feat(core): Slice 0 — core-runtime ↔ ReferenceRunner parity probe (th…

[cukas]

…in adapter)

The first cross-check of KERN's two disjoint TS interpreters: core-runtime (the product tree-walking interpreter) and ReferenceRunner (the differential oracle). A THIN adapter reduces both legs to the same CoreFixtureValue through EXISTING serializers (kernValueToCoreFixtureValue on core; the assign-event value on reference) — no invented semantics — proving the parity gate is viable for the scalar surface (the class/value-model unification stays deferred).

The gate immediately earns its keep: it surfaces a BIDIRECTIONAL divergence map across 4 honest buckets —

• AGREE: arithmetic, same-type comparison, concat, short-circuit, templates, String() coercion, -0 preservation, JS remainder sign (-5 % 3 = -2).
• DIVERGENCE (core accepts, reference abstains): cross-type (in)equality (1 === "1" -> core false, matching the emitted TS/Py legs; reference throws on un-same-typed operands), bitwise (5 | 0, ~5).
• REVERSE-DIVERGENCE (reference accepts, core rejects): truthiness ! over a non-boolean (!"" -> reference true; core's ! is boolean-only).
• BOTH-REJECT: strict + mixed, /0, %0, cross-type comparison, **, non-finite (Infinity / NaN sources refused — KERN numbers are finite-only).

Helper hardened per agon review (agy/codex/kimi consensus): only the engines' genuine reject paths are caught — a probe-side serializer or trace-shape failure now propagates loudly instead of masquerading as a language reject, so the gate can never silently report false parity. A missing/non-assign first event throws rather than returning an ambiguous { ok: true, value: undefined }.

51 tests green. Helper is thin (no per-fixture / per-value-kind branching).

⚔️ Forged by Agon

What

Why

How

Checklist

• tsc -b passes
• pnpm test passes
• pnpm test:kern passes
• pnpm lint passes
• kern review packages/ --recursive checked