Use H3 cache constants in tests

tests/test_minimax_h3.py

@@ -88,9 +88,9 @@ def runtime_for(config, inner, *, packed_len=6, beta=0.7, selected=None):
         "sage_lse_unavailable": None,
         "debug_state": None,
         "debug_topology": (packed_len,),
-        "attention_bias_cache": {},
-        "text_index_cache": {},
-        "scaled_rope_cache": {},
+        h3._ATTENTION_BIAS_CACHE_KEY: {},
+        h3._TEXT_INDEX_CACHE_KEY: {},
+        h3._ROPE_CACHE_KEY: {},
     }
 
 
@@ -644,7 +644,7 @@ def test_scaled_rope_is_computed_once_per_forward_and_invalidated_by_tensor(monk
     for replacement in replacements:
         replacement(block_args(runtime, rope), {"original_block": original})
     assert count == 1
-    cached_source, cached_scaled = next(iter(runtime["scaled_rope_cache"].values()))
+    cached_source, cached_scaled = next(iter(runtime[h3._ROPE_CACHE_KEY].values()))
     assert cached_source is rope
     assert torch.is_tensor(cached_scaled)
     replacement = replacements[0]
@@ -754,7 +754,7 @@ def test_runtime_masks_are_isolated_between_configured_clones():
     mask_b = h3._compact_text_bias(runtime_b, q)
     assert mask_a is not mask_b
     assert not torch.equal(mask_a, mask_b)
-    assert runtime_a["attention_bias_cache"] is not runtime_b["attention_bias_cache"]
+    assert runtime_a[h3._ATTENTION_BIAS_CACHE_KEY] is not runtime_b[h3._ATTENTION_BIAS_CACHE_KEY]
 
 
 def test_interrupted_attention_call_does_not_poison_next_runtime():