suite, default parallel threads	FAILED, 1454 / 35

So it is not cross-module pollution and not the proxy failure cascading — skipping the proxy test leaves exactly 32 failures. It is the typed-feedback registry being process-global and not reset between tests in the module. RUST_TEST_THREADS=1 does not help, because sequential execution still shares the process; and the parallel run failing more (35) is the usual signature of shared mutable global state.

Practical consequence: these are order/population-dependent, so a filtered run can look green. Anyone bisecting this should use the full --lib invocation, not a filtered one.

[proggeramlug]

Corroborating A/B from #6948, which tripped over this: I ran cargo test --lib -p perry-runtime at 83a6767ff in a clean worktree on an idle machine, and at a branch head whose diff does not touch typed_feedback.rs at all. Both report 33 failed, and the extracted failure lists match — 28 typed_feedback::tests::* plus proxy::tests::object_array_numeric_write_guard_requires_complete_uniform_proof.

So this reproduces off CI, on a quiet host, with no parallelism in play — it is not a runner-contention flake.

Caveat on my own numbers: my name-extraction regex had no digit class, so it compared 29 of the 33 by name (the packed_i32/packed_u32 family and two others were dropped on both sides equally). The 33 failed totals are from the summary lines and are directly comparable.