test(python): close the three parity-review findings (native anchor in CI, NaN, cosine)

A cross-vendor review of the parity rework (2febbb81) found three issues; all
verified and fixed here.

1. HIGH — the native≈golden half never ran in CI, so a binding marshalling bug
   could pass. The golden vectors were regenerated through the Python binding,
   and only pytest (binding vs that golden) runs in CI — the native reference
   tests in `python/tests/aether_parity.rs` link against the PyO3 crate and no
   workflow runs them. A stable PyO3 conversion defect present at regeneration
   would therefore be baked into the golden and go undetected.

   Fix: a real native parity test IN the `wifi-densepose-aether` crate
   (`tests/golden_parity.rs`), which is std-only and a member of the v2
   workspace, so it runs under the existing `cargo test --workspace`. It
   recomputes the embedding with no PyO3/marshalling and asserts it matches the
   SAME committed golden within tolerance. Now native≈golden AND binding≈golden
   both run in CI ⇒ binding≈native, and a binding-specific artifact in the
   golden surfaces here as a native mismatch. serde_json added as a
   dev-dependency only (test-only; never linked into the lib or the wheel, so
   the crate stays runtime-dependency-free).

   Independently proven at this commit: native output equals the committed
   golden BIT-FOR-BIT (128/128 exact bits, Δ=0) for both base and loaded — the
   golden is native-faithful today; this test keeps it that way.

2. MEDIUM — the Python parity helper passed non-finite output. `abs(nan - b) >
   tol` is False, so an all-NaN embedding slipped through. Now every element
   must be `math.isfinite` first. Proven: an all-NaN vector is rejected
   ("element 0 is not finite (nan)"); the Rust helpers already caught it via
   `<=`.

3. LOW — a coherent shift inside the per-element tolerance could move the whole
   vector undetected. Added a whole-vector cosine-similarity bound (≥ 1 - 1e-6)
   alongside the per-element check.

Verified on aarch64/macOS: `test_aether` 13, full `python/tests/` suite 227
passed against a `--features sota` build; the new native test passes
(2/2) run standalone (the in-worktree `cargo test -p` only failed on an
un-checked-out submodule; ci.yml checks out submodules recursively).

Co-Authored-By: Ruflo & AQE

v2/crates/wifi-densepose-aether/Cargo.toml

@@ -16,6 +16,11 @@ categories.workspace = true
 # ruvector server tree. Keep it std-only.
 [dependencies]
 
+# Test-only: parses the committed golden fixtures shared with the Python parity
+# tests. Never linked into the library or the wheel.
+[dev-dependencies]
+serde_json = "1"
+
 [lib]
 name = "wifi_densepose_aether"
 path = "src/lib.rs"

v2/crates/wifi-densepose-aether/tests/golden_parity.rs

@@ -0,0 +1,115 @@
+//! ADR-185 §4.1 — NATIVE half of the AETHER parity gate, and the half that runs
+//! in CI.
+//!
+//! The committed golden vectors live under `python/tests/golden/` and are
+//! shared with `python/tests/test_aether.py` (the binding half). That pytest
+//! runs in python-ci; the native reference tests in `python/tests/*.rs` link
+//! against the PyO3 crate and are NOT run by any workflow. This test closes that
+//! gap: it recomputes the embedding through THIS std-only crate — no PyO3, no
+//! marshalling — and asserts it matches the same golden within tolerance.
+//!
+//! Together with the pytest half: native≈golden AND binding≈golden ⇒
+//! binding≈native, portably. And because this side has no marshalling, a
+//! binding-specific defect baked into the golden would surface here as a native
+//! mismatch — which is the failure mode a binding-derived golden + pytest-only
+//! CI would otherwise hide.
+//!
+//! Runs under the repo's `cargo test --workspace` (this crate is a member).
+
+use std::path::PathBuf;
+
+use wifi_densepose_aether::embedding::{EmbeddingConfig, EmbeddingExtractor};
+use wifi_densepose_aether::graph_transformer::TransformerConfig;
+
+// Same tolerance as the Python and .rs parity tests. f32 + transcendentals are
+// not bit-reproducible across arch, so the golden (generated on one machine) is
+// compared within a bound, not by hash.
+const PARITY_ATOL: f32 = 1e-4;
+const PARITY_RTOL: f32 = 1e-4;
+
+fn golden_dir() -> PathBuf {
+    // This crate lives at v2/crates/wifi-densepose-aether; the shared golden
+    // fixtures are the single source of truth under python/tests/golden.
+    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+        .join("../../../python/tests/golden")
+}
+
+fn read_vec(name: &str) -> Vec<f32> {
+    let raw = std::fs::read_to_string(golden_dir().join(name))
+        .unwrap_or_else(|e| panic!("read {name}: {e}"));
+    serde_json::from_str(&raw).unwrap_or_else(|e| panic!("parse {name}: {e}"))
+}
+
+fn load_input() -> Vec<Vec<f32>> {
+    let raw = std::fs::read_to_string(golden_dir().join("aether_input.json"))
+        .expect("read aether_input.json");
+    let rows: Vec<Vec<f64>> = serde_json::from_str(&raw).expect("parse aether_input.json");
+    rows.into_iter()
+        .map(|r| r.into_iter().map(|x| x as f32).collect())
+        .collect()
+}
+
+fn extractor() -> EmbeddingExtractor {
+    let e = EmbeddingConfig { d_model: 64, d_proj: 128, temperature: 0.07, normalize: true };
+    let t = TransformerConfig {
+        n_subcarriers: 56,
+        n_keypoints: 17,
+        d_model: 64,
+        n_heads: 4,
+        n_gnn_layers: 2,
+    };
+    EmbeddingExtractor::new(t, e)
+}
+
+fn formula_weights(n: usize) -> Vec<f32> {
+    (0..n)
+        .map(|i| ((i as u64 * 1103515245 + 12345) % 65536) as f32 / 65536.0 - 0.5)
+        .collect()
+}
+
+fn assert_matches_golden(embedding: &[f32], name: &str) {
+    let golden = read_vec(name);
+    assert_eq!(embedding.len(), golden.len(), "{name}: length mismatch");
+    for (i, (&a, &b)) in embedding.iter().zip(&golden).enumerate() {
+        assert!(a.is_finite(), "{name}: element {i} is not finite ({a})");
+        let tol = PARITY_ATOL + PARITY_RTOL * b.abs();
+        assert!(
+            (a - b).abs() <= tol,
+            "{name}: element {i} diverged beyond tolerance \
+             (got {a}, golden {b}, |Δ|={}) — real regression, not arch drift",
+            (a - b).abs()
+        );
+    }
+}
+
+#[test]
+fn native_base_embedding_matches_committed_golden() {
+    let emb = extractor().extract(&load_input());
+    assert_matches_golden(&emb, "aether_embedding.json");
+}
+
+#[test]
+fn native_loaded_embedding_matches_committed_golden() {
+    let input = load_input();
+    let mut ext = extractor();
+    let baseline = ext.extract(&input);
+
+    let weights = formula_weights(ext.param_count());
+    let mut buf = Vec::new();
+    buf.extend_from_slice(b"AETHERW1");
+    buf.extend_from_slice(&(weights.len() as u32).to_le_bytes());
+    for w in &weights {
+        buf.extend_from_slice(&w.to_le_bytes());
+    }
+    let wpath = std::env::temp_dir().join(format!("aether_golden_parity_{}.bin", std::process::id()));
+    std::fs::write(&wpath, &buf).expect("write weights");
+    ext.load_weights(&wpath).expect("load_weights");
+    let _ = std::fs::remove_file(&wpath);
+
+    let loaded = ext.extract(&input);
+    assert!(
+        baseline.iter().zip(&loaded).any(|(a, b)| (a - b).abs() > 1e-6),
+        "load_weights had no effect vs the random-init baseline"
+    );
+    assert_matches_golden(&loaded, "aether_loaded_embedding.json");
+}