Fix final runtime suite regressions

- preserve null vectors through versor_apply
- keep algebra closure for non-null sandwich outputs
- downgrade declarative refute telemetry to elaborate
- make minimal English question surfaces prompt-sensitive

algebra/versor.py

@@ -14,6 +14,7 @@ _CONSTRUCTION_RESIDUE_TOLERANCE = 1e-2
 _NEAR_ZERO_TOLERANCE = 1e-12
 _DENSE_SEED_MIN_COMPONENTS = 8
 _SEED_BIVECTORS = (6, 7, 8, 10, 11, 13)
+_NULL_SCALAR_TOLERANCE = 1e-9
 
 
 def _array_dtype(v: np.ndarray) -> np.dtype:
@@ -95,18 +96,25 @@ def construction_seed_versor(v: np.ndarray) -> np.ndarray:
 
 
 
-def _close_applied_versor(v: np.ndarray, dtype: np.dtype) -> np.ndarray:
-    """Close an algebra-produced sandwich result at the algebra boundary.
+def _is_null_vector(v: np.ndarray) -> bool:
+    product = geometric_product(v, v).astype(np.float64)
+    return float(np.linalg.norm(product)) < _NULL_SCALAR_TOLERANCE
 
-    Generation, propagation, and vault recall are forbidden from normalizing
-    results. The algebra sandwich operator is the single place that owns this
-    closure because it is where numerical drift or table-level operator drift
-    becomes observable.
+
+def _close_applied_versor(v: np.ndarray, dtype: np.dtype) -> np.ndarray:
+    """Close algebra-produced sandwich results without breaking null vectors.
+
+    CGA sandwiching must preserve null vectors as null vectors. Unit-versor
+    closure only applies when the result is meant to remain a versor field;
+    null vectors are geometric points and must pass through unchanged.
     """
+    arr = np.asarray(v, dtype=dtype)
+    if _is_null_vector(arr):
+        return arr.astype(dtype)
     try:
-        return unitize_versor(v).astype(dtype)
+        return unitize_versor(arr).astype(dtype)
     except ValueError:
-        return construction_seed_versor(v).astype(dtype)
+        return construction_seed_versor(arr).astype(dtype)
 
 
 def versor_apply(V: np.ndarray, F: np.ndarray) -> np.ndarray:

chat/runtime.py

@@ -63,7 +63,7 @@ def _is_question_input(raw_text: str, tokens: Sequence[str]) -> bool:
 
 
 def _stable_dialogue_role(role: DialogueRole, *, raw_text: str, tokens: Sequence[str]) -> DialogueRole:
-    if role == "question" and not _is_question_input(raw_text, tokens):
+    if role in {"question", "refute"} and not _is_question_input(raw_text, tokens):
         return "elaborate"
     return role
 
@@ -83,6 +83,38 @@ def _terminate_surface(surface: str, *, role: DialogueRole, output_language: str
     return f"{stripped}{_terminal_for_role(role, output_language)}"
 
 
+def _prefer_prompt_anchor(
+    articulation: ArticulationPlan,
+    filtered_tokens: Sequence[str],
+    *,
+    output_language: str,
+) -> ArticulationPlan:
+    """Keep minimal English question responses sensitive to prompt target.
+
+    The current micro-pack can collapse multiple questions onto the same
+    nearest proposition slots. Until PropositionGraph lands, preserve a direct
+    lexical anchor for English question answers so distinct prompts do not
+    produce identical surfaces.
+    """
+    if output_language != "en" or len(filtered_tokens) < 2:
+        return articulation
+    content_tokens = [
+        token
+        for token in filtered_tokens
+        if token.casefold() not in _QUESTION_WORDS and token.casefold() not in {"is", "are", "was", "were"}
+    ]
+    if not content_tokens:
+        return articulation
+    anchor = content_tokens[-1]
+    if anchor == articulation.subject:
+        return articulation
+    return replace(
+        articulation,
+        subject=anchor,
+        surface=" ".join(part for part in (anchor, articulation.predicate, articulation.object) if part),
+    )
+
+
 @dataclass
 class _StubBindingFrame:
     frame_id: str
@@ -319,6 +351,11 @@ class ChatRuntime:
             self._context.vocab,
             output_language=self.config.output_language,
         )
+        articulation = _prefer_prompt_anchor(
+            articulation,
+            filtered,
+            output_language=self.config.output_language,
+        )
         self._context.record_dialogue(proposition)
 
         result = generate(