test(core): execute UWR profile KAT vectors

[Gio2050]

Summary

Executes the afi-config UWR profile KAT vectors against the afi-core engines — the executable half of the UP-11 KAT/golden gate. Additions only: 6 new files, 0 modified, no dependency/lockfile/tsconfig/CI changes.

• validators/__tests__/computeUwrScore.kat.test.ts — 10 score vectors vs computeUwrScore (16 tests)
• src/decay/__tests__/applyTimeDecay.kat.test.ts — 32 decay vectors vs applyTimeDecay (36 tests)
• Byte-identical vendored KAT copies + provenance sidecar READMEs under colocated kats/ fixture dirs

Authorization chain

1. UP-11 clarification (afi-governance/decisions/uwr-profile-pin-v0.1.md): KAT vectors (pure data) live in afi-config; KAT execution against afi-core code is "a separately-authorized afi-core PR (PR-UWR-KAT-EXEC), following the established vendored-golden-with-source-pin pattern used by afi-mint for the emissions goldens."
2. §7 follow-up table row: PR-UWR-KAT-EXEC — "afi-core: execute vendored KAT vectors against computeUwrScore / applyTimeDecay (vendored-golden-with-source-pin pattern)" — marked "No — separate authorization."
3. Owner authorization: the owner's direct instruction to implement this PR (2026-07-12, per the approved planning report) is that separate authorization. Merging this PR is the owner-review acceptance act.

Provenance pins

Vendored from AFI-Protocol/afi-config kats/uwr-profile/v0/ at merge commit fe329164919f0c1c9dc24bb5c279978fb680e983 (PR #17), content commit 7809af4b0308f7db47488947770a0ab9f236268d:

File	git blob SHA-1	sha256
compute-uwr-score.kat.json (10 vectors)	276be9e9348b4e0c2e92e5a9c59d05b50bfeb75d	ff2f63956038b1d5c01109184e2ddddb2606c11405d2420c35b7decb5eb36546
apply-time-decay.kat.json (32 vectors)	1199f705ee65eef70e6c0c58fffcd10206693cf0	1a1240cc9f8cc8ed70b22f4aa00e7cdfe1176f4f1d3c6b1e10e886b0cfe81b78

Executed against afi-core 254185381c3c9be91da303454c1f7a27e8818983 (main, PR #16 merge). The KATs' engine.sourceCommit: 390b440 pin: UniversalWeightingRule.ts is blob-identical between 390b440 and 2541853; GreeksDecayTemplate.ts differs only by the PR #16 kernel delegation, proven bit-exact in applyTimeDecay.mathEquivalence.test.ts — documented in the sidecars, vendored bytes left untouched per change control.

Drift detection: always-on sha256 self-integrity checks against pinned constants (hermetic in CI), plus it.skipIf-guarded full byte-compares against a sibling afi-config checkout on dev machines. Failure messages state the governance implication: divergence requires a new scoped authorization; never silently re-vendor.

Comparison policy (owned by this PR)

Both KAT files deliberately define no float-comparison policy; comparisonPolicyNote assigns it to PR-UWR-KAT-EXEC scope. Policy adopted:

• Decay: Object.is bit-exact, all 32 vectors — per the file's exactnessRule (dyadic base scores × integer half-life multiples ⇒ every expected value exact in binary64).
• Score: Object.is bit-exact for 9/10 vectors.
• uwr-score-anchor-0650 — dual-pinned characterized exception. The engine's multiply-accumulate order over axes (0.6, 0.7, 0.6, 0.7) with 0.25 weights yields 0.6499999999999999, exactly 1 ULP below the decimal anchor 0.65. This is deterministic, correctly-rounded IEEE-754 binary64 arithmetic (no transcendentals) — identical on every conforming engine, not flake. The test asserts both |actual − 0.65| ≤ 1 ULP and Object.is(actual, 0.6499999999999999), pinning the engine's true output verbatim so any future engine change fails loudly (drift-detection strength equal to bit-exact). 0.65 is not decision-pinned (only the 0.1875 golden anchor and the 0.25 weights are decision-named). Rejected alternatives: editing the vendored expected value (violates afi-config change control) and blanket tolerances (weaken the 41 vectors that can be exact). Optional future item for the owner (separately authorized, not this PR): a one-vector afi-config erratum swapping the 0650 axes to e.g. (0.5, 0.6, 0.7, 0.8) would collapse the exception to uniform bit-exactness.
• uwr-score-d2m2-golden-anchor (0.1875) — the decision-named D2 M2 golden replay anchor — reproduces bit-exactly and is called out in a dedicated test (UP-5/UP-11 byte-stability criterion).

What this does NOT unlock

Execution ≠ wiring. This PR runs KAT vectors in tests and nothing more:

• No reward, mint, or qualification wiring — IValidatorScorer/ValidatorDaemon and all reward/mint paths untouched; no eligibility assertions anywhere (the 0.5 = minDecayScoreThreshold coincidence in the decay data stays commentary).
• No PR-UWR-STAMP — no reactor stamping; separately authorized.
• No runtime registry consumption — the KAT JSON is imported only from *.test.ts; the executed config is built from the KAT's own weights block; defaultUwrConfig unchanged.
• No UP-8 resolution — decay-engine canonicality remains OPEN; these tests pin the GreeksDecayTemplate surface only.
• Status: testnet-provisional (x-afiStatus: draft-non-implementation). Each follow-up requires its own scoped authorization.

Validation

• git check-ignore on new paths: nothing ignored (fixture dirs named kats/, not vendor*/)
• cmp + sha256sum: vendored copies byte-identical to afi-config @ fe32916, hashes match pins
• npm run test:run: 196 passed (existing 144 unchanged + 16 score-KAT + 36 decay-KAT), including live sibling byte-compares
• npm run build, npm run typecheck, scripts/esm-check.sh: clean (tests are tsc-excluded; no config edits)
• git status: exactly 6 additions, 0 modifications; git diff --stat empty
• Adversarial verification pass: 0 blockers across 4 independent lenses — (1) provenance/byte-equivalence re-derived from the pinned afi-config commit (git-show bytes, blob SHAs, content-commit ancestry, every hash constant in tests/sidecars cross-checked; UniversalWeightingRule.ts confirmed blob-identical 390b440↔2541853, GreeksDecayTemplate.ts delta confirmed to be exactly the PR refactor(decay): delegate applyTimeDecay kernel to afi-math (PR-7, bit-exact) #16 delegation); (2) additions-only scope audit (status purity, no config/CI/lockfile edits, no vendor*/, KAT JSON referenced only from tests/sidecars, all six sibling repos byte-verified clean); (3) language guards (0 banned-phrase hits; required testnet-provisional / execution≠wiring / separately-authorized / never-silently-re-vendor language present; UP-8 stated OPEN); (4) test-strength review + mutation testing in an isolated copy — 4/4 mutations killed (vendored-byte flip, 0650 pin change, decay expected-value edit, weights swap), baseline green both with and without a sibling checkout (skip path verified)

🤖 Generated with Claude Code