fix(router): imperativeVerbs was weighted zero — restore the dimension key

CHANGELOG.md

@@ -38,6 +38,18 @@ All notable changes to blockrun-llm will be documented in this file.
   never retried, so a second model cannot sign a second transfer for one call.
 
 ### Fixed
+- **One scoring dimension was silently weighted zero.** The config transpile that
+  produced `router_core/config.py` snake_cased key names, and `imperativeVerbs`
+  is both a keyword-list field *and* a dimension name — so its 0.03 weight
+  landed under `imperative_verbs` while the classifier emitted `imperativeVerbs`,
+  and `weights.get(name, 0)` scored it zero. Build/deploy-shaped requests were
+  under-classified: 3 of 8 sampled imperative prompts ("Create and deploy the
+  service", "Set up the config and deploy it", "Develop a CLI that generates
+  reports") landed in SIMPLE where they should have been ambiguous and defaulted
+  up to MEDIUM. Now guarded by a test asserting the weight keys and the emitted
+  dimension names are the same set, plus the verbatim upstream weight table.
+  Cross-SDK parity re-verified at 24/24 after the fix.
+
 - **A 429 now walks the fallback chain instead of failing the call.** Both
   clients treated only 5xx as retriable, so a saturated upstream ended the
   request even with capable models left in the chain. Found live: a rate-limited

blockrun_llm/router_core/config.py

@@ -1048,7 +1048,7 @@
             "simpleIndicators": 0.02,  # Reduced from 0.12 to make room for agenticTask
             "multiStepPatterns": 0.12,
             "questionComplexity": 0.05,
-            "imperative_verbs": 0.03,
+            "imperativeVerbs": 0.03,
             "constraintCount": 0.04,
             "outputFormat": 0.03,
             "referenceComplexity": 0.02,

tests/unit/test_router_core.py

@@ -1235,3 +1235,46 @@ def test_does_not_treat_host_tool_descriptions_as_a_per_turn_requirement(self):
         )
 
         assert not infer_tool_requirement("What is 17 times 9?", system_prompt)
+
+
+class TestDimensionWeightKeys:
+    """Every scored dimension must find its weight.
+
+    The config transpile that produced ``config.py`` snake_cased key names, and
+    ``imperativeVerbs`` is both a keyword-list field *and* a dimension name — so
+    the weight landed under ``imperative_verbs`` while the classifier emitted
+    ``imperativeVerbs``. ``weights.get(name, 0)`` then silently scored that
+    dimension at zero, diverging from the TypeScript SDK on any prompt whose
+    imperative verbs would have crossed a tier boundary.
+    """
+
+    def test_every_emitted_dimension_has_a_weight(self):
+        from blockrun_llm.router_core.rules import classify_by_rules
+
+        scoring = DEFAULT_ROUTING_CONFIG["scoring"]
+        result = classify_by_rules("Build and deploy the service", None, 10, scoring)
+        emitted = {dimension["name"] for dimension in result["dimensions"]}
+        weighted = set(scoring["dimension_weights"])
+
+        assert emitted - weighted == set(), "scored dimensions with no weight"
+        assert weighted - emitted == set(), "weights that match no scored dimension"
+
+    def test_the_weights_match_the_upstream_values(self):
+        # Ported verbatim from router-core config.ts at 18bf4ab.
+        assert DEFAULT_ROUTING_CONFIG["scoring"]["dimension_weights"] == {
+            "tokenCount": 0.08,
+            "codePresence": 0.15,
+            "reasoningMarkers": 0.18,
+            "technicalTerms": 0.1,
+            "creativeMarkers": 0.05,
+            "simpleIndicators": 0.02,
+            "multiStepPatterns": 0.12,
+            "questionComplexity": 0.05,
+            "imperativeVerbs": 0.03,
+            "constraintCount": 0.04,
+            "outputFormat": 0.03,
+            "referenceComplexity": 0.02,
+            "negationComplexity": 0.01,
+            "domainSpecificity": 0.02,
+            "agenticTask": 0.04,
+        }