## Summary `conformance-smoke` is a **required** status context and it fails **8 of 8 shards** on every recent PR. I pulled all eight shard logs from #8090's completed run and tallied the actual gate output. It is not flake, and it is not one cause — there are **23 status changes** in four distinct categories. CLAUDE.md warns that these shards are flaky and that a red shard should be re-run and A/B'd before being believed. **Eight of eight is not the flake signature**, which is why I went looking. ## The tally ``` 10 node_fail -> parity_fail 8 pass -> parity_fail 3 pass -> crash 1 pass -> compile_fail 1 parity_fail -> pass ``` ### 10 × `node_fail -> parity_fail` — previously invisible, now exposed ``` test_gap_4510_enum_forward_ref test_gap_moment_methods test_gap_backoff_options test_gap_prop_plan_cache_invalidation test_gap_cron_cronjob test_gap_ratelimiter_memory test_gap_dayjs_factory_arg test_gap_slugify_options test_gap_derived_param_props test_gap_enum_in_function_body ``` This is exactly the #6364 pattern CLAUDE.md documents: when the oracle cannot run a test it exits non-zero, the harness classifies it `node_fail`, and **the test is silently dropped from the gate rather than going red**. These ten were in that hidden state and are now running — so these are not new breakage, they are newly *visible* breakage. The shard header says so itself: *"STATUS CHANGES — including tests the oracle stopped covering"*. That makes them the least urgent of the four groups and arguably good news, but they must be triaged and recorded rather than left to hold the gate red. ### 12 × genuine regressions from `pass` ``` pass -> crash test_gap_6386_dataview_concat_regex_fastpaths test_gap_7238_i64_specialization_exactness test_gap_buffer_own_props pass -> compile_fail test_gap_zlib_4917_level pass -> parity_fail test_gap_array_methods test_gap_builtin_alias_construct_7524 test_gap_collection_foreach_member_receiver_thisarg test_gap_dynamic_builtin_construct_dispatch test_gap_new_globalthis_builtin_6726 test_gap_set_map_foreach_fused_receiver test_gap_specabi_reassign test_gap_sso_concat_string_index ``` These are the ones that matter. Three `pass -> crash` are the worst of them. `test_gap_specabi_reassign` is #8100, fixed by #8109 — so at least one of the twelve is already resolved, which suggests the rest are individually tractable. **I have not attributed these to a commit.** They are red with similar counts on every PR I checked, so they are almost certainly main-side rather than any one PR's doing, but that is an inference and someone should bisect rather than trust it. ### 1 × `parity_fail -> pass` — a fix that fails the gate `test_gap_iterator_helpers_2874` now passes and the snapshot still lists it as failing. The snapshot gates divergence in **either** direction by design, so an improvement turns the gate red until the snapshot is regenerated. Working as intended, but worth naming: part of why this context is red is that something got better. ## Why this matters beyond the tests themselves `conformance-smoke-complete` is one of seven required contexts. With it red, **no PR can merge without admin bypass**, which is the "required and always red trains bypass" failure mode CLAUDE.md calls out — and it is how #8100 (a silent wrong-answer miscompile in the shipped default) sat unnoticed. Its shard was cancelled early on recent runs, so nothing surfaced it. Tracking the required-context picture in #8092. The others: `lint` is fixed as of #8087; `cargo-test` fails on four ELF-only byte-identity tests; `parity`/`compile-smoke` report `SKIPPED` by design and are `continue-on-error`. ## Suggested split 1. Triage the 10 `node_fail -> parity_fail` and record them in `test-parity/gap_snapshot.json` with issue links — they were always broken and are cheap to classify. 2. Bisect the 12 genuine regressions. Three crashes first. 3. Regenerate the snapshot for `test_gap_iterator_helpers_2874`. 4. Only then is the gate meaningful enough to keep required. Reproduce with `./scripts/run_gap_tests.sh --shard N/8`.