fix: route production hot paths through algebra.backend (Commit 2)

Three surgical import changes. No behavior change. No new semantics.
Backend decides Rust vs pure-Python transparently.

- field/propagate.py: versor_apply <- algebra.backend
- vocab/manifold.py: cga_inner    <- algebra.backend
- vault/store.py:    recall loop replaced with vault_recall() from
  algebra.backend; public result shape preserved ({versor, score,
  metadata, index}). null_project stays on algebra.cga (not the
  recall hot path). store() and reproject() unchanged.

Rust path for vault_recall uses Rayon parallel scan and releases
the GIL. Python fallback is sequential and behaviorally identical.
No batching introduced; that is Commit 3+.

field/propagate.py

@@ -5,9 +5,12 @@ Each step: F <- versor_apply(V, F)
 
 V is the rotor for the current node's outgoing edge in the vocab manifold.
 No correction. No normalization. No conditional branching. The loop is tight.
+
+Hot path routes through algebra.backend, which dispatches to the Rust
+extension (core_rs) when available and falls back to pure Python silently.
 """
 
-from algebra.versor import versor_apply
+from algebra.backend import versor_apply
 from field.state import FieldState
 
 

vault/store.py

@@ -4,10 +4,19 @@ VaultStore — exact memory via CGA inner product scan.
 No HNSW. No approximate nearest neighbor. No index rebuild.
 Recall is exact: argmax_i { cga_inner(query, X_i) } over stored versors.
 Periodic null_project() prevents floating-point null-cone drift in long sessions.
+
+Hot path: recall() routes through algebra.backend.vault_recall(), which
+dispatches to a Rayon parallel scan (releases GIL) when core_rs is available
+and falls back to a sequential Python scan silently. Public result shape
+is unchanged: list of {versor, score, metadata, index}.
+
+null_project() remains on algebra.cga — it is not the recall hot path
+and does not benefit from the same batching pattern.
 """
 
 import numpy as np
-from algebra.cga import cga_inner, null_project
+from algebra.backend import vault_recall
+from algebra.cga import null_project
 
 
 class VaultStore:
@@ -30,25 +39,31 @@ class VaultStore:
         """
         Return top_k closest stored versors by CGA inner product.
         Each result: {versor, score, metadata, index}
+
+        Routes through algebra.backend.vault_recall():
+          Rust path  — Rayon parallel scan, GIL released.
+          Python path — sequential, behaviorally identical.
         """
         if not self._versors:
             return []
-        scores = [cga_inner(query, v) for v in self._versors]
-        top_indices = sorted(range(len(scores)), key=lambda i: -scores[i])[:top_k]
+
+        ranked = vault_recall(self._versors, query, top_k)
+
         return [
             {
                 "versor": self._versors[i],
-                "score": scores[i],
+                "score": float(score),
                 "metadata": self._metadata[i],
                 "index": i,
             }
-            for i in top_indices
+            for i, score in ranked
         ]
 
     def reproject(self) -> None:
         """
         Re-project all stored versors onto the null cone.
         Corrects floating-point drift. Run between turns or asynchronously.
+        null_project stays on algebra.cga — not the recall hot path.
         """
         self._versors = [null_project(v) for v in self._versors]
 

vocab/manifold.py

@@ -21,10 +21,13 @@ Indexed access:
   get_word_at(idx)    — returns the word string by integer index.
   These are the primitives generation uses; VocabManifold does not build
   operators. Algebra builds operators. Vocab stores points.
+
+Hot path: nearest() routes cga_inner through algebra.backend, which
+dispatches to the Rust extension when available.
 """
 
 import numpy as np
-from algebra.cga import cga_inner
+from algebra.backend import cga_inner
 from algebra.cl41 import geometric_product, reverse
 
 
@@ -86,6 +89,8 @@ class VocabManifold:
         Find the word whose versor is closest to F by CGA inner product.
         Returns (word, index). O(|vocab|), exact, no approximation.
         cga_inner(X, Y) = -d^2 / 2 for null vectors: maximizing = minimizing distance.
+
+        Hot path: cga_inner routes through algebra.backend.
         """
         best_score = -np.inf
         best_idx = 0