Add combined 50K cross-model persistent-channel summary

analysis/cross_model_persistent_scan_50k_qwen35_2b_data/README.md

@@ -0,0 +1,3 @@
+# 50K Cross-model Persistent Residual Channel Scan
+
+Each model was scanned on the exact same 50,000 prompt strings generated with the original Qwen3.5-2B activation-capture pipeline: UltraChat 200K, FineWeb-Edu sample-10BT, OpenWebMath, and Cosmopedia-v2; the original 20 prompt formats; seed 3407; and the original 48-128 token beta-biased target-length distribution. The detector searches each model's native residual basis for channels with the same qualitative signature as Qwen3.5-2B feature 1201. Feature indices are architecture-specific and are not expected to match 1201 across models.

analysis/cross_model_persistent_scan_50k_qwen35_2b_data/cross_model_summary.csv

@@ -0,0 +1,7 @@
+model,n_layers,hidden_size,best_feature_index,best_persistent_channel_score,best_near_universal_layer_fraction,best_min_worst_group_sign_consistency,best_min_worst_group_strong_coverage,best_mean_topk_coverage,best_median_relative_strength,best_median_magnitude_cv,remote_dir
+Qwen/Qwen3.5-0.8B,24,1024,0,0.9781647439073424,1.0,1.0,1.0,0.9999925,105.74991198312023,0.09454580501397016,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/Qwen__Qwen3.5-0.8B
+Qwen/Qwen3.5-4B,32,2560,0,0.8942107762910956,0.96875,0.8361581920903954,0.5645604395604396,0.99459,264.5687047574181,0.09673759899649652,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/Qwen__Qwen3.5-4B
+Qwen/Qwen3.5-9B,32,4096,3994,0.9832802838292989,1.0,0.9922330097087378,0.9922330097087378,0.9999125,782.0693185305736,0.057514104723956634,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/Qwen__Qwen3.5-9B
+LiquidAI/LFM2.5-2.6B,30,2048,472,0.957948093710141,1.0,0.9985486211901307,0.9985486211901307,0.999996,159.14144180255914,0.23360832535227566,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/LiquidAI__LFM2.5-2.6B
+meta-llama/Llama-3.2-3B-Instruct,28,3072,1731,0.9404529119389515,1.0,1.0,0.9978768577494692,0.96481,233.8051619122549,0.41601373727237057,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/meta-llama__Llama-3.2-3B-Instruct
+google/gemma-4-E2B-it,35,1536,563,0.4960222094100736,0.4,0.5031712473572939,0.0,0.34009542857142855,5.357317412247288,0.22047117760737114,analysis/cross_model_persistent_scan_50k_qwen35_2b_data/google__gemma-4-E2B-it