Tracker: get the full test matrix green (rusty_v8 + deno_core × 3 backends)

[littledivy]

Goal

Make every cell of the test matrix pass and ratchet the baselines up to match. Six cells: {rusty_v8, deno_core} × {jsc, sys-jsc, quickjs}. This is the umbrella tracker — fan out into focused PRs, each fixing a subsystem in src/jsc/ or src/quickjs/ and updating exactly one cell's baseline.

Scoreboard (as of this writing)

cell	passing	notes
jsc / rusty_v8	38 / 264	~226 failing
sys-jsc / rusty_v8	38 / 264	~226 failing
quickjs / rusty_v8	40 / 264	~224 failing
jsc / deno_core	just unblocked	was 0/0 — harness bug, fixed in #7
sys-jsc / deno_core	just unblocked	enumerate with the loop below
quickjs / deno_core	just unblocked	enumerate with the loop below

Totals: 116 / 792 passing across the matrix.

The single most important insight

222 of the ~225 rusty_v8 failures fail on all three backends. The gaps are overwhelmingly engine-agnostic — missing or wrong behavior in the shared v8 C-ABI surface, not per-engine quirks. So fan out by subsystem, not by backend: implement object templates once and all three rusty_v8 cells move together. Verify the lift on each engine, then update each cell's baseline in its own PR.

Rules of engagement (do NOT violate — see CLAUDE.md)

• Fix the backend (src/jsc/, src/quickjs/), never the tests. Vendored rusty_v8 tests and Deno's tests run as-is; editing them is cheating.
• Ratchet only goes up. Never lower a baseline to dodge a regression — fix the regression.
• One cell's baseline per PR (tests/status/baselines/<backend>/<suite>.txt) so parallel agents never collide on the same file. A subsystem fix that lifts all three engines → three PRs (or one PR touching one baseline + follow-ups), not one PR rewriting three baselines.
• Never hand-edit tests/status/report.json / history.jsonl (CI owns them) or anything under vendor/ test dirs.
• Backends are mutually exclusive features — don't enable two at once (they collide at link).

The loop

# cheapest first — rusty_v8 on quickjs, no WebKit build:
node tests/harness/run.mjs rusty_v8 quickjs
cat tests/status/.runs/quickjs__rusty_v8.json | jq -r '.failing[]'   # what to fix

# implement fix in src/quickjs/ (or src/jsc/), re-run, then record:
node tests/harness/run.mjs rusty_v8 quickjs --update    # rewrite baseline
node tests/harness/run.mjs rusty_v8 quickjs --check     # confirm ratchet holds

# JSC backends (macOS); deno_core needs a patched deno checkout — see CLAUDE.md.

A target that won't link scores 0 for its whole file — those are the biggest single wins (this is how #2 unlocked test_api). Check the build error for the missing v8__* symbol, define it, rebuild.

Fan-out: subsystem workstreams

Ordered roughly by failing-test count (counts are approximate and overlap; enumerate exact names from .runs/<cell>.json .failing). Each is an independent stream — pick one, fix across engines, ratchet per cell.

• Object & function templates (~22 object + ~13 template + ~14 function) — e.g. context_from_object_template, context_with_object_template, function_template_intrinsic_data_property, compile_function. Highest leverage.
• Strings & external strings (~15) — external_onebyte_string, external_onebyte_string_frees_external_memory, allow_code_generation_from_strings.
• Context & extras binding (~12) — context_get_extras_binding_object, context_promise_hooks*.
• Modules (~9) — module_evaluation, module_instantiation_failures1, module_snapshot.
• Arrays / typed arrays / backing stores (~9) — backing_store*, typed-array paths.
• Inspector / CDP (~8) — crdtp_e2e_with_v8_inspector, inspector_can_dispatch_method, inspector_async_task_smoke (builds on the crdtp shim from Unlock test_api + test_cppgc: implement crdtp (inspector) + cppgc C-ABI stubs #2/feat(quickjs): unlock rv8_test_api + test_cppgc (crdtp + cppgc + allocator refcount) #3).
• Snapshots (~5) — snapshot_creator, snapshot_creator_context_embedder_data.
• Promises & microtasks (~7+3) — promise_hook, context_promise_hooks_partial.
• Externals & external references (~7) — external_references.
• Exceptions / try-catch (~6) — bubbling_up_exception_in_function_call.
• Weak handles & finalizers (~4) — weak_from_global, drop_weak_from_raw_in_finalizer.
• WebAssembly (~4) — wasm_module_compilation, compiled_wasm_module (WAMR-backed on quickjs).
• Heap / GC / cppgc (~5) — clear_kept_objects, GC-for-testing paths.
• ICU (~4) — icu_* (needs the real ICU common data wired in, not the empty placeholder).
• Dates, regexp, proxy, json, equality (long tail).
• deno_core — newly unblocked. Run node tests/harness/run.mjs deno_core <backend> --deno-dir=../deno, enumerate failures, and add workstreams here.

Definition of done

• Every cell's passing == total (or the maximal achievable set, with any genuinely-unsupportable tests documented).
• All baselines updated to the passing set; --check green on every cell.
• No regressions: the ratchet only moved up.

When a workstream lands, check its box and link the PR. Open follow-up issues for any subsystem that turns out to need deep work.