land #8885: ECS command path, composed with #8876 - #8888

Lands #8885, rebased onto the main that now contains #8876 and #8872. All six conflicts resolved here rather than asking for a rebase.

The conflicts, and how they were resolved

Two were genuine design overlaps between #8885 and #8876, and both compose — neither side was discarded:

• builtins/arithmetic.rs — perf(ecs): guarded store follows forwarding edge, inline typeof/typed-array/subclass fast paths (wolf-ecs -16.5% / -20.9%) #8876 split typeof classification into a ValueTypeofTag enum with a js_value_typeof_tag fast path; perf: ECS command path follow-up to #8872 — −40% on the codehz/ecs 5k-entity sync row #8885 collapsed eight separate thread_local! slots into one perry_thread_local! array indexed by slot constants, removing a _tlv_get_addr resolution per call. Kept perf(ecs): guarded store follows forwarding edge, inline typeof/typed-array/subclass fast paths (wolf-ecs -16.5% / -20.9%) #8876's enum and classifier, adopted perf: ECS command path follow-up to #8872 — −40% on the codehz/ecs 5k-entity sync row #8885's array, rewrote the call sites to pass slot constants.
• expr/compare.rs — perf(ecs): guarded store follows forwarding edge, inline typeof/typed-array/subclass fast paths (wolf-ecs -16.5% / -20.9%) #8876 handles "exactly one operand statically proven numeric"; perf: ECS command path follow-up to #8872 — −40% on the codehz/ecs 5k-entity sync row #8885 handles "neither proven, but both dynamically numbers". Complementary, so the static-proof path runs first and falls through to perf: ECS command path follow-up to #8872 — −40% on the codehz/ecs 5k-entity sync row #8885's inline fcmp instead of the bare helper call.

Four resolved toward main, including one where #8885's side was simply empty (it predates #8858's frozen-array throw helpers).

Four follow-on defects surfaced and were fixed: a stale array_has_named_properties import (the merged header.rs only defines _resolved), a dropped AtomicBool/AtomicU8 import, a dropped local QNAN_PREFIX_I64 const, and — the substantive one — five statics that #8885 relocated to a new indexing_support.rs were restored in indexing.rs by hunk-level resolution, producing duplicate #[no_mangle] symbols. Hunk resolution cannot reconstruct a file-level reorganisation when both sides edit the same files.

The soundness test: the assertion was over-broad, not the code unsound

noncanonical_bitwise_method_does_not_gain_raw_number_truthiness failed. It is not pre-existing — it exists on main but not in #8885's base, so a clean rebase would hit it too.

Diagnosis: fcmp one had exactly one source when that test was written — the constructive-proof shortcut (ConstructiveMethodTruthiness::RawNumber), which is genuinely unsound for an arbitrary a & b return. #8885 introduces a second, unrelated source: a general dynamic-truthiness diamond that emits fcmp one only inside its own truthy.num block, after a bit test has proved the value is a plain untagged non-NaN double. There it is exactly correct, and truthy.slow still calls js_is_truthy for strings, BigInt, pointers, handles, int32 boxes and NaN.

I did not invent that distinction. type_analysis/numeric/tests.rs:796 already encodes it, with a doc comment stating it verbatim: "decides a plain (untagged, non-NaN) double inline with fcmp one — but only inside its truthy.num block… An fcmp one anywhere else is the unguarded numeric claim these tests forbid." The failing test's assertion was mirrored onto that helper.

The test's claim is preserved. It still forbids unguarded numeric truthiness, and its first assertion — that @js_is_truthy( must still appear — is untouched and passing. Only the mechanism of checking changed, to account for a second legitimate emitter. match_u32_bitset_test, the structural proof the test guards, is byte-identical to main and still rejects a bare a & b.

Validation (merged tree)

• all 30 lint-job gates pass
• perry-codegen 1322, perry-runtime 2749, perry-stdlib 124, perry-hir 348 — all 0 failed
• moving-GC arm (PERRY_GC_FORCE_EVACUATE=1 PERRY_GC_VERIFY_EVACUATION=1): 16 failed on the batch and 16 on clean main — same-commit A/B, none introduced
• df checked before and after; no result produced under ENOSPC

Performance claim needs re-measuring

#8885's −40% was measured without #8876, and both campaigns optimise the same ECS hot paths — compare.rs, lower_conditional.rs, array/indexing.rs, arithmetic.rs all took changes from both. Whatever this composed tree performs at, it is not that number. The functional result stands on its own; the figure should be re-taken on the bench mini before it is repeated.

Summary by CodeRabbit

• Performance

  • Improved ECS command-path performance, with benchmark results showing approximately 36% lower execution time.
  • Added faster handling for comparisons, truthiness checks, arrays, maps, closures, strings, and garbage-collection bookkeeping.

• Bug Fixes

  • Preserved correct array layout information when arrays are truncated to zero length.
  • Improved handling of stale array-growth references in maps.
  • Corrected type resolution for branded, imported, and generic type aliases.

• Reliability

  • Strengthened write-barrier behavior and expanded regression coverage for memory management and array operations.