Stabilize holonomy accumulation

algebra/holonomy.py

@@ -2,65 +2,77 @@
 Holonomy prompt encoding.
 
 A prompt w1, w2, ..., wn is encoded as the geometric holonomy of its
-forward+reverse versor walk. The walk closes, producing a versor that
+forward+reverse versor walk. The walk closes, producing a bounded algebraic
-is bounded by construction and invariant to global phase.
+summary of the prompt path.
 
-The holonomy IS a versor — it drops directly into versor_apply with
+The input word objects must already be valid construction-time versors.
-no bridging code. The fuel and the engine are the same substance.
+Holonomy may unitize intermediate construction products to prevent float32
+scale blow-up, but never repairs propagation state.
 """
 
+from __future__ import annotations
+
 import numpy as np
+
 from .cl41 import geometric_product, reverse as cl_reverse
-from .versor import normalize_to_versor
+from .versor import unitize_versor
 from .cga import cga_inner
 
 
+def _renorm_if_needed(H: np.ndarray, step: int, renorm_every: int) -> np.ndarray:
+    """Bound accumulator scale to prevent float32 overflow on long prompts."""
+    if renorm_every <= 0 or step % renorm_every != 0:
+        return H
+    norm = float(np.linalg.norm(H))
+    if not np.isfinite(norm) or norm < 1e-12:
+        raise ValueError("holonomy accumulator became null/non-finite during encoding.")
+    return (H / norm).astype(np.float32)
+
+
 def holonomy_encode(
     word_versors: list,
     alpha: float = 0.5,
-    weights: list = None,
+    weights: list | None = None,
+    renorm_every: int = 8,
 ) -> np.ndarray:
     """
     Compute the holonomy of the word versor sequence.
 
     Forward walk:  F = w1 * w2 * ... * wn  (weighted by word frequency inverse)
     Reverse walk:  R = (1-alpha) * reverse(wn) * ... * reverse(w1)
-    Holonomy:      H = geometric_product(F, R)
+    Holonomy:      H = F * R
 
-    H is a versor. For alpha=0.5, the holonomy captures the geometric
+    Construction-time unitization is used at the boundary and at the final
-    curvature of the prompt path. Prompts with different semantic content
+    product. A bounded Euclidean renormalization is also applied every
-    produce geometrically distinct holonomies even at the same length.
+    `renorm_every` steps to prevent long prompt overflow in float32.
-
-    weights: optional list of float scalars (e.g. inverse token frequency).
-             Rare content words rotate more than common function words.
-             If None, uniform weights are used.
     """
     if not word_versors:
         raise ValueError("Cannot encode empty prompt.")
+    if not 0.0 <= alpha <= 1.0:
+        raise ValueError("alpha must be in [0, 1].")
 
     n = len(word_versors)
     if weights is None:
         weights = [1.0] * n
-    assert len(weights) == n
+    if len(weights) != n:
+        raise ValueError("weights length must match word_versors length.")
 
-    # Forward accumulation
+    # Forward accumulation.
-    F = word_versors[0].copy() * weights[0]
+    F = unitize_versor(np.asarray(word_versors[0], dtype=np.float32) * weights[0])
-    F = normalize_to_versor(F)
     for k in range(1, n):
-        w = word_versors[k] * weights[k]
+        w = unitize_versor(np.asarray(word_versors[k], dtype=np.float32) * weights[k])
-        w = normalize_to_versor(w)
         F = geometric_product(F, w)
+        F = _renorm_if_needed(F, k, renorm_every)
 
-    # Reverse accumulation with alpha damping
+    # Reverse accumulation with alpha damping.
-    R = cl_reverse(word_versors[-1]) * (1.0 - alpha)
+    R = unitize_versor(cl_reverse(word_versors[-1]) * (1.0 - alpha))
-    R = normalize_to_versor(R)
     for k in range(n - 2, -1, -1):
-        r = cl_reverse(word_versors[k])
+        r = unitize_versor(cl_reverse(word_versors[k]))
-        r = normalize_to_versor(r)
         R = geometric_product(r, R)
+        R = _renorm_if_needed(R, n - 1 - k, renorm_every)
 
     H = geometric_product(F, R)
-    return normalize_to_versor(H)
+    return unitize_versor(H)
 
 
 def holonomy_similarity(H1: np.ndarray, H2: np.ndarray) -> float:
@@ -68,4 +80,4 @@ def holonomy_similarity(H1: np.ndarray, H2: np.ndarray) -> float:
     Compare two holonomies via CGA inner product.
     Used for prompt-level semantic similarity without embedding lookup.
     """
-    return cga_inner(H1, H2)
+    return cga_inner(unitize_versor(H1), unitize_versor(H2))