tnfr 4.5.2__py3-none-any.whl → 7.0.0__py3-none-any.whl

tnfr/metrics/sense_index.py

@@ -1,41 +1,321 @@
-"""Sense index helpers."""
+"""Sense index helpers for coherence monitoring.
+
+This module implements the Sense index (Si) as a structural telemetry channel
+that blends three signals: the node's structural frequency ``νf`` (how quickly
+it reorganises), its phase coupling with neighbours (whether it stays locked to
+their resonance), and the damping imposed by ``ΔNFR`` (internal reorganisation
+pressure). By combining these contributions we can monitor how each node
+maintains coherence inside a TNFR graph and surface whether the network is
+favouring rapid reorganisation, synchrony, or ΔNFR attenuation.
+
+Examples
+--------
+Build a minimal resonance graph where Si highlights how the structural weights
+steer the interpretation of coherence.
+
+>>> import networkx as nx
+>>> from tnfr.metrics.sense_index import compute_Si
+>>> G = nx.Graph()
+>>> G.add_edge("sensor", "relay")
+>>> G.nodes["sensor"].update({"nu_f": 0.9, "delta_nfr": 0.3, "phase": 0.0})
+>>> G.nodes["relay"].update({"nu_f": 0.4, "delta_nfr": 0.05, "phase": 0.1})
+>>> G.graph["SI_WEIGHTS"] = {"alpha": 0.5, "beta": 0.3, "gamma": 0.2}
+>>> result = compute_Si(G, inplace=False)
+>>> round(result["sensor"], 3), round(result["relay"], 3)
+(0.767, 0.857)
+
+The heavier ``alpha`` weight privileges the ``sensor`` node's fast ``νf`` even
+though it suffers a larger ``ΔNFR``. Conversely, the ``relay`` keeps Si high
+thanks to a calmer ``ΔNFR`` profile despite slower frequency, illustrating how
+Si exposes the trade-off between structural cadence, phase alignment, and
+internal reorganisation pressure.
+"""
 
 from __future__ import annotations
 
 import math
+from concurrent.futures import ProcessPoolExecutor
 from functools import partial
-from typing import Any
+from time import perf_counter
+from typing import Any, Callable, Iterable, Iterator, Mapping, MutableMapping, cast
 
 from ..alias import get_attr, set_attr
-from ..collections_utils import normalize_weights
-from ..constants import get_aliases
-from ..cache import edge_version_cache, stable_json
-from ..helpers.numeric import angle_diff, clamp01
-from .trig import neighbor_phase_mean_list
-from ..import_utils import get_numpy
-from ..types import GraphLike
-
+from ..constants.aliases import ALIAS_DNFR, ALIAS_SI, ALIAS_VF
+from ..utils import angle_diff, angle_diff_array, clamp01
+from ..types import GraphLike, NodeAttrMap
+from ..utils import (
+    edge_version_cache,
+    get_numpy,
+    normalize_weights,
+    resolve_chunk_size,
+    stable_json,
+)
 from .common import (
+    _coerce_jobs,
+    _get_vf_dnfr_max,
     ensure_neighbors_map,
     merge_graph_weights,
-    _get_vf_dnfr_max,
 )
+from .trig import neighbor_phase_mean_bulk, neighbor_phase_mean_list
 from .trig_cache import get_trig_cache
 
-ALIAS_VF = get_aliases("VF")
-ALIAS_DNFR = get_aliases("DNFR")
-ALIAS_SI = get_aliases("SI")
-ALIAS_THETA = get_aliases("THETA")
-
+PHASE_DISPERSION_KEY = "dSi_dphase_disp"
+_SI_APPROX_BYTES_PER_NODE = 64
+_VALID_SENSITIVITY_KEYS = frozenset(
+    {"dSi_dvf_norm", PHASE_DISPERSION_KEY, "dSi_ddnfr_norm"}
+)
 __all__ = ("get_Si_weights", "compute_Si_node", "compute_Si")
 
 
+class _SiStructuralCache:
+    """Cache aligned ``νf`` and ``ΔNFR`` arrays for vectorised Si."""
+
+    __slots__ = ("node_ids", "vf_values", "dnfr_values", "vf_snapshot", "dnfr_snapshot")
+
+    def __init__(self, node_ids: tuple[Any, ...]):
+        self.node_ids = node_ids
+        self.vf_values: Any | None = None
+        self.dnfr_values: Any | None = None
+        self.vf_snapshot: list[float] = []
+        self.dnfr_snapshot: list[float] = []
+
+    def rebuild(
+        self,
+        node_ids: Iterable[Any],
+        node_data: Mapping[Any, NodeAttrMap],
+        *,
+        np: Any,
+    ) -> tuple[Any, Any]:
+        node_tuple = tuple(node_ids)
+        count = len(node_tuple)
+        if count == 0:
+            self.node_ids = node_tuple
+            self.vf_values = np.zeros(0, dtype=float)
+            self.dnfr_values = np.zeros(0, dtype=float)
+            self.vf_snapshot = []
+            self.dnfr_snapshot = []
+            return self.vf_values, self.dnfr_values
+
+        vf_arr = np.fromiter(
+            (float(get_attr(node_data[n], ALIAS_VF, 0.0)) for n in node_tuple),
+            dtype=float,
+            count=count,
+        )
+        dnfr_arr = np.fromiter(
+            (float(get_attr(node_data[n], ALIAS_DNFR, 0.0)) for n in node_tuple),
+            dtype=float,
+            count=count,
+        )
+
+        self.node_ids = node_tuple
+        self.vf_values = vf_arr
+        self.dnfr_values = dnfr_arr
+        self.vf_snapshot = [float(value) for value in vf_arr]
+        self.dnfr_snapshot = [float(value) for value in dnfr_arr]
+        return self.vf_values, self.dnfr_values
+
+    def ensure_current(
+        self,
+        node_ids: Iterable[Any],
+        node_data: Mapping[Any, NodeAttrMap],
+        *,
+        np: Any,
+    ) -> tuple[Any, Any]:
+        node_tuple = tuple(node_ids)
+        if node_tuple != self.node_ids:
+            return self.rebuild(node_tuple, node_data, np=np)
+
+        for idx, node in enumerate(node_tuple):
+            nd = node_data[node]
+            vf = float(get_attr(nd, ALIAS_VF, 0.0))
+            if vf != self.vf_snapshot[idx]:
+                return self.rebuild(node_tuple, node_data, np=np)
+            dnfr = float(get_attr(nd, ALIAS_DNFR, 0.0))
+            if dnfr != self.dnfr_snapshot[idx]:
+                return self.rebuild(node_tuple, node_data, np=np)
+
+        return self.vf_values, self.dnfr_values
+
+
+def _build_structural_cache(
+    node_ids: Iterable[Any],
+    node_data: Mapping[Any, NodeAttrMap],
+    *,
+    np: Any,
+) -> _SiStructuralCache:
+    cache = _SiStructuralCache(tuple(node_ids))
+    cache.rebuild(node_ids, node_data, np=np)
+    return cache
+
+
+def _ensure_structural_arrays(
+    G: GraphLike,
+    node_ids: Iterable[Any],
+    node_data: Mapping[Any, NodeAttrMap],
+    *,
+    np: Any,
+) -> tuple[Any, Any]:
+    node_key = tuple(node_ids)
+
+    def builder() -> _SiStructuralCache:
+        return _build_structural_cache(node_key, node_data, np=np)
+
+    cache = edge_version_cache(G, ("_si_structural", node_key), builder)
+    return cache.ensure_current(node_key, node_data, np=np)
+
+
+def _ensure_si_buffers(
+    G: GraphLike,
+    *,
+    count: int,
+    np: Any,
+) -> tuple[Any, Any, Any]:
+    """Return reusable NumPy buffers sized for ``count`` nodes."""
+
+    def builder() -> tuple[Any, Any, Any]:
+        return (
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+        )
+
+    return edge_version_cache(
+        G,
+        ("_si_buffers", count),
+        builder,
+    )
+
+
+def _ensure_chunk_workspace(
+    G: GraphLike,
+    *,
+    mask_count: int,
+    np: Any,
+) -> tuple[Any, Any]:
+    """Return reusable scratch buffers sized to the masked neighbours."""
+
+    if mask_count <= 0:
+        mask_count = 1
+
+    def builder() -> tuple[Any, Any]:
+        return (
+            np.empty(mask_count, dtype=float),
+            np.empty(mask_count, dtype=float),
+        )
+
+    return edge_version_cache(
+        G,
+        ("_si_chunk_workspace", mask_count),
+        builder,
+    )
+
+
+def _ensure_neighbor_bulk_buffers(
+    G: GraphLike,
+    *,
+    count: int,
+    np: Any,
+) -> tuple[Any, Any, Any, Any, Any]:
+    """Return reusable buffers for bulk neighbour phase aggregation."""
+
+    def builder() -> tuple[Any, Any, Any, Any, Any]:
+        return (
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+            np.empty(count, dtype=float),
+        )
+
+    return edge_version_cache(
+        G,
+        ("_si_neighbor_buffers", count),
+        builder,
+    )
+
+
+def _normalise_si_sensitivity_mapping(
+    mapping: Mapping[str, float], *, warn: bool
+) -> dict[str, float]:
+    """Preserve structural sensitivities compatible with the Si operator.
+
+    Parameters
+    ----------
+    mapping : Mapping[str, float]
+        Mapping of raw sensitivity weights keyed by structural derivatives.
+    warn : bool
+        Compatibility flag kept for trace helpers. It is not used directly but
+        retained so upstream logging keeps a consistent signature.
+
+    Returns
+    -------
+    dict[str, float]
+        Sanitised mapping containing only the supported sensitivity keys.
+
+    Raises
+    ------
+    ValueError
+        If the mapping defines keys outside of the supported sensitivity set.
+
+    Examples
+    --------
+    >>> _normalise_si_sensitivity_mapping({"dSi_dvf_norm": 1.0}, warn=False)
+    {'dSi_dvf_norm': 1.0}
+    >>> _normalise_si_sensitivity_mapping({"unknown": 1.0}, warn=False)
+    Traceback (most recent call last):
+        ...
+    ValueError: Si sensitivity mappings accept only {dSi_ddnfr_norm, dSi_dphase_disp, dSi_dvf_norm}; unexpected key(s): unknown
+    """
+
+    normalised = dict(mapping)
+    _ = warn  # kept for API compatibility with trace helpers
+    unexpected = sorted(k for k in normalised if k not in _VALID_SENSITIVITY_KEYS)
+    if unexpected:
+        allowed = ", ".join(sorted(_VALID_SENSITIVITY_KEYS))
+        received = ", ".join(unexpected)
+        raise ValueError(
+            "Si sensitivity mappings accept only {%s}; unexpected key(s): %s"
+            % (allowed, received)
+        )
+    return normalised
+
+
 def _cache_weights(G: GraphLike) -> tuple[float, float, float]:
-    """Normalise and cache Si weights, delegating persistence."""
+    """Normalise and persist Si weights attached to the graph coherence.
+
+    Parameters
+    ----------
+    G : GraphLike
+        Graph structure whose global Si sensitivities must be harmonised.
+
+    Returns
+    -------
+    tuple[float, float, float]
+        Ordered tuple ``(alpha, beta, gamma)`` with normalised Si weights.
+
+    Raises
+    ------
+    ValueError
+        Propagated if the graph stores unsupported sensitivity keys.
+
+    Examples
+    --------
+    >>> import networkx as nx
+    >>> G = nx.Graph()
+    >>> G.graph["SI_WEIGHTS"] = {"alpha": 0.2, "beta": 0.5, "gamma": 0.3}
+    >>> tuple(round(v, 2) for v in _cache_weights(G))
+    (0.2, 0.5, 0.3)
+    """
 
     w = merge_graph_weights(G, "SI_WEIGHTS")
     cfg_key = stable_json(w)
 
+    existing = G.graph.get("_Si_sensitivity")
+    if isinstance(existing, Mapping):
+        migrated = _normalise_si_sensitivity_mapping(existing, warn=True)
+        if migrated != existing:
+            G.graph["_Si_sensitivity"] = migrated
+
     def builder() -> tuple[float, float, float]:
         weights = normalize_weights(w, ("alpha", "beta", "gamma"), default=0.0)
         alpha = weights["alpha"]
@@ -45,7 +325,7 @@ def _cache_weights(G: GraphLike) -> tuple[float, float, float]:
         G.graph["_Si_weights_key"] = cfg_key
         G.graph["_Si_sensitivity"] = {
             "dSi_dvf_norm": alpha,
-            "dSi_ddisp_fase": -beta,
+            PHASE_DISPERSION_KEY: -beta,
             "dSi_ddnfr_norm": -gamma,
         }
         return alpha, beta, gamma
@@ -54,7 +334,25 @@ def _cache_weights(G: GraphLike) -> tuple[float, float, float]:
 
 
 def get_Si_weights(G: GraphLike) -> tuple[float, float, float]:
-    """Obtain and normalise weights for the sense index."""
+    """Expose the normalised Si weights associated with ``G``.
+
+    Parameters
+    ----------
+    G : GraphLike
+        Graph that carries optional ``SI_WEIGHTS`` metadata.
+
+    Returns
+    -------
+    tuple[float, float, float]
+        The ``(alpha, beta, gamma)`` weights after normalisation.
+
+    Examples
+    --------
+    >>> import networkx as nx
+    >>> G = nx.Graph()
+    >>> get_Si_weights(G)
+    (0.0, 0.0, 0.0)
+    """
 
     return _cache_weights(G)
 
@@ -68,10 +366,73 @@ def compute_Si_node(
     gamma: float,
     vfmax: float,
     dnfrmax: float,
-    disp_fase: float,
+    phase_dispersion: float | None = None,
     inplace: bool,
+    **kwargs: Any,
 ) -> float:
-    """Compute ``Si`` for a single node."""
+    """Evaluate how a node's structure tilts Si within its local resonance.
+
+    Parameters
+    ----------
+    n : Any
+        Node identifier whose structural perception is computed.
+    nd : dict[str, Any]
+        Mutable node attributes containing cached structural magnitudes.
+    alpha : float
+        Normalised weight applied to the node's structural frequency, boosting
+        Si when the node reorganises faster than the network baseline.
+    beta : float
+        Normalised weight applied to the phase alignment term so that tighter
+        synchrony raises the index.
+    gamma : float
+        Normalised weight applied to the ΔNFR attenuation term, rewarding nodes
+        that keep internal turbulence under control.
+    vfmax : float
+        Maximum structural frequency used for normalisation.
+    dnfrmax : float
+        Maximum |ΔNFR| used for normalisation.
+    phase_dispersion : float, optional
+        Phase dispersion ratio in ``[0, 1]`` for the node against its
+        neighbours. The value must be supplied by the caller.
+    inplace : bool
+        Whether to write the resulting Si back to ``nd``.
+    **kwargs : Any
+        Additional keyword arguments are not accepted and will raise.
+
+    Returns
+    -------
+    float
+        The clamped Si value in ``[0, 1]``.
+
+    Raises
+    ------
+    TypeError
+        If ``phase_dispersion`` is missing or unsupported keyword arguments
+        are provided.
+
+    Examples
+    --------
+    >>> nd = {"nu_f": 1.0, "delta_nfr": 0.1}
+    >>> compute_Si_node(
+    ...     "n0",
+    ...     nd,
+    ...     alpha=0.4,
+    ...     beta=0.3,
+    ...     gamma=0.3,
+    ...     vfmax=1.0,
+    ...     dnfrmax=1.0,
+    ...     phase_dispersion=0.2,
+    ...     inplace=False,
+    ... )
+    0.91
+    """
+
+    if kwargs:
+        unexpected = ", ".join(sorted(kwargs))
+        raise TypeError(f"Unexpected keyword argument(s): {unexpected}")
+
+    if phase_dispersion is None:
+        raise TypeError("Missing required keyword-only argument: 'phase_dispersion'")
 
     vf = get_attr(nd, ALIAS_VF, 0.0)
     vf_norm = clamp01(abs(vf) / vfmax)
@@ -79,42 +440,625 @@ def compute_Si_node(
     dnfr = get_attr(nd, ALIAS_DNFR, 0.0)
     dnfr_norm = clamp01(abs(dnfr) / dnfrmax)
 
-    Si = alpha * vf_norm + beta * (1.0 - disp_fase) + gamma * (1.0 - dnfr_norm)
+    Si = alpha * vf_norm + beta * (1.0 - phase_dispersion) + gamma * (1.0 - dnfr_norm)
     Si = clamp01(Si)
     if inplace:
         set_attr(nd, ALIAS_SI, Si)
     return Si
 
 
-def compute_Si(G: GraphLike, *, inplace: bool = True) -> dict[Any, float]:
-    """Compute ``Si`` per node and optionally store it on the graph."""
+def _compute_si_python_chunk(
+    chunk: Iterable[tuple[Any, tuple[Any, ...], float, float, float]],
+    *,
+    cos_th: dict[Any, float],
+    sin_th: dict[Any, float],
+    alpha: float,
+    beta: float,
+    gamma: float,
+    vfmax: float,
+    dnfrmax: float,
+) -> dict[Any, float]:
+    """Propagate Si contributions for a node chunk using pure Python.
+
+    The fallback keeps the νf/phase/ΔNFR balance explicit so that structural
+    effects remain traceable even without vectorised support.
+
+    Parameters
+    ----------
+    chunk : Iterable[tuple[Any, tuple[Any, ...], float, float, float]]
+        Iterable of node payloads ``(node, neighbors, theta, vf, dnfr)``.
+    cos_th : dict[Any, float]
+        Cached cosine values keyed by node identifiers.
+    sin_th : dict[Any, float]
+        Cached sine values keyed by node identifiers.
+    alpha : float
+        Normalised weight for structural frequency.
+    beta : float
+        Normalised weight for phase dispersion.
+    gamma : float
+        Normalised weight for ΔNFR dispersion.
+    vfmax : float
+        Maximum |νf| reference for normalisation.
+    dnfrmax : float
+        Maximum |ΔNFR| reference for normalisation.
+
+    Returns
+    -------
+    dict[Any, float]
+        Mapping of node identifiers to their clamped Si values.
+
+    Examples
+    --------
+    >>> _compute_si_python_chunk(
+    ...     [("n0", ("n1",), 0.0, 0.5, 0.1)],
+    ...     cos_th={"n1": 1.0},
+    ...     sin_th={"n1": 0.0},
+    ...     alpha=0.5,
+    ...     beta=0.3,
+    ...     gamma=0.2,
+    ...     vfmax=1.0,
+    ...     dnfrmax=1.0,
+    ... )
+    {'n0': 0.73}
+    """
+
+    results: dict[Any, float] = {}
+    for n, neigh, theta, vf, dnfr in chunk:
+        th_bar = neighbor_phase_mean_list(
+            neigh, cos_th=cos_th, sin_th=sin_th, np=None, fallback=theta
+        )
+        phase_dispersion = abs(angle_diff(theta, th_bar)) / math.pi
+        vf_norm = clamp01(abs(vf) / vfmax)
+        dnfr_norm = clamp01(abs(dnfr) / dnfrmax)
+        Si = (
+            alpha * vf_norm
+            + beta * (1.0 - phase_dispersion)
+            + gamma * (1.0 - dnfr_norm)
+        )
+        results[n] = clamp01(Si)
+    return results
+
+
+def _iter_python_payload_chunks(
+    nodes_data: Iterable[tuple[Any, NodeAttrMap]],
+    *,
+    neighbors: Mapping[Any, Iterable[Any]],
+    thetas: Mapping[Any, float],
+    chunk_size: int,
+    ) -> Iterator[tuple[tuple[Any, tuple[Any, ...], float, float, float], ...]]:
+    """Yield lazily constructed Si payload chunks for the Python fallback.
+
+    Each batch keeps the structural triad explicit—θ, νf, and ΔNFR—so that the
+    downstream worker preserves the coherence balance enforced by the Si
+    operator.  Streaming prevents a single monolithic buffer that would skew
+    memory pressure on dense graphs while still producing deterministic ΔNFR
+    sampling. The iterator is consumed lazily by :func:`compute_Si` so that the
+    Python fallback can submit and harvest chunk results incrementally, keeping
+    both memory usage and profiling telemetry representative of the streamed
+    execution.
+    """
+
+    if chunk_size <= 0:
+        return
+
+    buffer: list[tuple[Any, tuple[Any, ...], float, float, float]] = []
+    for node, data in nodes_data:
+        theta = thetas.get(node, 0.0)
+        vf = float(get_attr(data, ALIAS_VF, 0.0))
+        dnfr = float(get_attr(data, ALIAS_DNFR, 0.0))
+        neigh = tuple(neighbors[node])
+        buffer.append((node, neigh, theta, vf, dnfr))
+        if len(buffer) >= chunk_size:
+            yield tuple(buffer)
+            buffer.clear()
+
+    if buffer:
+        yield tuple(buffer)
+
+
+def compute_Si(
+    G: GraphLike,
+    *,
+    inplace: bool = True,
+    n_jobs: int | None = None,
+    chunk_size: int | None = None,
+    profile: MutableMapping[str, Any] | None = None,
+) -> dict[Any, float] | Any:
+    """Compute the Si metric for each node by integrating structural drivers.
+
+    Si (sense index) quantifies how effectively a node sustains coherent
+    reorganisation within the TNFR triad. The metric aggregates three
+    structural contributions: the node's structural frequency (weighted by
+    ``alpha``), its phase alignment with neighbours (weighted by ``beta``),
+    and the attenuation of disruptive ΔNFR (weighted by ``gamma``). The
+    weights therefore bias Si towards faster reorganisation, tighter phase
+    coupling, or reduced dissonance respectively, depending on the scenario.
+
+    Parameters
+    ----------
+    G : GraphLike
+        Graph that exposes ``νf`` (structural frequency), ``ΔNFR`` and phase
+        attributes for each node.
+    inplace : bool, default: True
+        If ``True`` the resulting Si values are written back to ``G``.
+    n_jobs : int or None, optional
+        Maximum number of worker processes for the pure-Python fallback. Use
+        ``None`` to auto-detect the configuration.
+    chunk_size : int or None, optional
+        Maximum number of nodes processed per batch when building the Si
+        mapping. ``None`` derives a safe value from the node count, the
+        available CPUs, and conservative memory heuristics. Non-positive values
+        fall back to the automatic mode. Graphs may also provide a default via
+        ``G.graph["SI_CHUNK_SIZE"]``.
+    profile : MutableMapping[str, Any] or None, optional
+        Mutable mapping that aggregates wall-clock durations for the internal
+        stages of the computation. The mapping receives the keys
+        ``"cache_rebuild"``, ``"neighbor_phase_mean_bulk"``,
+        ``"normalize_clamp"`` and ``"inplace_write"`` accumulating seconds for
+        each step, plus ``"path"`` describing whether the vectorised (NumPy)
+        or fallback implementation executed the call. When the Python fallback
+        streams chunk execution, ``"fallback_chunks"`` records how many payload
+        batches completed. Reusing the mapping across invocations accumulates
+        the timings and chunk counts.
+
+    Returns
+    -------
+    dict[Any, float] | numpy.ndarray
+        Mapping from node identifiers to their Si scores when ``inplace`` is
+        ``False``. When ``inplace`` is ``True`` and the NumPy accelerated path
+        is available the function updates the graph in place and returns the
+        vector of Si values as a :class:`numpy.ndarray`. The pure-Python
+        fallback always returns a mapping for compatibility.
+
+    Raises
+    ------
+    ValueError
+        Propagated if graph-level sensitivity settings include unsupported
+        keys or invalid weights.
+
+    Examples
+    --------
+    Build a minimal resonance graph with two nodes sharing a phase-locked
+    edge. The structural weights bias the result towards phase coherence.
+
+    >>> import networkx as nx
+    >>> from tnfr.metrics.sense_index import compute_Si
+    >>> G = nx.Graph()
+    >>> G.add_edge("a", "b")
+    >>> G.nodes["a"].update({"nu_f": 0.8, "delta_nfr": 0.2, "phase": 0.0})
+    >>> G.nodes["b"].update({"nu_f": 0.6, "delta_nfr": 0.1, "phase": 0.1})
+    >>> G.graph["SI_WEIGHTS"] = {"alpha": 0.3, "beta": 0.5, "gamma": 0.2}
+    >>> {k: round(v, 3) for k, v in compute_Si(G, inplace=False).items()}
+    {'a': 0.784, 'b': 0.809}
+    """
+
+    if profile is not None:
+        for key in (
+            "cache_rebuild",
+            "neighbor_phase_mean_bulk",
+            "normalize_clamp",
+            "inplace_write",
+            "fallback_chunks",
+        ):
+            profile.setdefault(key, 0.0)
+
+        def _profile_start() -> float:
+            return perf_counter()
+
+        def _profile_stop(key: str, start: float) -> None:
+            profile[key] = float(profile.get(key, 0.0)) + (perf_counter() - start)
+
+        def _profile_mark_path(path: str) -> None:
+            profile["path"] = path
+
+    else:
+
+        def _profile_start() -> float:
+            return 0.0
+
+        def _profile_stop(key: str, start: float) -> None:
+            return None
+
+        def _profile_mark_path(path: str) -> None:
+            return None
 
     neighbors = ensure_neighbors_map(G)
     alpha, beta, gamma = get_Si_weights(G)
-    vfmax, dnfrmax = _get_vf_dnfr_max(G)
-
     np = get_numpy()
     trig = get_trig_cache(G, np=np)
     cos_th, sin_th, thetas = trig.cos, trig.sin, trig.theta
 
-    pm_fn = partial(
-        neighbor_phase_mean_list, cos_th=cos_th, sin_th=sin_th, np=np
+    pm_fn = partial(neighbor_phase_mean_list, cos_th=cos_th, sin_th=sin_th, np=np)
+
+    if n_jobs is None:
+        n_jobs = _coerce_jobs(G.graph.get("SI_N_JOBS"))
+    else:
+        n_jobs = _coerce_jobs(n_jobs)
+
+    supports_vector = (
+        np is not None
+        and hasattr(np, "ndarray")
+        and all(
+            hasattr(np, attr)
+            for attr in (
+                "fromiter",
+                "abs",
+                "clip",
+                "remainder",
+                "zeros",
+                "add",
+                "bincount",
+                "arctan2",
+                "where",
+                "divide",
+                "errstate",
+                "max",
+            )
+        )
     )
 
-    out: dict[Any, float] = {}
-    for n, nd in G.nodes(data=True):
-        neigh = neighbors[n]
-        th_bar = pm_fn(neigh, fallback=thetas[n])
-        disp_fase = abs(angle_diff(thetas[n], th_bar)) / math.pi
-        out[n] = compute_Si_node(
-            n,
-            nd,
-            alpha=alpha,
-            beta=beta,
-            gamma=gamma,
-            vfmax=vfmax,
-            dnfrmax=dnfrmax,
-            disp_fase=disp_fase,
-            inplace=inplace,
+    nodes_view = G.nodes
+    nodes_data = list(nodes_view(data=True))
+    if not nodes_data:
+        return {}
+
+    node_mapping = cast(Mapping[Any, NodeAttrMap], nodes_view)
+    node_count = len(nodes_data)
+
+    trig_order = list(getattr(trig, "order", ()))
+    node_ids: list[Any]
+    node_idx: dict[Any, int]
+    using_cache_order = False
+    if trig_order and len(trig_order) == node_count:
+        node_ids = trig_order
+        node_idx = dict(getattr(trig, "index", {}))
+        using_cache_order = len(node_idx) == len(node_ids)
+        if not using_cache_order:
+            node_idx = {n: i for i, n in enumerate(node_ids)}
+    else:
+        node_ids = [n for n, _ in nodes_data]
+        node_idx = {n: i for i, n in enumerate(node_ids)}
+
+    chunk_pref = chunk_size if chunk_size is not None else G.graph.get("SI_CHUNK_SIZE")
+
+    if supports_vector:
+        _profile_mark_path("vectorized")
+        node_key = tuple(node_ids)
+        count = len(node_key)
+
+        cache_theta = getattr(trig, "theta_values", None)
+        cache_cos = getattr(trig, "cos_values", None)
+        cache_sin = getattr(trig, "sin_values", None)
+
+        trig_index_map = dict(getattr(trig, "index", {}) or {})
+        index_arr: Any | None = None
+        cached_mask = None
+        if trig_index_map and count:
+            index_values: list[int] = []
+            mask_values: list[bool] = []
+            for node in node_ids:
+                cached_idx = trig_index_map.get(node)
+                if cached_idx is None:
+                    index_values.append(-1)
+                    mask_values.append(False)
+                else:
+                    index_values.append(int(cached_idx))
+                    mask_values.append(True)
+            cached_mask = np.asarray(mask_values, dtype=bool)
+            if cached_mask.any():
+                index_arr = np.asarray(index_values, dtype=np.intp)
+        if cached_mask is None:
+            cached_mask = np.zeros(count, dtype=bool)
+
+        def _gather_values(cache_values: Any | None, fallback_getter: Callable[[Any], float]) -> Any:
+            if (
+                index_arr is not None
+                and cache_values is not None
+                and cached_mask.size
+                and cached_mask.any()
+            ):
+                out = np.empty(count, dtype=float)
+                cached_indices = np.nonzero(cached_mask)[0]
+                if cached_indices.size:
+                    out[cached_indices] = np.take(
+                        np.asarray(cache_values, dtype=float), index_arr[cached_indices]
+                    )
+                missing_indices = np.nonzero(~cached_mask)[0]
+                if missing_indices.size:
+                    missing_nodes = [node_ids[i] for i in missing_indices]
+                    out[missing_indices] = np.fromiter(
+                        (fallback_getter(node) for node in missing_nodes),
+                        dtype=float,
+                        count=missing_indices.size,
+                    )
+                return out
+            return np.fromiter(
+                (fallback_getter(node) for node in node_ids),
+                dtype=float,
+                count=count,
+            )
+
+        cache_timer = _profile_start()
+
+        if using_cache_order and cache_theta is not None:
+            theta_arr = np.asarray(cache_theta, dtype=float)
+        else:
+            theta_arr = _gather_values(cache_theta, lambda node: thetas.get(node, 0.0))
+
+        if using_cache_order and cache_cos is not None:
+            cos_arr = np.asarray(cache_cos, dtype=float)
+        else:
+            cos_arr = _gather_values(
+                cache_cos,
+                lambda node: cos_th.get(node, math.cos(thetas.get(node, 0.0))),
+            )
+
+        if using_cache_order and cache_sin is not None:
+            sin_arr = np.asarray(cache_sin, dtype=float)
+        else:
+            sin_arr = _gather_values(
+                cache_sin,
+                lambda node: sin_th.get(node, math.sin(thetas.get(node, 0.0))),
+            )
+
+        cached_edge_src = None
+        cached_edge_dst = None
+        if using_cache_order:
+            cached_edge_src = getattr(trig, "edge_src", None)
+            cached_edge_dst = getattr(trig, "edge_dst", None)
+            if cached_edge_src is not None and cached_edge_dst is not None:
+                cached_edge_src = np.asarray(cached_edge_src, dtype=np.intp)
+                cached_edge_dst = np.asarray(cached_edge_dst, dtype=np.intp)
+                if cached_edge_src.shape != cached_edge_dst.shape:
+                    cached_edge_src = None
+                    cached_edge_dst = None
+
+        if cached_edge_src is not None and cached_edge_dst is not None:
+            edge_src = cached_edge_src
+            edge_dst = cached_edge_dst
+        else:
+
+            def _build_edge_arrays() -> tuple[Any, Any]:
+                edge_src_list: list[int] = []
+                edge_dst_list: list[int] = []
+                for node in node_ids:
+                    dst_idx = node_idx[node]
+                    for neighbor in neighbors[node]:
+                        src_idx = node_idx.get(neighbor)
+                        if src_idx is None:
+                            continue
+                        edge_src_list.append(src_idx)
+                        edge_dst_list.append(dst_idx)
+                src_arr = np.asarray(edge_src_list, dtype=np.intp)
+                dst_arr = np.asarray(edge_dst_list, dtype=np.intp)
+                return src_arr, dst_arr
+
+            edge_src, edge_dst = edge_version_cache(
+                G,
+                ("_si_edges", node_key),
+                _build_edge_arrays,
+            )
+            if using_cache_order:
+                trig.edge_src = edge_src
+                trig.edge_dst = edge_dst
+
+        (
+            neighbor_cos_sum,
+            neighbor_sin_sum,
+            neighbor_counts,
+            mean_cos_buf,
+            mean_sin_buf,
+        ) = _ensure_neighbor_bulk_buffers(
+            G,
+            count=count,
+            np=np,
         )
+
+        vf_arr, dnfr_arr = _ensure_structural_arrays(
+            G,
+            node_ids,
+            node_mapping,
+            np=np,
+        )
+        raw_vfmax = float(np.max(np.abs(vf_arr))) if getattr(vf_arr, "size", 0) else 0.0
+        raw_dnfrmax = (
+            float(np.max(np.abs(dnfr_arr))) if getattr(dnfr_arr, "size", 0) else 0.0
+        )
+        G.graph["_vfmax"] = raw_vfmax
+        G.graph["_dnfrmax"] = raw_dnfrmax
+        vfmax = 1.0 if raw_vfmax == 0.0 else raw_vfmax
+        dnfrmax = 1.0 if raw_dnfrmax == 0.0 else raw_dnfrmax
+
+        (
+            phase_dispersion,
+            raw_si,
+            si_values,
+        ) = _ensure_si_buffers(
+            G,
+            count=count,
+            np=np,
+        )
+
+        _profile_stop("cache_rebuild", cache_timer)
+
+        neighbor_timer = _profile_start()
+        mean_theta, has_neighbors = neighbor_phase_mean_bulk(
+            edge_src,
+            edge_dst,
+            cos_values=cos_arr,
+            sin_values=sin_arr,
+            theta_values=theta_arr,
+            node_count=count,
+            np=np,
+            neighbor_cos_sum=neighbor_cos_sum,
+            neighbor_sin_sum=neighbor_sin_sum,
+            neighbor_counts=neighbor_counts,
+            mean_cos=mean_cos_buf,
+            mean_sin=mean_sin_buf,
+        )
+        _profile_stop("neighbor_phase_mean_bulk", neighbor_timer)
+        norm_timer = _profile_start()
+        # Reuse the Si buffers as scratch space to avoid transient allocations during
+        # the normalization pass and keep the structural buffers coherent with the
+        # cached layout.
+        np.abs(vf_arr, out=raw_si)
+        np.divide(raw_si, vfmax, out=raw_si)
+        np.clip(raw_si, 0.0, 1.0, out=raw_si)
+        vf_norm = raw_si
+        np.abs(dnfr_arr, out=si_values)
+        np.divide(si_values, dnfrmax, out=si_values)
+        np.clip(si_values, 0.0, 1.0, out=si_values)
+        dnfr_norm = si_values
+        phase_dispersion.fill(0.0)
+        neighbor_mask = np.asarray(has_neighbors, dtype=bool)
+        neighbor_count = int(neighbor_mask.sum())
+        use_chunked = False
+        if neighbor_count:
+            effective_chunk = resolve_chunk_size(
+                chunk_pref,
+                neighbor_count,
+                approx_bytes_per_item=_SI_APPROX_BYTES_PER_NODE,
+            )
+            if effective_chunk <= 0 or effective_chunk >= neighbor_count:
+                effective_chunk = neighbor_count
+            else:
+                use_chunked = True
+
+        if neighbor_count and not use_chunked:
+            angle_diff_array(
+                theta_arr,
+                mean_theta,
+                np=np,
+                out=phase_dispersion,
+                where=neighbor_mask,
+            )
+            np.abs(phase_dispersion, out=phase_dispersion, where=neighbor_mask)
+            np.divide(
+                phase_dispersion,
+                math.pi,
+                out=phase_dispersion,
+                where=neighbor_mask,
+            )
+        elif neighbor_count and use_chunked:
+            neighbor_indices = np.nonzero(neighbor_mask)[0]
+            chunk_theta, chunk_values = _ensure_chunk_workspace(
+                G,
+                mask_count=neighbor_count,
+                np=np,
+            )
+            for start in range(0, neighbor_count, effective_chunk):
+                end = min(start + effective_chunk, neighbor_count)
+                slice_indices = neighbor_indices[start:end]
+                chunk_len = end - start
+                theta_view = chunk_theta[:chunk_len]
+                values_view = chunk_values[:chunk_len]
+                np.take(theta_arr, slice_indices, out=theta_view)
+                np.take(mean_theta, slice_indices, out=values_view)
+                angle_diff_array(theta_view, values_view, np=np, out=values_view)
+                np.abs(values_view, out=values_view)
+                np.divide(values_view, math.pi, out=values_view)
+                phase_dispersion[slice_indices] = values_view
+        else:
+            np.abs(phase_dispersion, out=phase_dispersion)
+            np.divide(
+                phase_dispersion,
+                math.pi,
+                out=phase_dispersion,
+                where=neighbor_mask,
+            )
+
+        np.multiply(vf_norm, alpha, out=raw_si)
+        np.subtract(1.0, phase_dispersion, out=phase_dispersion)
+        np.multiply(phase_dispersion, beta, out=phase_dispersion)
+        np.add(raw_si, phase_dispersion, out=raw_si)
+        np.subtract(1.0, dnfr_norm, out=si_values)
+        np.multiply(si_values, gamma, out=si_values)
+        np.add(raw_si, si_values, out=raw_si)
+        np.clip(raw_si, 0.0, 1.0, out=si_values)
+
+        _profile_stop("normalize_clamp", norm_timer)
+
+        if inplace:
+            write_timer = _profile_start()
+            for idx, node in enumerate(node_ids):
+                set_attr(G.nodes[node], ALIAS_SI, float(si_values[idx]))
+            _profile_stop("inplace_write", write_timer)
+            return np.copy(si_values)
+
+        return {node: float(value) for node, value in zip(node_ids, si_values)}
+
+    vfmax, dnfrmax = _get_vf_dnfr_max(G)
+
+    out: dict[Any, float] = {}
+    _profile_mark_path("fallback")
+    if n_jobs is not None and n_jobs > 1:
+        node_count = len(nodes_data)
+        if node_count:
+            effective_chunk = resolve_chunk_size(
+                chunk_pref,
+                node_count,
+                approx_bytes_per_item=_SI_APPROX_BYTES_PER_NODE,
+            )
+            if effective_chunk <= 0:
+                effective_chunk = node_count
+            payload_chunks = _iter_python_payload_chunks(
+                nodes_data,
+                neighbors=neighbors,
+                thetas=thetas,
+                chunk_size=effective_chunk,
+            )
+            chunk_count = 0
+            with ProcessPoolExecutor(max_workers=n_jobs) as executor:
+                worker = partial(
+                    _compute_si_python_chunk,
+                    cos_th=cos_th,
+                    sin_th=sin_th,
+                    alpha=alpha,
+                    beta=beta,
+                    gamma=gamma,
+                    vfmax=vfmax,
+                    dnfrmax=dnfrmax,
+                )
+                payload_iter = iter(payload_chunks)
+                futures: list[Any] = []
+                for chunk in payload_iter:
+                    futures.append(executor.submit(worker, chunk))
+                    if len(futures) >= n_jobs:
+                        future = futures.pop(0)
+                        chunk_result = future.result()
+                        chunk_count += 1
+                        out.update(chunk_result)
+                for future in futures:
+                    chunk_result = future.result()
+                    chunk_count += 1
+                    out.update(chunk_result)
+            if profile is not None:
+                profile["fallback_chunks"] = float(profile.get("fallback_chunks", 0.0)) + float(
+                    chunk_count
+                )
+    else:
+        for n, nd in nodes_data:
+            theta = thetas.get(n, 0.0)
+            neigh = neighbors[n]
+            th_bar = pm_fn(neigh, fallback=theta)
+            phase_dispersion = abs(angle_diff(theta, th_bar)) / math.pi
+            norm_timer = _profile_start()
+            out[n] = compute_Si_node(
+                n,
+                nd,
+                alpha=alpha,
+                beta=beta,
+                gamma=gamma,
+                vfmax=vfmax,
+                dnfrmax=dnfrmax,
+                phase_dispersion=phase_dispersion,
+                inplace=False,
+            )
+            _profile_stop("normalize_clamp", norm_timer)
+
+    if inplace:
+        write_timer = _profile_start()
+        for n, value in out.items():
+            set_attr(G.nodes[n], ALIAS_SI, value)
+        _profile_stop("inplace_write", write_timer)
     return out