feat(cwb): add focused merge presets

docs/consensus_mergers.md

@@ -177,13 +177,17 @@ Dense nodes offer `balanced_mean`, `robust_medn`, `selective_mean`,
 
 Embedding nodes offer `balanced_idx_mean`, `balanced_sim_mean`,
 `robust_idx_medn`, `robust_sim_medn`, `varied_sim_mean_rn_softcb`, and
-`diverse_sim_medn_rn_dsc_softcb`.
+`diverse_sim_medn_rn_dsc_softcb`, `focused_strong_sim_medn`,
+`focused_balance_sim_medn`, `focused_soft_sim_medn`, and
+`focused_weak_sim_medn`.
 
 LoRA nodes offer `broad_sim_medn_rn_softcb`,
 `moderate_sim_medn_rn_softcb`, `conservative_sim_medn_rn_softcb`,
 `direct_idx_medn_rn_softcb`, `broad_sim_mean_rn_softcb`,
 `neutral_sim_medn_rn`, `focused_sim_medn_rn_dsc_softcb`, and
-`strongfocus_sim_medn_rn_dsc_softcb`.
+`strongfocus_sim_medn_rn_dsc_softcb`, `focused_strong_sim_medn`,
+`focused_balance_sim_medn`, `focused_soft_sim_medn`, and
+`focused_weak_sim_medn`.
 
 The LoRA default is `broad_sim_medn_rn_softcb`: median consensus,
 similarity alignment, zero similarity threshold, alpha 2, beta 4, norm
@@ -204,6 +208,13 @@ Every LoRA preset fixes all 12 controls explicitly:
 | `neutral_sim_medn_rn` | median | similarity | 0 | 0 | 2 | 0 | on | 1 | off | off | 0.05 | off | Similarity weighting without diversity modulation |
 | `focused_sim_medn_rn_dsc_softcb` | median | similarity | 0 | 0 | 2 | 4 | on | 1 | on | on | 0.05 | off | Moderate DSC concentration |
 | `strongfocus_sim_medn_rn_dsc_softcb` | median | similarity | 0 | 0 | 2 | 7 | on | 1 | on | on | 0.05 | off | Strong DSC concentration |
+| `focused_strong_sim_medn` | median | similarity | 0.85 | 0.60 | 1.25 | 0 | off | 1 | off | off | 0.20 | off | Strongly focused matching |
+| `focused_balance_sim_medn` | median | similarity | 0.75 | 0.55 | 1.25 | 0 | off | 1 | off | off | 0.20 | off | Balanced focused matching |
+| `focused_soft_sim_medn` | median | similarity | 0.55 | 0.50 | 1.25 | 0 | off | 1 | off | off | 0.20 | off | Soft focused matching |
+| `focused_weak_sim_medn` | median | similarity | 0.25 | 0.35 | 1.25 | 0 | off | 1 | off | off | 0.20 | off | Weak focused matching |
+
+The four focused median similarity presets have identical settings in the
+embedding and LoRA registries.
 
 The LoRA Multi-Merge counterfactual weight sweep reuses each consensus vector
 already computed by the active merge. It reports mean and median consensus

nodes/consensus_merger.py

@@ -157,6 +157,38 @@ EMBEDDING_CWB_PRESETS = {
         dynamic_similarity_contrast=True, soft_comfort_bandpass=True,
         position_weight=0.1, preserve_common_prefix=False,
     ),
+    "focused_strong_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.85, similarity_threshold=0.60,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_balance_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.75, similarity_threshold=0.55,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_soft_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.55, similarity_threshold=0.50,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_weak_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.25, similarity_threshold=0.35,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
 }
 
 LORA_CWB_PRESETS = {
@@ -224,6 +256,38 @@ LORA_CWB_PRESETS = {
         soft_comfort_bandpass=True, position_weight=0.05,
         preserve_common_prefix=False,
     ),
+    "focused_strong_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.85, similarity_threshold=0.60,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_balance_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.75, similarity_threshold=0.55,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_soft_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.55, similarity_threshold=0.50,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
+    "focused_weak_sim_medn": _preset(
+        consensus_type="median", alignment_method="similarity",
+        alignment_threshold=0.25, similarity_threshold=0.35,
+        power_alpha=1.25, diversity_beta=0.0, rescale_norm=False,
+        global_scale=1.0, dynamic_similarity_contrast=False,
+        soft_comfort_bandpass=False, position_weight=0.20,
+        preserve_common_prefix=False,
+    ),
 }
 
 

tests/test_consensus_merger.py

@@ -135,6 +135,39 @@ def test_lora_alignment_presets_use_sweep_verified_thresholds(cwb):
     ] == pytest.approx(0.0025)
 
 
+@pytest.mark.parametrize(
+    ("registry_name", "preset", "alignment_threshold", "similarity_threshold"),
+    [
+        (registry_name, preset, alignment_threshold, similarity_threshold)
+        for registry_name in ("EMBEDDING_CWB_PRESETS", "LORA_CWB_PRESETS")
+        for preset, alignment_threshold, similarity_threshold in (
+            ("focused_strong_sim_medn", 0.85, 0.60),
+            ("focused_balance_sim_medn", 0.75, 0.55),
+            ("focused_soft_sim_medn", 0.55, 0.50),
+            ("focused_weak_sim_medn", 0.25, 0.35),
+        )
+    ],
+)
+def test_focused_median_similarity_presets(
+    cwb, registry_name, preset, alignment_threshold, similarity_threshold
+):
+    settings = getattr(cwb, registry_name)[preset]
+    assert settings == {
+        "consensus_type": "median",
+        "alignment_method": "similarity",
+        "alignment_threshold": alignment_threshold,
+        "similarity_threshold": similarity_threshold,
+        "power_alpha": 1.25,
+        "diversity_beta": 0.0,
+        "rescale_norm": False,
+        "global_scale": 1.0,
+        "dynamic_similarity_contrast": False,
+        "soft_comfort_bandpass": False,
+        "position_weight": 0.20,
+        "preserve_common_prefix": False,
+    }
+
+
 def test_cwb_math_index_prefix_and_scale(cwb):
     settings = _settings(
         cwb,
@@ -647,7 +680,7 @@ def test_use_case_preset_names_match_complete_settings(cwb):
         cwb.EMBEDDING_CWB_PRESETS,
         cwb.LORA_CWB_PRESETS,
     ):
-        assert 6 <= len(registry) <= 8
+        assert len(registry) >= 6
         for name, values in registry.items():
             assert set(values) == fields
             assert ("_rn" in name) is values["rescale_norm"]