De-flake test_flat_panorama across SIMD levels (#5361)

Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/5361

`test_batch_boundaries` flaked (~36%) on the `AVX512_SPR` dispatch variant. The reference `IndexFlat` distance kernel is SIMD-dispatched, so its float reduction order changes with the runtime SIMD level (wide FMA under `AVX512_SPR`), while the `IndexFlatPanorama` kernel is compiled once at baseline and uses the algebraically-equivalent `a*a + b*b - 2ab` decomposition instead of `(a-b)^2`. Both are exact but round differently, so near-tied neighbors reorder at batch boundaries. The earlier scalar de-flake (D109488316 / D109488320, T276491016) only widened the `atol`/`otol` tolerance for the `_NONE` gap, leaving the larger `AVX512_SPR` gap.

Fix: replace the brittle exact-id overlap-rate gate in `assert_search_results_equal` with the tie-aware comparator `faiss.contrib.evaluation.check_ref_knn_with_draws`, which requires distances to match within `rtol` and ids to agree only as sets within each distance tie-group. This is variant-agnostic (covers `NONE` / `AVX2` / `AVX512` / `AVX512_SPR`) and is the standard faiss kNN comparison helper used across the suite.

Reviewed By: mnorris11

Differential Revision: D110092203

fbshipit-source-id: ec6fc8c6114d7dc9bb7b6911478a6d5b0b993b5c

tests/test_flat_panorama.py

@@ -20,6 +20,7 @@ import faiss
 import numpy as np
 from common_faiss_tests import for_all_simd_levels
 from faiss.contrib.datasets import SyntheticDataset
+from faiss.contrib.evaluation import check_ref_knn_with_draws
 
 
 @for_all_simd_levels
@@ -58,20 +59,17 @@ class TestIndexFlatPanorama(unittest.TestCase):
         I_panorama,
         rtol=1e-5,
         atol=1e-4,
-        otol=1e-3,
     ):
-        # Allow small tolerance in overlap rate to account for
-        # floating-point errors in distance computations that can affect
-        # ordering when distances are nearly equal.
-        # Faiss: (a - b) * (a - b) vs. Panorama: a * a + b * b - 2(a * b);
-        # these differ at the float32 ULP level, so atol absorbs the gap.
-        overlap_rate = np.mean(I_regular == I_panorama)
-
-        self.assertGreater(
-            overlap_rate,
-            1 - otol,
-            f"Overlap rate {overlap_rate:.6f} is not > {1 - otol:.3f}. ",
-        )
+        # The reference IndexFlat kernel is SIMD-dispatched, so its float
+        # reduction order changes with the runtime SIMD level (wide FMA under
+        # AVX512_SPR), while the Panorama kernel is compiled once at baseline
+        # and uses the algebraically equivalent a * a + b * b - 2(a * b)
+        # decomposition instead of (a - b) * (a - b). Both are exact but round
+        # differently, so near-tied neighbors can swap order. Compare with the
+        # tie-aware kNN comparator: distances must match within tolerance and
+        # ids only need to agree as sets within each distance tie-group. This is
+        # variant-agnostic (NONE / AVX2 / AVX512 / AVX512_SPR) rather than tuned
+        # to one dispatch level's gap.
         np.testing.assert_allclose(
             D_regular,
             D_panorama,
@@ -79,6 +77,9 @@ class TestIndexFlatPanorama(unittest.TestCase):
             atol=atol,
             err_msg="Distances mismatch",
         )
+        check_ref_knn_with_draws(
+            D_regular, I_regular, D_panorama, I_panorama, rtol=rtol
+        )
 
     def assert_range_results_equal(
         self,