Fix Hamming attention scale factor: sqrt(pi/2)/m → (pi/2)/m

ClawTeam research findings:
1. Sign-sign agreement estimator uses (pi/2)/m, not sqrt(pi/2)/m
   sqrt(pi/2) is for QJL random-projection-then-sign (different estimator)
2. RHT seed-invariance: changing seed doesn't change sign agreement
   (random diagonal cancels in sign comparison)
3. Multi-hash with Permute+RHT K=4: cosine 0.854 for dim=64

Current architecture: int_attn disabled (FP32 attention + 1-bit storage).
Scale fix is for future Hamming attention restoration.

33/33 tests pass. PPL unchanged (FP32 attention path).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

src/core/tq_turbo_kv.c

@@ -768,8 +768,11 @@ void tq_turbo_kv_1b_attention_ref(const float* query, const void* kv_cache,
     int sketch_dim = dim;
     if (sketch_dim < TQ_BK) sketch_dim = TQ_BK;
 
-    /* Scale factor uses sketch_dim (total sign bits), not dim */
-    float scale_factor = sqrtf(TQ_PI_2) / (float)sketch_dim;
+    /* Scale factor for sign-sign agreement estimator: (pi/2) / m.
+     * Note: sqrt(pi/2)/m is for random-projection-then-sign (QJL).
+     * sign-sign (Hamming) uses pi/2 per the arcsin law.
+     * Currently int_attn is disabled, but fix for future use. */
+    float scale_factor = TQ_PI_2 / (float)sketch_dim;
 
     /* Step 1: RHT(query) with expansion matching quantize */
     float q_rot[TQ_BK];

tests/test_neon_scalar.cpp

@@ -295,8 +295,8 @@ TEST(NeonScalarConsistency, HammingAttentionReference) {
     tq_turbo_kv_1b_attention_ref(query.data(), kv_blocks.data(),
                                    actual_scores.data(), seq_len, dim);
 
-    /* Manual reference computation */
-    float scale_factor = sqrtf((float)M_PI / 2.0f) / (float)dim;
+    /* Manual reference computation — sign-sign estimator uses (pi/2)/m */
+    float scale_factor = ((float)M_PI / 2.0f) / (float)dim;
 
     /* RHT(query) */
     std::vector<float> q_rot(dim);