Conformance regression on main since 2026-07-08 ~08:15 UTC: 30 untriaged gap failures (7/8 shards) — window 3a46815aa..d6635879c, misattributed to innocent PRs

[proggeramlug]

Every PR merge-ref built after 2026-07-08 ~08:15 UTC fails 7 of 8 conformance shards (shard 6 passes) plus compiler-output-regression, with 30 NEW untriaged gap failures — a mix of compile errors and output mismatches. Since main pushes don't run conformance, the failures show up misattributed on whatever PR runs next (first victims: #6115, a Dependabot console bump that only touches perry-container-compose, and the in-flight runs 28939125930 / 28940216860 / 28940322061).

Evidence for the regression window

• PR fix(runtime): arm64_32 ILP32 — box oversized thread_locals, guard GC pointer classifier #6118 (fix/arm64_32-box-oversized-thread-locals): push 1cf73c105 passed the full suite at 06:15 UTC (run 28921820312). A final push 1061eb4a5 landed ~08:15; its run (28927997601) started 08:15:18Z and the PR was merged 18 seconds later (08:15:36Z, d6635879c) without waiting — that run then failed 7 shards + compiler-output-regression. The final push's content was never conformance-tested before landing.
• Same merge-ref window also absorbed the 08:14 UTC batch: 3a46815aa (fix(runtime): cap object keys-array length at capacity in property walks (prevents multi-GB / spin on a corrupt keys length) #6101), b760be65c (fix(runtime): register the per-thread globalThis cache as a GC root #6103), 0f64e8a05 (fix(codegen): resolve private-member declaring class by unique id, not name #6104). Each was green individually against pre-08:00 main, so the culprit is either the untested delta between 1cf73c105 → 1061eb4a5 in fix(runtime): arm64_32 ILP32 — box oversized thread_locals, guard GC pointer classifier #6118 (prime suspect), or a semantic interaction within the batch.
• Independent confirmation: Dependabot PR deps(deps): bump console from 0.15.11 to 0.16.3 #6115 passed conformance on 2026-07-07 (pre-rebase), was auto-rebased at 08:19 UTC, and now fails (run 28928233767) with the identical failure signature (same shards, same gap tests). Its diff (console 0.15→0.16 in perry-container-compose, a crate the compiler doesn't link) cannot cause this.

Failing gap tests (union across shards, untriaged in known_failures.json)

Several of the compile-error tests (diagchannel, zlib_3285_params, …) need non-default feature-combo static-lib builds, so a perry-runtime/perry-stdlib compile error under specific --features combos would explain the "compile error" class while default-combo tests keep passing (and shard 6 stays green).

30 tests

test_gap_1831_yield_star_array               (compile error)
test_gap_arguments_object_3553plus           (compile error)
test_gap_1840_class_iterator_for_of_spread
test_gap_2021_json_stringify_grown_array
test_gap_2308_unrolled_loop_var
test_gap_2438_object_key_order
test_gap_2442_object_literal_accessors
test_gap_2792_2814_2794_2796_array_validation
test_gap_2823_2825_2822_promise_semantics
test_gap_5345_dup_static_field
test_gap_5588_object_assign_proxy_source
test_gap_5592_yield_star_value_get_abrupt
test_gap_array_callbacks_2797_2798_2799
test_gap_array_iterator_manual_next
test_gap_array_methods
test_gap_async_advanced
test_gap_class_field_named_stack_index
test_gap_class_ref_numeric_array_canon
test_gap_class_symbol_async_iterator
test_gap_closures
test_gap_fetch_response
test_gap_fn_prototype_own_keys_5024
test_gap_for_of_untyped_map_set
test_gap_generators
test_gap_http_overloads_3226plus
test_gap_json_stringify_replacer_tojson
test_gap_map_instance_reflection_4576
test_gap_mapset_iter_2831_2830_2856
test_gap_param_prop_array_index
test_gap_typed_array_proto_methods

Suggested next steps

1. Diff 1cf73c105..1061eb4a5 on fix(runtime): arm64_32 ILP32 — box oversized thread_locals, guard GC pointer classifier #6118's branch — that delta was never conformance-tested.
2. For attribution across the batch: byte-compare perry outputs across the four commits (CI node is constant, so perry-vs-perry comparison avoids node-version noise).
3. Check the shelled-out feature-combo builds (cargo build --release -p perry-runtime-static -p perry-stdlib-static --features <combo>) for the compile-error tests — likely a straight rustc error under a non-default combo.

Until fixed, expect every newly-rebased PR to fail conformance-smoke-complete through no fault of its own.

[proggeramlug]

Resolved on main. The specific 30-test regression in the window 3a46815aa..d6635879c passes again on current main — not via a single revert, but absorbed by the July 8–10 fix cascade culminating in the incremental-GC flip (PR #6243).

Evidence:

• test-parity/known_failures.json has zero diff since d6635879c, so this wasn't papered over by triaging the tests away — they actually pass now.
• PR feat(gc): incremental collection by default — bounded pauses, RSS parity, STW as escape hatch (#6180) #6243's conformance run was green on all 8 shards; none of Conformance regression on main since 2026-07-08 ~08:15 UTC: 30 untriaged gap failures (7/8 shards) — window 3a46815aa..d6635879c, misattributed to innocent PRs #6124's 30 named tests appear as failures. Spot-checks in fix(runtime): stringify/JSON of a fetch/Blob handle SIGSEGV (fixes #6240, #6241) #6244/fix(runtime): dispatch dynamic native-module namespace method calls (fixes #6230) #6238 runs show 8 of the originally-listed tests explicitly PASS (test_gap_1831_yield_star_array, test_gap_class_field_named_stack_index, test_gap_mapset_iter_2831_2830_2856, test_gap_array_iterator_manual_next, test_gap_class_symbol_async_iterator, test_gap_for_of_untyped_map_set, test_gap_generators, test_gap_map_instance_reflection_4576).

Closing as fixed.

Heads-up (separate, out of scope here): test_gap_zlib_4917_level is a distinct, currently-active untriaged failure showing up in shard 6/7 across recent PR runs (#6238/#6244/#6245/#6248), with no existing issue. Worth filing on its own.