"""NdFeB SPS physics-model tool: predict, p_meets_spec, eig, condition.

Evidence class: owner-built conditional executable physics model (Dynamical's model fitted to RPI's released records
and declared literature priors; not an RPI or collaborator model); not a physical observation; does not calibrate
current-body coercivity. The fit's assumptions about those records are listed in the README ("Fitted assumptions").

Dependencies: Python standard library + numpy. No JAX. Loads hash-bound artifacts:
  physics.py   - schedule discretisation and rate laws (identical bytes to the fitted model)
  forward.py   - mechanism forward equations (the fitted model's, plus the C4 feedstock particle-size layer, a declared
                 and uncalibrated hook that the fit never reached)
  posterior    - .npz of thinned posterior and prior draws + metadata
  manifest     - TOOL_MANIFEST.json with the sha256 of this module and the three files above (verified on load)

Usage inside a science workspace (files are mounted by sha256, without names or extensions):
    import importlib.util
    from importlib.machinery import SourceFileLoader
    raw = '/workspace/raw/'
    loader = SourceFileLoader('ndfeb_physics_tool', raw + TOOL_SHA)
    spec = importlib.util.spec_from_loader('ndfeb_physics_tool', loader)
    tool = importlib.util.module_from_spec(spec); loader.exec_module(tool)
    pm = tool.load(physics=raw + PHYSICS_SHA, forward=raw + FORWARD_SHA, posterior=raw + POSTERIOR_SHA,
                   manifest=raw + MANIFEST_SHA)
    pm.predict({'T_peak_C': 800, 'pressure_mode': 'cooling', 'P_MPa': 60})

Units: deg C, minutes, MPa, relative density (fraction of 7.62 g/cm3), kOe, kGs, MGOe, um.
"""
from __future__ import annotations

import hashlib
import importlib.util
import json
import math
import sys
import types
from importlib.machinery import SourceFileLoader
from pathlib import Path

import numpy as np

TOOL_VERSION = "ndfeb-sps-physics-tool-20261013.5.0"
EVIDENCE_CLASS = ("owner-built conditional executable physics model (Dynamical's model fitted to RPI's released records "
                  "and declared literature priors; not an RPI or collaborator model); not a physical observation; does "
                  "not calibrate current-body coercivity")
OBJECTIVE = {"Hci_kOe_min": 20.0, "BHmax_MGOe_min": 45.0, "note": "Dy/Tb-free, verified at operating temperature; "
             "10 kOe / 11 MGOe is a historical simulated benchmark from sample E, not the objective"}
RHO_TD = 7.62
HERE = Path(__file__).resolve().parent
_PKG = "ndfeb_physics_model"

QUANTILE_LEVELS = (0.01, 0.025, 0.05, 0.10, 0.25, 0.50, 0.75, 0.90, 0.95, 0.975, 0.99)
QUANTILE_KEYS = ("q01", "q025", "q05", "q10", "q25", "q50", "q75", "q90", "q95", "q975", "q99")
H_MEASUREMENT_LOG_SD = 0.25  # declared, uncalibrated: hydrogen content vs the model's residual-hydrogen index
H_DETECTION_INDEX = 0.01  # declared detection limit of a hydrogen measurement on the residual-hydrogen index scale
# (1 = non-degassed powder); a body below it (e.g. degas_completeness "complete") reads at the limit
# degas_completeness d is a declared, uncalibrated input (a number, a uniform or beta distribution, or "complete").
# The route defaults below are the v4 build defaults; a package may override them through meta["route_defaults"]
# (each time slice states its own defaults from information released by its cutoff).
ROUTE_DEFAULTS = dict(T_peak_C=None, heating_rate_C_per_min=10.0, dwell_min=5.0, P_MPa=60.0, pressure_mode="cooling",
                      cooling_rate_C_per_min=100.0, load_ramp_min=0.3, degas=1.0, anneal="none", alignment_cos=0.5,
                      T_measure_C=20.0, prep="new", degas_completeness=1.0, current_path="compact",
                      feedstock_d50_um=4.0)
# C4 feedstock particle-size layer (declared, uncalibrated): feedstock_d50_um = 4.0 is RPI's stated feedstock, the
# feedstock of every fitted record; any other value is a declared extrapolation (forward.py, "C4 feedstock").
FEEDSTOCK_KEY = "feedstock_d50_um"
FEEDSTOCK_REF_UM = 4.0
FEEDSTOCK_RANGE_UM = (0.5, 8.0)
FEEDSTOCK_PARAMS = ("fd_beta", "fd_sg", "fd_dox", "fd_gox", "fd_kH")
STATUS_FITTED = "fitted_feedstock"
STATUS_DECLARED = "declared_extrapolation_uncalibrated"
OUTPUTS = ("density_rel", "grain_um", "Hci_kOe", "Br_kGs", "Hk_over_Hci", "BHmax_MGOe")
PRESSURE_MODES = ("heating", "cooling", "dwell")
ANNEALS = ("none", "900", "900+520")
CURRENT_PATHS = ("compact", "insulated")
LT_LEVELS = ("trap", "load", "current")
SPEC_KEYS = ("Hci_kOe_min", "BHmax_MGOe_min", "Br_kGs_min", "density_rel_min", "T_operating_C")


def _validate_spec(spec):
    """Reject unknown or non-numeric spec keys (a misspelled key must not silently drop a requirement)."""
    if not isinstance(spec, dict) or not spec:
        raise ValueError(f"spec must be a non-empty dict with keys from {SPEC_KEYS}")
    bad = sorted(k for k in spec if k not in SPEC_KEYS)
    if bad:
        raise ValueError(f"unknown spec keys {bad}; allowed: {list(SPEC_KEYS)}")
    for k, v in spec.items():
        if not _is_number(v):
            raise ValueError(f"spec[{k!r}] must be a number")
    return spec


def _sha256(path) -> str:
    return hashlib.sha256(Path(path).read_bytes()).hexdigest()


def _load_module(name, path):
    loader = SourceFileLoader(name, str(path))
    spec = importlib.util.spec_from_loader(name, loader)
    mod = importlib.util.module_from_spec(spec)
    sys.modules[name] = mod
    loader.exec_module(mod)
    return mod


def _compose(physics_path, forward_path):
    """Compose physics.py and forward.py into a synthetic package (forward does `from . import physics`)."""
    pkg = types.ModuleType(_PKG)
    pkg.__path__ = []
    sys.modules[_PKG] = pkg
    ph = _load_module(_PKG + ".physics", physics_path)
    pkg.physics = ph
    fw = _load_module(_PKG + ".forward", forward_path)
    return ph, fw


def load(physics=None, forward=None, posterior=None, manifest=None, verify=True):
    """Load the tool. With a manifest (default: TOOL_MANIFEST.json next to this file) every artifact's sha256 is
    checked and a mismatch raises. Without one the model loads with `verified = False`, reported in every output."""
    physics = physics or HERE / "physics.py"
    forward = forward or HERE / "forward.py"
    posterior = posterior or HERE / "posterior_tool.npz"
    if manifest is None and (HERE / "TOOL_MANIFEST.json").exists():
        manifest = HERE / "TOOL_MANIFEST.json"
    hashes = {"physics": _sha256(physics), "forward": _sha256(forward), "posterior": _sha256(posterior),
              "tool": _sha256(Path(__file__))}
    verified = False
    if verify and manifest is not None:
        man = json.loads(Path(manifest).read_text())
        missing = [k for k in ("physics", "forward", "posterior", "tool") if k not in man.get("files", {})]
        if missing:
            raise ValueError(f"manifest lacks entries {missing}; a v3.1 manifest lists physics, forward, posterior and tool")
        if man.get("tool_version") != TOOL_VERSION:
            raise ValueError(f"manifest is for {man.get('tool_version')}, this module is {TOOL_VERSION}")
        want = {k: man["files"][k]["sha256"] for k in ("physics", "forward", "posterior", "tool")}
        bad = {k: (hashes[k], want[k]) for k in want if hashes[k] != want[k]}
        if bad:
            raise ValueError(f"artifact hashes differ from the manifest: {bad}")
        verified = True
    ph, fw = _compose(physics, forward)
    return PhysicsModel(ph, fw, posterior, hashes, verified=verified)


def _n_logpdf(y, mu, s):
    return -0.5 * ((y - mu) / s) ** 2 - np.log(s) - 0.5 * math.log(2 * math.pi)


def _t_logpdf(y, mu, s, nu=5.0):
    z = (y - mu) / s
    return (math.lgamma((nu + 1) / 2) - math.lgamma(nu / 2) - 0.5 * math.log(nu * math.pi) - np.log(s)
            - (nu + 1) / 2 * np.log1p(z ** 2 / nu))


def _lse(a, axis=None):
    a = np.asarray(a, float)
    m = np.max(a, axis=axis, keepdims=True)
    m = np.where(np.isfinite(m), m, 0.0)
    out = m + np.log(np.sum(np.exp(a - m), axis=axis, keepdims=True))
    return np.squeeze(out, axis=axis) if axis is not None else float(np.squeeze(out))


def _lme(a, axis):
    a = np.asarray(a, float)
    return _lse(a, axis) - math.log(a.shape[axis])


def _iqr(x):
    q25, q75 = np.quantile(x, [0.25, 0.75])
    return float(q75 - q25)


def _distribution(x):
    """Quantiles (the primary summary) and a quantile-matched log-normal approximation of a predictive sample
    (mu_ln = ln median, sd_ln = half the 15.9-84.1 % log spread, over the positive draws), with its adequacy:
    lognormal_max_log_error is the largest |ln q - (mu_ln + z_q sd_ln)| over q05..q95, and lognormal_adequate is true
    when it is below 0.25. Mixtures over mechanisms and degassing states are often not log-normal."""
    from statistics import NormalDist
    x = np.asarray(x, float)
    qs = np.quantile(x, QUANTILE_LEVELS)
    pos = x[x > 0]
    mu = sd = worst = None  # JSON-safe: None where a log-normal summary is not computable (never NaN or inf)
    if len(pos) >= 10:  # quantile-matched (robust): a few collapsed draws must not distort the summary
        l16, l50, l84 = np.log(np.quantile(pos, [0.158655, 0.5, 0.841345]))
        mu, sd = float(l50), float((l84 - l16) / 2.0)
        central = [(q, lv) for q, lv in zip(qs, QUANTILE_LEVELS) if 0.05 <= lv <= 0.95]
        if sd > 0 and all(q > 0 for q, _ in central):
            worst = float(max(abs(math.log(q) - (mu + NormalDist().inv_cdf(lv) * sd)) for q, lv in central))
    return dict(quantiles={k: float(v) for k, v in zip(QUANTILE_KEYS, qs)}, mu_ln=mu, sd_ln=sd,
                lognormal_max_log_error=worst, lognormal_adequate=bool(worst is not None and worst < 0.25),
                frac_nonpositive=float(np.mean(x <= 0)))


def _load_on_cooling(r):
    """Hull feature: 1 load on cooling, 0 load in heating, 0.5 load from the start of the dwell (declared scenario)."""
    return {"cooling": 1.0, "heating": 0.0, "dwell": 0.5}[r["pressure_mode"]]


def _is_number(v):
    import numbers
    return isinstance(v, numbers.Real) and not isinstance(v, bool)


def _check_prep_feedstock(seen, r, what="routes"):
    """Routes (or observations) sharing a preparation share one powder batch, hence one feedstock size."""
    v = float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM))
    if seen.setdefault(r["prep"], v) != v:
        raise ValueError(f"{what} sharing preparation {r['prep']!r} must share {FEEDSTOCK_KEY} (one powder batch)")


def _scenarios(r):
    dc = r["degas_completeness"]
    dc_num = float(dc) if _is_number(dc) else None
    complete = dc == "complete"
    extrap = complete or (dc_num is not None and dc_num > 1.0) or (isinstance(dc, dict) and (
        (dc.get("distribution") == "uniform" and float(dc.get("high", 0)) > 1.0)))
    calibrated = bool(abs(float(r["alignment_cos"]) - 0.5) < 1e-9 and abs(float(r["T_measure_C"]) - 20) < 1e-9
                      and dc_num is not None and dc_num in (0.0, 1.0) and r["current_path"] == "compact"
                      and r["pressure_mode"] in ("heating", "cooling")
                      and float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM)) == FEEDSTOCK_REF_UM)
    return {"alignment_cos": float(r["alignment_cos"]), "T_measure_C": float(r["T_measure_C"]),
            "degas_occurrence_probability": float(r["degas"]), "degas_completeness": dc,
            "degas_extrapolation_beyond_fitted_state": bool(extrap),
            "degas_complete_declared": bool(complete), "current_path": r["current_path"],
            "pressure_mode_dwell_declared": r["pressure_mode"] == "dwell",
            FEEDSTOCK_KEY: float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM)),
            "feedstock_declared": float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM)) != FEEDSTOCK_REF_UM,
            "feedstock_note": ("feedstock_d50_um other than 4 um (the feedstock of every fitted record) is a declared, "
                               "uncalibrated extrapolation: grain size, coercivity (an empirical grain-size factor in "
                               "every mechanism, an oxidation term and a hydrogen term) and the flags change; density "
                               "is the current feedstock's density model, unchanged"),
            "calibrated": calibrated,
            "note": "model assumptions, not records: calibrated only at isotropic alignment, room temperature, the "
                    "two degassing states the fit assigns (the model assumes A-C correspond to completeness 0 and D/E "
                    "to completeness 1; RPI's record does not state A-C's dehydrogenation), a current-carrying compact "
                    "(conductive graphite die) and load in heating or on cooling. Alignment, temperature, intermediate "
                    "completeness (interpolated), completeness above 1 (extrapolated), complete degassing "
                    "(no residual hydrogen; generic degassing terms held at the fitted D/E state), an insulated "
                    "compact and load from the start of the dwell are declared scenarios; batch records or a hydrogen "
                    "measurement would pin the completeness."}


# ---------------------------------------------------------------------------------------------------- eig (v4) helpers
EIG_METHOD_V4 = "bracketed-ba-v4"
EIG_QUANTITIES = ("Hci", "Br", "sq", "density", "grain", "H_ppm", "Hci_ratio_T")
EIG_EPS = 0.01            # every fitted posterior is mixed with the prior weights at this rate (bounds log q below)
EIG_ESS_POOL = (20.0, 200.0)  # pooling weight on the exact nested posterior: 0 below ESS 20, 1 above ESS 200 (log-linear)
EIG_QUAD_DIMS = 36        # quadratic classifier terms use the leading principal components of the latent means
GOAL_BODIES_POOL, GOAL_BODIES_OUTER = 4, 32  # simulated bodies per draw for the success probability p(goal | draw)
GOAL_QUANTILES = (0.05, 0.25, 0.5, 0.75, 0.95)


def _eig_spec_check(spec):
    """Same keys and checks as p_meets_spec; unknown keys raise (a misspelled key must not drop a requirement)."""
    v = globals().get("_validate_spec")
    if v is not None:
        return v(spec)
    allowed = ("Hci_kOe_min", "BHmax_MGOe_min", "Br_kGs_min", "density_rel_min", "T_operating_C")
    if not isinstance(spec, dict) or not spec:
        raise ValueError(f"spec must be a non-empty dict with keys from {allowed}")
    bad = sorted(k for k in spec if k not in allowed)
    if bad:
        raise ValueError(f"unknown spec keys {bad}; allowed: {list(allowed)}")
    for k, val in spec.items():
        if not _is_number(val):
            raise ValueError(f"spec[{k!r}] must be a number")
    return spec


def _lbfgs(fg, x0, iters=200, m=10, tol=1e-8):
    """Minimise a smooth convex f given fg(x) -> (f, grad): two-loop L-BFGS with a backtracking Armijo search."""
    x = np.array(x0, float)
    f, g = fg(x)
    S, Y = [], []
    for _ in range(iters):
        q = g.copy()
        hist = []
        for s, y in reversed(list(zip(S, Y))):
            r = 1.0 / (y @ s)
            a = r * (s @ q)
            q -= a * y
            hist.append((r, a, s, y))
        if S:
            q *= (S[-1] @ Y[-1]) / (Y[-1] @ Y[-1])
        for r, a, s, y in reversed(hist):
            q += s * (a - r * (y @ q))
        d = -q
        gd = g @ d
        if not gd < 0:
            d, gd, S, Y = -g, -(g @ g), [], []
        t = 1.0
        while True:
            xn = x + t * d
            fn, gn = fg(xn)
            if fn <= f + 1e-4 * t * gd or t < 1e-12:
                break
            t *= 0.5
        s, y = xn - x, gn - g
        done = abs(f - fn) <= tol * max(1.0, abs(f))
        if y @ s > 1e-14:
            S.append(s)
            Y.append(y)
            if len(S) > m:
                S.pop(0)
                Y.pop(0)
        x, f, g = xn, fn, gn
        if done:
            break
    return x


class _Features:
    """Classifier / regression inputs: standardised summaries (clipped at 6 sd) and quadratic terms of the leading
    EIG_QUAD_DIMS principal components of the first n_quad columns (the per-latent means). Fitted on training rows."""

    def __init__(self, X, n_quad):
        self.mu, self.sd = X.mean(0), X.std(0) + 1e-9
        Z = np.clip((X - self.mu) / self.sd, -6.0, 6.0)[:, :n_quad]
        k = min(EIG_QUAD_DIMS, Z.shape[1])
        if k:
            lam, V = np.linalg.eigh(np.cov(Z.T).reshape(Z.shape[1], Z.shape[1]))
            self.V = V[:, ::-1][:, :k]
        else:
            self.V = np.zeros((Z.shape[1], 0))
        self.n_quad = n_quad

    def __call__(self, X):
        Z = np.clip((X - self.mu) / self.sd, -6.0, 6.0)
        P = Z[:, :self.n_quad] @ self.V
        k = P.shape[1]
        iu = np.triu_indices(k)
        quad = (P[:, :, None] * P[:, None, :])[:, iu[0], iu[1]] / math.sqrt(2.0) if k else np.zeros((len(Z), 0))
        return np.concatenate([Z, np.clip(quad, -36.0, 36.0)], 1)


class _Softmax:
    """Weighted multinomial logistic regression on _Features, fitted by L-BFGS with a small L2 penalty. Targets may
    be soft (rows of class probabilities)."""

    def __init__(self, X, T, sw, n_quad, l2=1e-4, iters=200):
        self.feat = _Features(X, n_quad)
        Phi = self.feat(X)
        n, p = Phi.shape
        K = T.shape[1]
        swn = sw / sw.sum()

        def fg(x):
            W, b = x[:p * K].reshape(p, K), x[p * K:]
            A = Phi @ W + b
            A -= A.max(1, keepdims=True)
            E = np.exp(A)
            Ssum = E.sum(1, keepdims=True)
            G = (E / Ssum - T) * swn[:, None]
            f = -float(np.sum(swn[:, None] * T * (A - np.log(Ssum)))) + 0.5 * l2 * float(np.sum(W * W))
            return f, np.concatenate([(Phi.T @ G + l2 * W).ravel(), G.sum(0)])
        x = _lbfgs(fg, np.zeros(p * K + K), iters=iters)
        self.W, self.b = x[:p * K].reshape(p, K), x[p * K:]

    def log_prob(self, X):
        A = self.feat(X) @ self.W + self.b
        return A - _lse(A, 1)[:, None]


def _ridge(X, t, sw, Xnew, n_quad, l2=1e-3):
    """Weighted ridge regression of t on [1, _Features(X)]; returns predictions at Xnew."""
    feat = _Features(X, n_quad)
    Phi = np.concatenate([np.ones((len(X), 1)), feat(X)], 1)
    w = sw / sw.sum()
    beta = np.linalg.solve(Phi.T @ (Phi * w[:, None]) + l2 * np.eye(Phi.shape[1]), Phi.T @ (w * t))
    return np.concatenate([np.ones((len(Xnew), 1)), feat(Xnew)], 1) @ beta


def _systematic(p, n, rng):
    """n indices drawn by systematic resampling with probabilities p (every index appears floor/ceil(n p) times)."""
    if n <= 0:
        return np.zeros(0, int)
    c = np.cumsum(p / p.sum())
    u = (rng.random() + np.arange(n)) / n
    return np.minimum(np.searchsorted(c, u), len(p) - 1)


def _alloc(w, n, balanced):
    """Label counts summing to n: equal shares (balanced) or proportional to w (largest remainder)."""
    k = len(w)
    if balanced:
        c = np.full(k, n // k)
        c[: n - c.sum()] += 1
        return c
    raw = n * w / w.sum()
    c = np.floor(raw).astype(int)
    c[np.argsort(-(raw - c))[: n - c.sum()]] += 1
    return c


def _eig_halves(contrib, fold, offset):
    """Lower bound from each half of the simulated experiments and whether they agree within 2 combined SE."""
    est, var = [], []
    for f in (0, 1):
        c = contrib[fold == f]
        est.append(offset + float(c.mean()))
        var.append(float(c.var(ddof=1)) / len(c) if len(c) > 1 else float("inf"))
    return est, bool(abs(est[0] - est[1]) <= 2.0 * math.sqrt(var[0] + var[1]))


def _bern_entropy(p):
    p = np.clip(np.asarray(p, float), 1e-12, 1 - 1e-12)
    return -(p * np.log(p) + (1 - p) * np.log1p(-p))


# ----------------------------------------------------------------------------------------------------------------
class PhysicsModel:
    def __init__(self, ph, fw, posterior_path, hashes, verified=False, weights=None):
        self.ph, self.fw, self.hashes, self.verified = ph, fw, hashes, verified
        z = np.load(posterior_path, allow_pickle=False)
        self.meta = json.loads(str(z["__meta__"]))
        self.post, self.prior = {}, {}
        for key in z.files:
            if key == "__meta__":
                continue
            grp, mech, par = key.split("__", 2)
            tgt = self.post if grp == "post" else self.prior
            tgt.setdefault(mech, {})[par] = np.asarray(z[key], dtype=float)
        # every indicator's weights (density, magnetic and, from v4, load_timing)
        self.w = {k: dict(v) for k, v in self.meta["weights"].items()}
        # per-draw load-timing probabilities (v4): stored as post__<H>__ltp_<label>, kept apart from the parameters
        self.ltp = {}
        for m, v in self.post.items():
            if all(f"ltp_{k}" in v for k in LT_LEVELS):
                P = np.stack([v.pop(f"ltp_{k}") for k in LT_LEVELS], 1)
                self.ltp[m] = P / P.sum(1, keepdims=True)
        self.dw = weights or {m: np.ones(len(next(iter(v.values())))) for m, v in self.post.items()}
        self.preps = self.meta["density_preps"]
        self.conditioning = None

    def _stamp(self, res):
        res.update(tool_version=TOOL_VERSION, evidence_class=EVIDENCE_CLASS, artifact_sha256=self.hashes,
                   artifact_verified=self.verified)
        if self.conditioning is not None:
            res["conditioned_on"] = self.conditioning
        return res

    # ---------------------------------------------------------------- routes and hulls
    def _route(self, route):
        r = dict(ROUTE_DEFAULTS)
        r.update(self.meta.get("route_defaults", {}))
        r.update(route or {})
        if r["T_peak_C"] is None:
            raise ValueError("route needs T_peak_C")
        if r["pressure_mode"] not in PRESSURE_MODES:
            raise ValueError(f"pressure_mode must be one of {PRESSURE_MODES}")
        if r["anneal"] not in ANNEALS:
            raise ValueError(f"anneal must be one of {ANNEALS} (a 520 C step without the 900 C anneal is not modelled)")
        if r["current_path"] not in CURRENT_PATHS:
            raise ValueError(f"current_path must be one of {CURRENT_PATHS}")
        dc = r["degas_completeness"]
        if dc == "complete":
            pass
        elif _is_number(dc):
            if not 0.0 <= float(dc) <= 1.5:
                raise ValueError("degas_completeness must lie in [0, 1.5] (values above 1 are a declared extrapolation)")
            r["degas_completeness"] = float(dc)
        elif isinstance(dc, dict):
            kind = dc.get("distribution")
            if kind == "uniform":
                lo, hi = dc.get("low"), dc.get("high")
                if not (_is_number(lo) and _is_number(hi) and 0.0 <= float(lo) <= float(hi) <= 1.5):
                    raise ValueError("uniform degas_completeness needs numeric 0 <= low <= high <= 1.5")
            elif kind == "beta":
                a, b = dc.get("a"), dc.get("b")
                if not (_is_number(a) and _is_number(b) and float(a) > 0 and float(b) > 0):
                    raise ValueError("beta degas_completeness needs numeric a > 0 and b > 0")
            else:
                raise ValueError("degas_completeness dict needs distribution 'uniform' (low, high) or 'beta' (a, b)")
        else:
            raise ValueError("degas_completeness must be a number in [0, 1.5], a distribution dict "
                             "({'distribution': 'uniform', 'low', 'high'} or {'distribution': 'beta', 'a', 'b'}) "
                             "or 'complete'")
        fd = r[FEEDSTOCK_KEY]
        lo, hi = FEEDSTOCK_RANGE_UM
        if not (_is_number(fd) and lo <= float(fd) <= hi):
            raise ValueError(f"{FEEDSTOCK_KEY} must be a number in [{lo:g}, {hi:g}] (um; {FEEDSTOCK_REF_UM:g} = the "
                             "fitted feedstock; other values are a declared, uncalibrated extrapolation)")
        r[FEEDSTOCK_KEY] = float(fd)
        if r[FEEDSTOCK_KEY] != FEEDSTOCK_REF_UM and r["prep"] in self.preps:
            raise ValueError(f"fitted preparation {r['prep']!r} is the {FEEDSTOCK_REF_UM:g} um feedstock: a body made "
                             f"from other feedstock needs prep 'new' or another new label")
        return r

    @staticmethod
    def _feedstock_declared(r):
        return float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM)) != FEEDSTOCK_REF_UM

    def _feedstock_info(self, r):
        dec = self._feedstock_declared(r)
        return {FEEDSTOCK_KEY: float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM)), "declared": dec,
                "status": STATUS_DECLARED if dec else STATUS_FITTED,
                "note": ("declared feedstock particle size: grain size, an empirical grain-size factor on every "
                         "mechanism's coercivity, an oxidation penalty and a residual-hydrogen term follow uncalibrated "
                         "priors (README, 'Feedstock particle size'); no record constrains them") if dec else
                        "the fitted feedstock (about 4 um); the feedstock layer is inactive"}

    @staticmethod
    def _completeness_is_default(r):
        dc = r["degas_completeness"]
        return _is_number(dc) and float(dc) == 1.0

    @staticmethod
    def _sample_completeness(dc, size, rng):
        """Draws of the degassing completeness d (fraction of the fitted D/E state; > 1 = declared extrapolation).
        "complete" returns 1 (the generic degassing terms are held at the D/E state; hydrogen is zeroed separately)."""
        if dc == "complete":
            return np.ones(size)
        if _is_number(dc):
            return np.full(size, float(dc))
        if dc["distribution"] == "uniform":
            return rng.uniform(float(dc["low"]), float(dc["high"]), size)
        return rng.beta(float(dc["a"]), float(dc["b"]), size)

    def _expected_state(self, r):
        """Expected degassing state E[occurrence x min(d, 1)], used only for the hull distance."""
        dc = r["degas_completeness"]
        if dc == "complete":
            m = 1.0
        elif _is_number(dc):
            m = min(float(dc), 1.0)
        else:
            m = float(np.mean(np.minimum(self._sample_completeness(dc, 4000, np.random.default_rng(0)), 1.0)))
        return float(r["degas"]) * m

    def _degas_state(self, r, occ_uniform, rng_extra):
        """Per-draw degassing state (occurrence x completeness) and the extrapolation multiplier eps (only used where
        the state exceeds 1). occ_uniform: the uniforms v3 already drew for occurrence, so default inputs reproduce v3
        exactly; rng_extra is consulted only for non-default completeness."""
        occ = (occ_uniform < float(r["degas"])).astype(float)
        if self._completeness_is_default(r) or r["degas_completeness"] == "complete":
            return occ, None
        d = self._sample_completeness(r["degas_completeness"], occ_uniform.shape, rng_extra)
        eps = rng_extra.uniform(0.0, 1.0, occ_uniform.shape)
        return occ * d, eps

    def _arrays(self, r):
        R = self.ph.Route(T_peak_C=float(r["T_peak_C"]), heating_rate_C_per_min=float(r["heating_rate_C_per_min"]),
                          dwell_min=float(r["dwell_min"]), P_MPa=float(r["P_MPa"]), pressure_mode=r["pressure_mode"],
                          cooling_rate_C_per_min=float(r["cooling_rate_C_per_min"]),
                          load_ramp_min=float(r["load_ramp_min"]), degas=1.0, anneal=r["anneal"],
                          alignment_cos=float(r["alignment_cos"]), T_measure_C=float(r["T_measure_C"]),
                          current_path=r["current_path"])
        S = self.ph.stack([self.ph.schedule(R)])
        return dict(S=S, T_peak=np.array([R.T_peak_C]), rate=np.array([R.heating_rate_C_per_min]),
                    dwell=np.array([R.dwell_min]), P=np.array([R.P_MPa]),
                    cooling=np.array([1.0 if R.pressure_mode == "cooling" else 0.0]), degas=np.array([1.0]),
                    mode_dwell=np.array([1.0 if R.pressure_mode == "dwell" else 0.0]),
                    path=np.array([0.0 if R.current_path == "insulated" else 1.0]), hfac=np.array([1.0]),
                    a900=np.array([1.0 if R.anneal in ("900", "900+520") else 0.0]),
                    a520=np.array([1.0 if R.anneal == "900+520" else 0.0]), align=np.array([R.alignment_cos]),
                    Tmeas=np.array([R.T_measure_C]))

    def hull(self, r, rho=None):
        """Nearest-neighbour distance to the fitted routes in standardised process space; out_of_hull if the distance
        exceeds 1 or a feature leaves the fitted range by more than a quarter of its scale. The magnetic distance is the
        one H0's discrepancy uses (forward.mag_hull_distance)."""
        out = {}
        for module in ("density", "magnetic"):
            H = self.meta["hull"][module]
            feats = H["features"]
            x = self._features(r, rho, feats)
            pts = np.asarray(H["points"], float)
            sc = np.asarray(H["scales"], float)
            d = float(np.sqrt((((pts - x) / sc) ** 2).sum(1)).min())
            lo, hi = pts.min(0), pts.max(0)
            outside = [f for f, xi, l, h, s in zip(feats, x, lo, hi, sc) if xi < l - 0.25 * s or xi > h + 0.25 * s]
            if r["pressure_mode"] == "dwell":  # declared scenario: no fitted record has load from the dwell start
                outside.append("pressure_mode=dwell")
            if module == "magnetic" and r.get("current_path") == "insulated":
                outside.append("current_path=insulated")
            if module == "magnetic" and r.get("degas_completeness") == "complete":
                outside.append("degas_completeness=complete")
            if self._feedstock_declared(r):  # C4: no record was made from other feedstock
                outside.append(f"{FEEDSTOCK_KEY}={float(r[FEEDSTOCK_KEY]):g}")
            out[module] = dict(distance=d, out_of_hull=bool(d > 1.0 or outside), outside_features=outside)
        return out

    @staticmethod
    def _features(r, rho, feats):
        val = dict(T_peak_C=float(r["T_peak_C"]), log10_rate=math.log10(float(r["heating_rate_C_per_min"])),
                   dwell_min=float(r["dwell_min"]), P_MPa=float(r["P_MPa"]),
                   load_on_cooling=_load_on_cooling(r),
                   degas=float(r.get("_degas_expected", r["degas"])), rel_density=float(rho if rho is not None else 0.98))
        return np.array([val[f] for f in feats], float)

    def _hull_distance(self, r, rho):
        d = self.fw.mag_hull_distance(float(r["T_peak_C"]), float(rho), _load_on_cooling(r),
                                      float(r.get("_degas_expected", r["degas"])),
                                      float(r["heating_rate_C_per_min"]))
        if self._feedstock_declared(r):
            # C4 (declared): every record used the 4 um feedstock, so a halving or doubling of the feedstock size adds
            # one standardised unit (as 100 C does) to H0's nearest-neighbour distance
            d = math.sqrt(d * d + (math.log(float(r[FEEDSTOCK_KEY]) / FEEDSTOCK_REF_UM) / math.log(2.0)) ** 2)
        return d

    # ---------------------------------------------------------------- draw assignment
    def _mechs(self, kind, group="post"):
        src = self.post if group == "post" else self.prior
        return [m for m in self.w[kind] if m in src]

    def _assign_bma(self, kind, n, rng, group="post"):
        """n draws from the model average: mechanism labels by weight, parameter indices by the per-draw weights."""
        src = self.post if group == "post" else self.prior
        mechs = self._mechs(kind, group)
        wm = np.array([self.w[kind][m] for m in mechs], float) if group == "post" else np.ones(len(mechs))
        counts = rng.multinomial(n, wm / wm.sum())
        labels, idx = [], []
        for m, c in zip(mechs, counts):
            if c == 0:
                continue
            K = len(next(iter(src[m].values())))
            p = self.dw[m] / self.dw[m].sum() if group == "post" else np.full(K, 1.0 / K)
            labels += [m] * c
            idx.append(rng.choice(K, size=c, p=p))
        labels, idx = np.array(labels), np.concatenate(idx)
        perm = rng.permutation(n)
        return labels[perm], idx[perm]

    def _assign_strata(self, kind, n_per, rng):
        """n_per draws from every mechanism with non-zero weight (stratified inner sets for eig)."""
        labels, idx = [], []
        for m in self._mechs(kind):
            if self.w[kind][m] <= 0:
                continue
            K = len(next(iter(self.post[m].values())))
            p = self.dw[m] / self.dw[m].sum()
            labels += [m] * n_per
            idx.append(rng.choice(K, size=n_per, p=p))
        return np.array(labels), np.concatenate(idx)

    def _params(self, src, m, idx):
        return {k: v[idx][:, None] for k, v in src[m].items()}

    # ---------------------------------------------------------------- conditional indicators (v4: load timing)
    def _indicators(self):
        """Indicators whose labels are drawn conditionally on the magnetic draw (no parameter arrays of their own)."""
        return ["load_timing"] if self.ltp else []

    def _assign_indicator(self, name, mlab, midx, rng, group="post"):
        """Per-draw labels of a conditional indicator, aligned with (mlab, midx): load_timing labels are drawn from
        the draw's own posterior probabilities p(LT | theta, data) (posterior group) or uniformly (prior group)."""
        if name != "load_timing":
            raise ValueError(f"unknown conditional indicator {name!r}")
        n = len(mlab)
        u = rng.random(n)
        out = np.empty(n, dtype=object)
        for m in np.unique(mlab):
            pos = np.where(mlab == m)[0]
            if group == "post" and m in self.ltp:
                P = self.ltp[m][midx[pos]]
            else:
                P = np.full((len(pos), len(LT_LEVELS)), 1.0 / len(LT_LEVELS))
            c = np.cumsum(P, 1)
            k = np.minimum((u[pos][:, None] > c).sum(1), len(LT_LEVELS) - 1)
            out[pos] = np.array(LT_LEVELS, dtype=object)[k]
        return out.astype(str)

    def _assign_strata_indicator(self, name, n_per, rng):
        """n_per joint draws (mlab, midx, labels) per level of a conditional indicator: (m, i) sampled with
        probability proportional to w_m * dw_m[i] / sum(dw_m) * p_m[i, level] (exactly p(m, i | level))."""
        if name != "load_timing":
            raise ValueError(f"unknown conditional indicator {name!r}")
        mechs = [m for m in self._mechs("magnetic") if self.w["magnetic"][m] > 0]
        blocks = []
        for m in mechs:
            base = self.w["magnetic"][m] * self.dw[m] / self.dw[m].sum()
            blocks.append(base[:, None] * self.ltp[m])
        Wt = np.concatenate(blocks, 0)
        offs = np.cumsum([0] + [len(b) for b in blocks])
        labs, idxs, lts = [], [], []
        for k, lvl in enumerate(LT_LEVELS):
            pk = Wt[:, k] / Wt[:, k].sum()
            j = rng.choice(len(pk), size=n_per, p=pk)
            mi = np.searchsorted(offs, j, side="right") - 1
            labs.append(np.array(mechs, dtype=object)[mi])
            idxs.append(j - offs[mi])
            lts.append(np.full(n_per, lvl, dtype=object))
        return (np.concatenate(labs).astype(str), np.concatenate(idxs).astype(int),
                np.concatenate(lts).astype(str))

    # ---------------------------------------------------------------- per-draw latent quantities at a route
    def _density_mean(self, A, r, dlab, didx, prep_z, group="post"):
        """Per-draw route-level relative density (before specimen scatter) and its scatter scale."""
        src = self.post if group == "post" else self.prior
        n = len(dlab)
        mu, sig = np.empty(n), np.empty(n)
        for m in np.unique(dlab):
            pos = np.where(dlab == m)[0]
            p = self._params(src, m, didx[pos])
            if r["prep"] in self.preps and group == "post":
                off = p["tau_prep"] * p[f"z_prep_{self.preps.index(r['prep'])}"]
            else:
                off = p["tau_prep"] * prep_z[pos]
            mu[pos] = self.fw.density(p, m, A, lnk_offset=off, xp=np)[:, 0]
            sig[pos] = p["sig_rho"][:, 0]
        return mu, sig

    def _magnetic_latent(self, A, r, mlab, midx, rho, degas_ind, z_hull, u_spec=None, group="post", eps=None,
                         indicators=None):
        """Per-draw latent magnetic quantities. rho, degas_ind, z_hull, u_spec: (n,) arrays aligned with the draws.
        ln Hci includes u_spec (the body's specimen effect) when given; otherwise it is the route-level value.
        degas_ind is the continuous degassing state (occurrence x completeness). Where it exceeds 1 (a declared
        extrapolation beyond the fitted D/E state) the forward model is evaluated at 1 and ln Hci gains
        eps * (state - 1) * [ln Hci(state 1) - ln Hci(state 0)] for the same draw, eps ~ U(0, 1); the residual-hydrogen
        index is reduced by the factor 1 - eps * (state - 1). With degas_completeness "complete" a degassed draw has
        state 1 for the generic degassing terms and no hydrogen (H1 and the trap alternative see h0 = 0).
        indicators: {"load_timing": labels (n,)} aligned with the draws (required when the posterior carries LT; the
        same labels must be used for every route of a joint draw). lh_res is the physical residual-hydrogen index of
        the sintered body (pore closure by the load mode), floored at the declared detection limit H_DETECTION_INDEX."""
        src = self.post if group == "post" else self.prior
        n = len(mlab)
        if self.ltp:
            if not indicators or "load_timing" not in indicators:
                raise ValueError("this posterior carries the load-timing indicator: pass "
                                 "indicators={'load_timing': labels} from _assign_indicator")
            ltlab = np.asarray(indicators["load_timing"]).astype(str)
            if ltlab.shape != (n,):
                raise ValueError("load_timing labels must be aligned with the draws")
        else:
            ltlab = np.full(n, "load")
        complete = r.get("degas_completeness") == "complete"
        fd = float(r.get(FEEDSTOCK_KEY, FEEDSTOCK_REF_UM))
        milled = fd != FEEDSTOCK_REF_UM  # C4 feedstock layer; inactive (code path unchanged) at the fitted feedstock
        keys = ("lh", "lbr", "sq", "lgrain", "lh_res", "sig_spec", "sig_piece", "sig_Br", "sig_sq", "Js", "HA", "nB",
                "dB", "HR", "kR", "grain")
        out = {k: np.empty(n) for k in keys}
        for m in np.unique(mlab):
            pos = np.where(mlab == m)[0]
            p = self._params(src, m, midx[pos])
            rr = rho[pos][:, None]
            state = degas_ind[pos][:, None]
            over = np.maximum(state - 1.0, 0.0)
            Ad = dict(A)
            Ad["degas"] = np.minimum(state, 1.0)
            Ad["hfac"] = np.where(state >= 1.0, 0.0, 1.0) if complete else np.ones_like(state)
            if milled:
                miss = [k for k in FEEDSTOCK_PARAMS if k not in p]
                if miss:
                    raise ValueError(f"this posterior carries no feedstock-layer draws {miss}: {FEEDSTOCK_KEY} other "
                                     f"than {FEEDSTOCK_REF_UM:g} needs a package that declares the layer")
                # residual hydrogen of the milled powder, as the current-feedstock degassing state with the same index
                # (no hydrogen, hence no change, where degassing is complete: hfac = 0)
                Ad["degas"] = np.where(Ad["hfac"] > 0, self.fw.feedstock_degas_equivalent(p, Ad["degas"], fd, np),
                                       Ad["degas"])
                # sintered grain from the declared feedstock vs the same route and draw from the 4 um feedstock; one
                # empirical grain-size factor for every mechanism. Ad carries no grain hook, so H4's Kronmueller term
                # stays at g_ref and the feedstock change is counted once.
                g_fd = self.fw.grain(p, dict(A, fd_d0_factor=self.fw.feedstock_d0_factor(p, fd, np)), xp=np)
                g_ref = self.fw.grain(p, A, xp=np)
                fshift = (np.log(self.fw.feedstock_grain_factor(p, g_fd, g_ref, np))
                          - self.fw.feedstock_oxidation_penalty(p, fd, np))
            hd = np.array([self._hull_distance(r, x) for x in rho[pos]])[:, None] if m == "H0" else 0.0
            u = u_spec[pos][:, None] if u_spec is not None else 0.0
            zh = z_hull[pos][:, None]
            lt = ltlab[pos]
            if np.any(over > 0):
                e = eps[pos][:, None]
                raw = self.fw._log_hci_raw(p, m, Ad, rr, u, hd, zh, np, lt)
                one = dict(Ad, degas=np.ones_like(state), hfac=np.ones_like(state))
                zero = dict(Ad, degas=np.zeros_like(state), hfac=np.ones_like(state))
                slope = (self.fw._log_hci_raw(p, m, one, rr, 0.0, hd, zh, np, lt)
                         - self.fw._log_hci_raw(p, m, zero, rr, 0.0, hd, zh, np, lt))
                tot = raw + e * over * slope
                lh = self.fw.cap_log_hci(tot + fshift if milled else tot, np)
            else:
                raw = self.fw._log_hci_raw(p, m, Ad, rr, u, hd, zh, np, lt)
                lh = self.fw.cap_log_hci(raw + fshift if milled else raw, np)
            out["lh"][pos] = lh[:, 0]
            out["lbr"][pos] = np.log(self.fw.remanence(p, A, rr, np.exp(lh), xp=np))[:, 0]
            out["sq"][pos] = (self.fw.squareness_logit(p, A, xp=np) * np.ones((len(pos), 1)))[:, 0]
            g = (g_fd if milled else self.fw.grain(p, A, xp=np))[:, 0]
            out["grain"][pos], out["lgrain"][pos] = g, np.log(g)
            hres = self.ph.desorption(A["S"], A["cooling"], Ad["degas"], p, mode_dwell=A.get("mode_dwell")) * Ad["hfac"]
            if np.any(over > 0):
                hres = hres * np.maximum(1.0 - eps[pos][:, None] * over, 1e-6)
            out["lh_res"][pos] = np.log(np.maximum(hres, H_DETECTION_INDEX))[:, 0]
            for k in ("sig_spec", "sig_piece", "sig_Br", "sig_sq", "Js", "HR", "kR"):
                out[k][pos] = p[k][:, 0]
            for k, default in (("HA", 75.0), ("nB", 0.17), ("dB", 0.004)):
                out[k][pos] = p[k][:, 0] if k in p else default
        return out

    # ---------------------------------------------------------------- forward simulation for predict / p_meets_spec
    def _temperature(self, lat, hci, br, T_C, rng):
        """Declared temperature scenario (not calibrated): Kuz'min Js(T), linear HA(T), Neff = nB ln(D/dB), alpha from
        the room-temperature Hci; remanence retention R(Hci) re-evaluated at temperature."""
        if abs(T_C - 20.0) < 1e-9:
            return hci, br
        n = len(hci)
        s = rng.uniform(0.5, 0.6, n)
        cHA = rng.uniform(0.0026, 0.0032, n)
        raw, rJ = self._temperature_map(lat, hci, T_C, s, cHA)
        hci_T = np.maximum(raw, 0.0)
        R = lambda h: 1.0 / (1.0 + (lat["HR"] / np.maximum(h, 1e-3)) ** lat["kR"])
        return hci_T, br * rJ * R(hci_T) / R(hci)

    def _temperature_map(self, lat, hci, T_C, s, cHA):
        """Deterministic part of the temperature scenario: unfloored Hci(T) for a body whose room-temperature Hci is
        `hci` (alpha is derived from it, so the ratio Hci(T)/Hci(20 C) depends on the body's own Hci), and the Js ratio."""
        rJ = self.ph.kuzmin_js(T_C + 273.15, 1.0, s=s) / self.ph.kuzmin_js(293.15, 1.0, s=s)
        rH = np.clip(1.0 - cHA * (T_C - 20.0), 0.05, None)
        neff = lat["nB"] * np.log(np.maximum(lat["grain"], 0.01) / lat["dB"])
        alpha = (hci + neff * lat["Js"]) / lat["HA"]
        return alpha * lat["HA"] * rH - neff * lat["Js"] * rJ, rJ

    def _bh(self, br, hci, sq, rng):
        cal = self.meta["bhmax_shape"]
        f = cal["shape_factor"] * np.exp(cal["log_sd"] * rng.standard_normal(len(br)))
        raw = np.asarray(self.ph.bhmax(np.asarray(br), np.maximum(np.asarray(hci), 1e-3), np.asarray(sq), shape=1.0))
        return np.minimum(raw * f, np.asarray(br) ** 2 / 4.0)  # the Br^2/4 bound holds after the shape factor

    @staticmethod
    def _rho_scenario(rs):
        """Validate a declared density scenario {"mean": .., "sd": ..} (g/cm3 if mean > 1.5, else relative)."""
        if not (isinstance(rs, dict) and _is_number(rs.get("mean")) and _is_number(rs.get("sd"))
                and float(rs["mean"]) > 0 and float(rs["sd"]) >= 0):
            raise ValueError("rho_scenario must be {'mean': number > 0, 'sd': number >= 0} (g/cm3 or relative)")
        scale = RHO_TD if float(rs["mean"]) > 1.5 else 1.0
        return float(rs["mean"]) / scale, float(rs["sd"]) / scale

    def _simulate(self, route, n, seed, group="post", rho_observed=None, rho_scenario=None):
        r = self._route(route)
        r["_degas_expected"] = self._expected_state(r)
        rng = np.random.default_rng(seed)
        A = self._arrays(r)
        hull = self.hull(r)
        if rho_observed is not None and rho_scenario is not None:
            raise ValueError("pass either rho_observed (a measured body) or rho_scenario (an assumed density), not both")
        if rho_observed is not None:
            rho_obs = float(rho_observed)
            rho_obs = rho_obs / RHO_TD if rho_obs > 1.5 else rho_obs
            rho = np.full(n, rho_obs) + 0.005 * rng.standard_normal(n)
            dlab = np.array(["observed"] * n)
        elif rho_scenario is not None:
            mu, sd = self._rho_scenario(rho_scenario)
            rho = np.clip(mu + sd * rng.standard_normal(n), 0.3, 1.02)
            dlab = np.array(["scenario"] * n)
        else:
            dlab, didx = self._assign_bma("density", n, rng, group)
            mu, sig = self._density_mean(A, r, dlab, didx, rng.standard_normal((n, 1)), group)
            rho = np.clip(mu + sig * rng.standard_t(5.0, n), 0.3, 1.02)
        mlab, midx = self._assign_bma("magnetic", n, rng, group)
        ind = {k: self._assign_indicator(k, mlab, midx, np.random.default_rng([int(seed), 47]), group)
               for k in self._indicators()}
        degas_ind, eps = self._degas_state(r, rng.random(n), np.random.default_rng([int(seed), 31]))
        z_hull = rng.standard_normal(n)
        sig_spec = np.array([self.post[m]["sig_spec"][i] if group == "post" else self.prior[m]["sig_spec"][i]
                             for m, i in zip(mlab, midx)])
        u = sig_spec * rng.standard_normal(n)  # a new physical body: its own specimen effect (before the Hci cap)
        lat = self._magnetic_latent(A, r, mlab, midx, rho, degas_ind, z_hull, u, group, eps=eps, indicators=ind)
        hci_meas = np.exp(lat["lh"]) * np.exp(lat["sig_piece"] * rng.standard_normal(n))
        br = np.exp(lat["lbr"] + lat["sig_Br"] * rng.standard_normal(n))
        sq = 1.0 / (1.0 + np.exp(-(lat["sq"] + lat["sig_sq"] * rng.standard_normal(n))))
        hci_T, br_T = self._temperature(lat, hci_meas, br, float(r["T_measure_C"]), rng)
        bh = self._bh(br_T, hci_T, sq, rng)
        sims = dict(density_rel=rho, grain_um=lat["grain"], Hci_kOe=hci_T, Br_kGs=br_T, Hk_over_Hci=sq, BHmax_MGOe=bh)
        hull_m = self.hull(r, rho=float(np.median(rho)))["magnetic"]
        r.pop("_degas_expected", None)
        return r, sims, dlab, mlab, dict(density=hull["density"], magnetic=hull_m), ind

    # ---------------------------------------------------------------- public API
    def predict(self, route, n_draws=4000, seed=0, rho_observed=None, include_draws=False, rho_scenario=None):
        """Posterior predictive of a *measured* as-processed body (specimen-to-specimen and measurement scatter
        included). rho_observed (g/cm3 or relative): condition the magnetics on a measured density (an existing body).
        rho_scenario {"mean", "sd"}: an assumed density with its uncertainty, for a body not yet made."""
        r, sims, dlab, mlab, hull, ind = self._simulate(route, n_draws, seed, "post", rho_observed, rho_scenario)
        _, prs, _, _, _, _ = self._simulate(route, max(1000, n_draws // 2), seed + 1, "prior", rho_observed,
                                            rho_scenario)
        res = {"route": r, "mechanism_weights": self.w, "out_of_hull": hull, "outputs": {}}
        for k in OUTPUTS:
            x = np.asarray(sims[k], float)
            logk = k not in ("density_rel", "Hk_over_Hci")
            tr = (lambda v: np.log(np.maximum(v, 1e-6))) if logk else (lambda v: v)
            lab = dlab if k == "density_rel" else mlab
            groups = {m: x[lab == m] for m in np.unique(lab)}
            tx, tp = tr(x), tr(np.asarray(prs[k], float))
            v_post = float(np.var(tx))
            between = float(np.var(np.concatenate([np.full(len(g), np.mean(tr(g))) for g in groups.values()]))) \
                if len(groups) > 1 else 0.0
            share = between / v_post if v_post > 0 else 0.0
            ratio = (_iqr(tx) / _iqr(tp)) ** 2 if _iqr(tp) > 0 else 1.0
            module = "density" if k == "density_rel" else "magnetic"
            observed = k == "density_rel" and rho_observed is not None
            extra = {}
            if module == "magnetic" and "load_timing" in ind:
                lt = ind["load_timing"]
                extra["by_load_timing"] = {L: dict(median=float(np.median(x[lt == L])),
                                                   q05=float(np.quantile(x[lt == L], 0.05)),
                                                   q95=float(np.quantile(x[lt == L], 0.95)), n=int((lt == L).sum()))
                                           for L in LT_LEVELS if (lt == L).any()}
            res["outputs"][k] = dict(
                mean=float(np.mean(x)), median=float(np.median(x)), sd=float(np.std(x)),
                q05=float(np.quantile(x, 0.05)), q25=float(np.quantile(x, 0.25)), q75=float(np.quantile(x, 0.75)),
                q95=float(np.quantile(x, 0.95)),
                by_mechanism={m: dict(median=float(np.median(g)), q05=float(np.quantile(g, 0.05)),
                                      q95=float(np.quantile(g, 0.95)), n=int(len(g))) for m, g in groups.items()},
                posterior_to_prior_spread_ratio=ratio, between_mechanism_share=share,
                prior_dominated=bool(False if observed else (ratio > 0.5 or hull[module]["out_of_hull"])),
                mechanism_dominated=bool(share > 0.5), conditioned_on_observed_density=observed,
                density_scenario_declared=bool(rho_scenario is not None), **extra, **_distribution(x),
                status=self._feedstock_info(r)["status"])
        d = res["outputs"]["density_rel"]
        res["outputs"]["density_g_cm3"] = {k: d[k] * RHO_TD for k in ("mean", "median", "sd", "q05", "q25", "q75", "q95")}
        res["outputs"]["density_g_cm3"]["quantiles"] = {k: v * RHO_TD for k, v in d["quantiles"].items()}
        js = 16.1
        br_iso = js * d["median"] * float(r["alignment_cos"])
        res["ceilings"] = {"Br_ideal_kGs_Js_rho_cos": br_iso,
                           "BHmax_bound_MGOe_Br2_over_4": br_iso ** 2 / 4.0,
                           "isotropic_full_density_BHmax_MGOe": (js * 0.5) ** 2 / 4.0,
                           "Br_needed_for_45_MGOe_kGs": math.sqrt(4 * 45.0),
                           "note": "ideal values use Js = 16.1 kG; RPI's E sits ~2 % above (kappa); isotropic "
                                   "<cos>=0.5 caps Br near Js/2; 45 MGOe needs Br >= 13.4 kG, i.e. alignment"}
        res["scenarios"] = _scenarios(r)
        res["feedstock"] = self._feedstock_info(r)
        res["flag_rules"] = ("prior_dominated: posterior-to-prior predictive spread ratio (IQR^2 on the transformed "
                             "scale) > 0.5, or the output's module is out of its data hull; mechanism_dominated: > 50 % "
                             "of the predictive variance lies between mechanism-conditional means")
        res["objective"] = OBJECTIVE
        if rho_observed is not None:
            res["rho_observed_note"] = ("magnetics conditioned on rho_observed as a MEASURED density (fixed 0.005 relative "
                                        "spread); for a body not yet made pass rho_scenario={'mean', 'sd'} instead")
        res["recording_note"] = ("record a forecast with its quantiles, not mean +/- sd: the magnetic predictives are "
                                 "skewed mixtures; mu_ln / sd_ln summarise them only where lognormal_adequate is true; "
                                 "forecast_record() returns a ready-to-store record")
        if include_draws:
            res["draws"] = {k: np.asarray(v).tolist() for k, v in sims.items()}
        return self._stamp(res)

    def p_meets_spec(self, route, spec, n_draws=6000, seed=0, rho_observed=None, rho_scenario=None):
        """Posterior probability that one body made by `route` meets `spec`, scored on *measured* values (the same
        measurement scatter as the RPI curves). spec keys (all optional): Hci_kOe_min, BHmax_MGOe_min, Br_kGs_min,
        density_rel_min, T_operating_C (evaluates Hci/Br/BHmax at that temperature as a declared scenario)."""
        _validate_spec(spec)
        route = dict(route)
        if "T_operating_C" in spec:
            route["T_measure_C"] = float(spec["T_operating_C"])
        r, sims, dlab, mlab, hull, ind = self._simulate(route, n_draws, seed, "post", rho_observed, rho_scenario)
        keys = (("Hci_kOe_min", "Hci_kOe"), ("BHmax_MGOe_min", "BHmax_MGOe"), ("Br_kGs_min", "Br_kGs"),
                ("density_rel_min", "density_rel"))
        ok = np.ones(n_draws, bool)
        checks = {}
        for key, out in keys:
            if key in spec:
                c = np.asarray(sims[out]) >= float(spec[key])
                checks[key] = float(c.mean())
                ok &= c
        p = float(ok.mean())
        by_m = {m: float(ok[mlab == m].mean()) for m in np.unique(mlab)}
        _, prs, _, _, _, _ = self._simulate(route, max(2000, n_draws // 2), seed + 1, "prior", rho_observed,
                                            rho_scenario)
        okp = np.ones(len(prs["Hci_kOe"]), bool)
        for key, out in keys:
            if key in spec:
                okp &= np.asarray(prs[out]) >= float(spec[key])
        mag_spec = any(k in spec for k in ("Hci_kOe_min", "BHmax_MGOe_min", "Br_kGs_min"))
        spread = max(by_m.values()) - min(by_m.values()) if by_m else 0.0
        res = {"route": r, "spec": spec, "p_meets_spec": p,
               "monte_carlo_se": math.sqrt(max(p * (1 - p), 1e-12) / n_draws), "marginals": checks,
               "by_magnetic_mechanism": by_m, "mechanism_weights": self.w["magnetic"],
               "by_load_timing": ({L: float(ok[ind["load_timing"] == L].mean()) for L in LT_LEVELS
                                   if (ind["load_timing"] == L).any()} if "load_timing" in ind else None),
               "load_timing_weights": self.w.get("load_timing"),
               "prior_predictive_p": float(okp.mean()),
               "prior_dominated": bool((mag_spec and hull["magnetic"]["out_of_hull"])
                                       or (hull["density"]["out_of_hull"] and "density_rel_min" in spec)
                                       or (abs(p - okp.mean()) < 0.02 and 0.0 < p < 1.0) or spread > 0.3),
               "no_draw_meets_spec": bool(p == 0.0 and float(okp.mean()) == 0.0),
               "p_upper_95": (3.0 / n_draws) if p == 0.0 else None,
               "out_of_hull": hull, "scenario_not_calibrated": not _scenarios(r)["calibrated"],
               "feedstock": self._feedstock_info(r),
               "scenarios": _scenarios(r),
               "scored_on": "measured values (piece, Br, squareness and curve-shape scatter included)"}
        return self._stamp(res)

    def forecast_record(self, route, body_id=None, rho_observed=None, n_draws=8000, seed=0,
                        outputs=("density_rel", "Hci_kOe", "Br_kGs", "BHmax_MGOe", "Hk_over_Hci"), rho_scenario=None):
        """A schema-ready committed-forecast record for one body (deterministic for a given seed, so it can be stored
        and reproduced). body_id: a neutral caller-supplied identifier. Contains, per output, the quantile set and
        log-normal parameters with the flags; the resolved route and scenario keys; the conditioning; and the tool
        and artifact hashes. The scoring harness reads this record directly."""
        res = self.predict(route, n_draws=n_draws, seed=seed, rho_observed=rho_observed, rho_scenario=rho_scenario)
        rec_out = {}
        for k in outputs:
            o = res["outputs"][k]
            rec_out[k] = {"quantiles": o["quantiles"], "mu_ln": o["mu_ln"], "sd_ln": o["sd_ln"],
                          "lognormal_adequate": o["lognormal_adequate"],
                          "lognormal_max_log_error": o["lognormal_max_log_error"], "median": o["median"],
                          "frac_nonpositive": o["frac_nonpositive"], "prior_dominated": o["prior_dominated"],
                          "mechanism_dominated": o["mechanism_dominated"],
                          "posterior_to_prior_spread_ratio": o["posterior_to_prior_spread_ratio"],
                          "conditioned_on_observed_density": o["conditioned_on_observed_density"],
                          "status": o["status"]}
        return {"schema": "ndfeb-forecast-record-v1", "body_id": body_id, "route": res["route"],
                "status": res["feedstock"]["status"], "feedstock": res["feedstock"],
                "scenarios": res["scenarios"], "conditioning": {"rho_observed": rho_observed,
                                                                "rho_scenario": rho_scenario},
                "outputs": rec_out, "out_of_hull": res["out_of_hull"], "mechanism_weights": res["mechanism_weights"],
                "n_draws": int(n_draws), "seed": int(seed), "quantile_levels": list(QUANTILE_LEVELS),
                "units": {"density_rel": "fraction of 7.62 g/cm3", "Hci_kOe": "kOe", "Br_kGs": "kG",
                          "BHmax_MGOe": "MGOe", "Hk_over_Hci": "ratio"},
                "scoring": "score from the quantiles (interpolated predictive on the log scale for positive quantities); "
                           "mu_ln / sd_ln only where lognormal_adequate; 50/90 % interval coverage",
                "tool_version": TOOL_VERSION, "evidence_class": EVIDENCE_CLASS, "artifact_sha256": self.hashes,
                "artifact_verified": self.verified, "conditioned_on": self.conditioning}

    def _latents(self, meas, target, dlab, didx, mlab, midx, rng, indicators=None):
        """Per-route latent means and noise scales for a set of joint draws (aligned arrays of length n). Shared across
        routes within a draw: new-preparation effects, the degas state of each preparation, H0's discrepancy direction."""
        n = len(mlab)
        routes = [mm["route"] for mm in meas] + [target]
        rho_obs = [mm.get("rho_observed") for mm in meas] + [None]
        prep_z, degas, eps, prep_dc, prep_fd = {}, {}, {}, {}, {}
        z_hull = rng.standard_normal(n)
        rows = []
        extra = None  # child generator, created only when a new v3.1 option needs randomness (keeps v3 equivalence)
        for j, route in enumerate(routes):
            r = self._route(route)
            r["_degas_expected"] = self._expected_state(r)
            A = self._arrays(r)
            if r["prep"] in prep_dc and prep_dc[r["prep"]] != json.dumps(r["degas_completeness"], sort_keys=True):
                raise ValueError(f"routes sharing preparation {r['prep']!r} must share degas_completeness")
            prep_dc[r["prep"]] = json.dumps(r["degas_completeness"], sort_keys=True)
            _check_prep_feedstock(prep_fd, r)
            if r["prep"] not in prep_z:
                prep_z[r["prep"]] = rng.standard_normal((n, 1))
                u = rng.random(n)
                if not self._completeness_is_default(r) and extra is None:
                    extra = rng.spawn(1)[0]
                degas[r["prep"]], eps[r["prep"]] = self._degas_state(r, u, extra)
            mu, sig = self._density_mean(A, r, dlab, didx, prep_z[r["prep"]])
            ro = rho_obs[j]
            rho = np.full(n, (ro / RHO_TD if ro > 1.5 else ro)) if ro else np.clip(mu, 0.3, 1.0)
            lat = self._magnetic_latent(A, r, mlab, midx, rho, degas[r["prep"]], z_hull, eps=eps[r["prep"]],
                                        indicators=indicators)
            row = dict(mu_rho=mu, sig_rho=sig, rho_known=ro is not None, **lat)
            if j < len(meas) and "Hci_ratio_T" in meas[j].get("quantities", []):
                if extra is None:
                    extra = rng.spawn(1)[0]
                row["lratioT"], row["lratioT_g"] = self._ln_hci_ratio_T(lat, float(meas[j].get("T2_C", 120.0)), extra)
            rows.append(row)
        return rows

    def _ln_hci_ratio_T(self, lat, T2_C, rng):
        """Declared measurement hook (uncalibrated): ln[Hci(T2) / Hci(20 C)] at the route-level latent under the tool's
        temperature scenario (the same map as predict), with per-draw Kuz'min s and HA(T) slope; floored at ln(0.01)
        (a loop whose coercivity collapses is field-limited, not exact). Returns the ratio and its sensitivity
        g = d ln(ratio) / d ln(Hci): the specimen effect does not cancel, because alpha follows the body's own Hci, so
        the eig noise adds (g * sig_spec)^2 to 2 * sig_piece^2."""
        n = len(lat["lh"])
        s = rng.uniform(0.5, 0.6, n)
        cHA = rng.uniform(0.0026, 0.0032, n)

        def ln_ratio(lh):
            hci = np.exp(lh)
            raw, _ = self._temperature_map(lat, hci, T2_C, s, cHA)
            return np.log(np.maximum(raw, 0.01 * hci) / hci)
        r0 = ln_ratio(lat["lh"])
        g = (ln_ratio(lat["lh"] + 0.01) - r0) / 0.01
        return r0, g

    # ---------------------------------------------------------------- experiment design
    def _eig_v31(self, candidates, n_outer=1500, n_inner_per_mechanism=6000, seed=0, target=None):
        """The v3.1-r3 estimator, unchanged (eig(..., method="nmc-v3.1")). Rank candidate measurements by expected
        information gain (nats).
        candidates: list of {"name": str, "measurements": [ {"route": {...}, "rho_observed": optional,
                    "quantities": ["Hci","Br","sq","density","grain","H_ppm","Hci_ratio_T"], "replicates": 1,
                    "T2_C": 120 (only for Hci_ratio_T)} ]}
        H_ppm and Hci_ratio_T use declared, uncalibrated measurement models (H_MEASUREMENT_LOG_SD; the temperature
        scenario), so their EIG is conditional on those assumptions.
        Mechanism EIG is estimated two ways from the same draws: nested Monte Carlo with the outer draw excluded
        (consistent; biased low when the inner set is small) and with the outer draw included in its own stratum
        (biased high for informative designs). The reported value is the midpoint of that bracket; `*_converged` is
        False when the bracket is wider than max(0.05 nats, 2 SE): then increase n_inner_per_mechanism. The
        all-parameter EIG is a prior-contrastive-style estimate (outer draw included; cf. Foster et al. 2019). The
        density-mechanism EIG counts the density measurements only (0 when no density is measured; a lower bound when
        magnetics are measured on a body of unknown density, which also carries density information). The midpoint
        uses max(low, 0), since a negative lower estimate is small-sample bias. p_identify is bracketed the same way;
        expected posterior weights come from the outer-draw-included estimator. Also: the
        expected posterior mechanism weights, the probability that the MAP mechanism is the generating one, and the
        expected change of the predictive sd of ln Hci at `target` (a route; default the first measurement's)."""
        out = []
        for ci, cand in enumerate(candidates):
            out.append(self._eig_one_v31(cand, n_outer, n_inner_per_mechanism,
                                         np.random.default_rng(seed + 7919 * (ci + 1)), target))
        for d in out:
            d["ln_Hci_sd_reduction_target"] = d["ln_Hci_sd_target_now"] - d["ln_Hci_sd_target_expected_after"]
        orderings = {key: [d["name"] for d in sorted(out, key=lambda d: -d[key])]
                     for key in ("eig_magnetic_mechanism_nats", "eig_density_mechanism_nats", "eig_all_parameters_nats",
                                 "ln_Hci_sd_reduction_target")}
        out.sort(key=lambda d: (-d["eig_magnetic_mechanism_nats"], -d["eig_all_parameters_nats"]))
        for i, d in enumerate(out):
            d["rank"] = i + 1
        return self._stamp({"ranking": out, "ranking_rule": "primary: EIG on the magnetic mechanism indicator "
                            "(bracket midpoint); tie-break: EIG on all parameters; see orderings for every metric",
                            "orderings": orderings,
                            "all_converged": bool(all(d["eig_magnetic_mechanism_converged"]
                                                      and d["eig_density_mechanism_converged"] for d in out)),
                            "method": "nested Monte Carlo, stratified inner sets per magnetic mechanism; mechanism "
                                      "EIG bracketed by excluding / including the outer draw in its stratum; Gaussian "
                                      "measurement models in log space (Student-t(5) for density); per-body specimen "
                                      "effects integrated; H0 discrepancy direction, new-preparation effects and "
                                      "degas state shared across routes within a draw"})

    def _eig_one_v31(self, cand, n_outer, n_per, rng, target, chunk=6000):
        meas = cand["measurements"]
        tgt = target or meas[0]["route"]
        dl_o, di_o = self._assign_bma("density", n_outer, rng)
        ml_o, mi_o = self._assign_bma("magnetic", n_outer, rng)
        ml_i, mi_i = self._assign_strata("magnetic", n_per, rng)
        n_inner = len(ml_i)
        dl_i, di_i = self._assign_bma("density", n_inner, rng)
        dlab, didx = np.concatenate([dl_o, dl_i]), np.concatenate([di_o, di_i])
        mlab, midx = np.concatenate([ml_o, ml_i]), np.concatenate([mi_o, mi_i])
        ind = {k: self._assign_indicator(k, mlab, midx, rng) for k in self._indicators()}
        rows = self._latents(meas, tgt, dlab, didx, mlab, midx, rng, indicators=ind)
        o = slice(0, n_outer)
        H_SD = H_MEASUREMENT_LOG_SD  # declared: log-sd of an H-content measurement relative to the desorption index
        # simulated outcomes, one per outer draw; per-term inner means and scales
        terms, own, own_d = [], np.zeros(n_outer), np.zeros(n_outer)
        for j, mm in enumerate(meas):
            R = rows[j]
            q = set(mm.get("quantities", ["Hci", "Br", "sq"]))
            reps = int(mm.get("replicates", 1))
            if "density" in q and not R["rho_known"]:
                y = R["mu_rho"][o] + R["sig_rho"][o] * rng.standard_t(5.0, n_outer)
                terms.append((_t_logpdf, y, R["mu_rho"][n_outer:], R["sig_rho"][n_outer:], True))
                own += _t_logpdf(y, R["mu_rho"][o], R["sig_rho"][o])
                own_d += _t_logpdf(y, R["mu_rho"][o], R["sig_rho"][o])
            spec = []
            if "Hci" in q:  # the body's specimen effect integrated: ln Hci ~ N(lh, sig_spec^2 + sig_piece^2/reps)
                spec.append(("lh", np.sqrt(R["sig_spec"] ** 2 + R["sig_piece"] ** 2 / reps)))
            if "Br" in q:
                spec.append(("lbr", R["sig_Br"] / math.sqrt(reps)))
            if "sq" in q:
                spec.append(("sq", R["sig_sq"] / math.sqrt(reps)))
            if "grain" in q:
                spec.append(("lgrain", np.full(len(R["lh"]), 0.15)))
            if "H_ppm" in q:
                spec.append(("lh_res", np.full(len(R["lh"]), H_SD)))
            if "Hci_ratio_T" in q:  # same body at two temperatures; delta-method specimen term (alpha follows the body)
                spec.append(("lratioT", np.sqrt(2.0 * R["sig_piece"] ** 2 / reps + (R["lratioT_g"] * R["sig_spec"]) ** 2)))
            ys = {}
            for key, sd in spec:
                if key == "lratioT" and "lh" in ys:
                    # Hci and Hci_ratio_T on the same body share its specimen effect u and the 20 C piece noise e20:
                    # ln Hci = lh + u + e20, ln ratio = r0 + g u + (eT - e20), so cov = g sig_spec^2 - sig_piece^2/reps.
                    # Modelled jointly as ln ratio | ln Hci (a conditional normal), not as an independent term.
                    v1 = R["sig_spec"] ** 2 + R["sig_piece"] ** 2 / reps
                    cov = R["lratioT_g"] * R["sig_spec"] ** 2 - R["sig_piece"] ** 2 / reps
                    cslope = cov / v1
                    sdc = np.sqrt(np.maximum(sd ** 2 - cov ** 2 / v1, 1e-12))
                    y1 = ys["lh"]
                    y = R[key][o] + cslope[o] * (y1 - R["lh"][o]) + sdc[o] * rng.standard_normal(n_outer)
                    terms.append((_n_logpdf, y, R[key][n_outer:], sdc[n_outer:], False,
                                  (y1, R["lh"][n_outer:], cslope[n_outer:])))
                    own += _n_logpdf(y, R[key][o] + cslope[o] * (y1 - R["lh"][o]), sdc[o])
                    continue
                y = R[key][o] + sd[o] * rng.standard_normal(n_outer)
                ys[key] = y
                terms.append((_n_logpdf, y, R[key][n_outer:], sd[n_outer:], False))
                own += _n_logpdf(y, R[key][o], sd[o])
        res = {"name": cand.get("name", ""), "measurements": meas, "cost": cand.get("cost"),
               "n_outer": n_outer, "n_inner_per_magnetic_mechanism": n_per}
        kinds = {}
        for kind, lab_o, lab_i in (("magnetic", ml_o, ml_i), ("density", dl_o, dl_i)):
            mechs = [m for m in self._mechs(kind) if self.w[kind][m] > 0]
            w = np.array([self.w[kind][m] for m in mechs])
            kinds[kind] = (mechs, w / w.sum(), lab_o, lab_i, np.full((n_outer, len(mechs)), -np.inf))
        # target-route moments (importance weights over the inner set, re-weighted to the model average)
        T = rows[-1]
        wbase = np.array([self.w["magnetic"][m] / (ml_i == m).sum() for m in ml_i])
        tgt_inner = T["lh"][n_outer:]
        M = np.full(n_outer, -np.inf)
        S0, S1, S2, Sq = (np.zeros(n_outer) for _ in range(4))
        for c0 in range(0, n_inner, chunk):
            c1 = min(n_inner, c0 + chunk)
            ll = np.zeros((n_outer, c1 - c0))
            ll_d = np.zeros((n_outer, c1 - c0))
            for tm in terms:
                f, y, mu, sd, is_density = tm[:5]
                if len(tm) > 5:  # conditional term: mean shifts with the paired observation y1 (same body)
                    y1, mu1, cs = tm[5]
                    mu_eff = mu[None, c0:c1] + cs[None, c0:c1] * (y1[:, None] - mu1[None, c0:c1])
                    term = f(y[:, None], mu_eff, sd[None, c0:c1])
                else:
                    term = f(y[:, None], mu[None, c0:c1], sd[None, c0:c1])
                ll += term
                if is_density:
                    ll_d += term
            for kind, (mechs, w, lab_o, lab_i, acc) in kinds.items():
                lab = lab_i[c0:c1]
                src = ll if kind == "magnetic" else ll_d
                for k, m in enumerate(mechs):
                    cols = lab == m
                    if cols.any():
                        acc[:, k] = np.logaddexp(acc[:, k], _lse(src[:, cols], 1))
            lw = ll + np.log(wbase[c0:c1])[None, :]
            cmax = np.max(lw, 1)
            newM = np.maximum(M, cmax)
            scale = np.exp(np.where(np.isfinite(M), M - newM, -np.inf))
            ew = np.exp(lw - newM[:, None])
            tt = tgt_inner[c0:c1]
            S0 = S0 * scale + ew.sum(1)
            S1 = S1 * scale + ew @ tt
            S2 = S2 * scale + ew @ (tt ** 2)
            Sq = Sq * scale ** 2 + (ew ** 2).sum(1)
            M = newM
        lpy_mag = None
        has_density = any(term[4] for term in terms)
        for kind, (mechs, w, lab_o, lab_i, acc) in kinds.items():
            idx = np.array([mechs.index(l) for l in lab_o])
            own_k = own if kind == "magnetic" else own_d
            est = {}
            for include_own in (False, True):
                logp = np.empty((n_outer, len(mechs)))
                for k, m in enumerate(mechs):
                    n_m = int((lab_i == m).sum())
                    mine = lab_o == m
                    if include_own:
                        logp[:, k] = (np.where(mine, np.logaddexp(acc[:, k], own_k), acc[:, k])
                                      - np.log(np.maximum(n_m + mine, 1)))
                    else:
                        logp[:, k] = acc[:, k] - math.log(max(n_m, 1))
                lpy = _lse(logp + np.log(w)[None, :], 1)
                contrib = logp[np.arange(n_outer), idx] - lpy
                ok = np.isfinite(contrib)
                n_ok = max(int(ok.sum()), 2)
                post = np.exp(logp + np.log(w)[None, :] - lpy[:, None])
                pid = float(np.mean((np.argmax(post, 1) == idx)[ok])) if ok.any() else float("nan")
                est[include_own] = (float(contrib[ok].mean()) if ok.any() else float("nan"),
                                    float(contrib[ok].std(ddof=1) / math.sqrt(n_ok)) if ok.sum() > 1 else float("nan"),
                                    post, lpy, pid, int((~ok).sum()))
            lo, hi = est[False][0], est[True][0]
            se = max(est[False][1], est[True][1])
            pid_lo, pid_hi = est[False][4], est[True][4]
            if kind == "density" and not has_density:  # nothing measured that the density mechanism predicts
                lo = hi = se = 0.0
            # the true EIG is >= 0; a negative lower estimate is small-sample bias, so the midpoint uses max(lo, 0)
            res[f"eig_{kind}_mechanism_nats"] = 0.5 * (max(lo, 0.0) + hi)
            res[f"eig_{kind}_mechanism_bracket"] = [lo, hi]
            res[f"eig_{kind}_mechanism_se"] = se
            res[f"eig_{kind}_mechanism_converged"] = bool(abs(hi - lo) <= max(0.05, 2 * se))
            res[f"n_nonfinite_outer_{kind}"] = max(est[False][5], est[True][5])
            post = est[True][2]
            res[f"expected_posterior_{kind}_weights"] = {m: float(post[:, k].mean()) for k, m in enumerate(mechs)}
            res[f"p_identify_{kind}_mechanism_map"] = 0.5 * (pid_lo + pid_hi)
            res[f"p_identify_{kind}_mechanism_bracket"] = [pid_lo, pid_hi]
            if kind == "magnetic":
                lpy_mag = est[True][3]
        contrib = own - lpy_mag
        res["eig_all_parameters_nats"] = float(contrib.mean())
        res["eig_all_parameters_se"] = float(contrib.std(ddof=1) / math.sqrt(n_outer))
        w0 = wbase / wbase.sum()
        now_var = float(w0 @ tgt_inner ** 2 - (w0 @ tgt_inner) ** 2)
        spec_var = float(w0 @ T["sig_spec"][n_outer:] ** 2)
        m1, m2 = S1 / S0, S2 / S0
        res["target_route"] = tgt
        res["ln_Hci_sd_target_now"] = math.sqrt(now_var + spec_var)
        res["ln_Hci_sd_target_expected_after"] = float(np.mean(np.sqrt(np.maximum(m2 - m1 ** 2, 0) + spec_var)))
        res["ess_inner_median"] = float(np.median(S0 ** 2 / Sq))
        return res

    # ---------------------------------------------------------------- experiment design (v4 estimator)
    def eig(self, candidates, n_outer=1500, n_inner_per_mechanism=None, seed=0, target=None, goal=None,
            method="v4"):
        """Rank candidate measurements by expected information gain (nats); see the eig section of the README.
        candidates: list of {"name": str, "measurements": [ {"route": {...}, "rho_observed": optional,
                    "quantities": ["Hci","Br","sq","density","grain","H_ppm","Hci_ratio_T"], "replicates": 1,
                    "T2_C": 120 (only for Hci_ratio_T)} ]}; each measurement is one body.
        For every per-draw discrete indicator the model carries (self.w: magnetic and density mechanisms, and any
        other, e.g. load timing) the EIG is bracketed by two estimates from the same simulated experiments:
          lower: a Barber-Agakov bound, H(prior weights) + E[log q(true label | y)], where q pools the exact nested
                 posterior over an inner set of stored draws that shares no stored draw with the simulated experiment
                 (weight rising with its effective sample size) with a classifier trained on simulated experiments
                 from the other half of the stored draws (cross-fitted), mixed with the prior at EIG_EPS. It is a
                 lower bound on the EIG in expectation whatever the quality of q;
          upper: the nested estimate with the generating draw included in its own stratum (as in v3.1, biased high
                 when the stored draws cannot resolve the design).
        The value (`*_nats`) is max(lower, 0): the design gains at least this much information about the indicator.
        `*_lower_bound_stable` (= `*_converged`): the two halves (independent simulated experiments, each scored by a
        classifier trained on the other half's stored draws) agree within 2 combined SE. `*_bracket_closed`: upper -
        lower <= max(0.05, 2 SE), so the value is the EIG itself (single-body designs); for multi-body designs the
        upper end stays loose because the stored posterior holds about 2,000 draws per mechanism.
        goal: optional {"route": {...}, "spec": {...p_meets_spec keys...}, "rho_observed": optional} adds the
        information about reaching the spec at that route (see `_eig_goal`).
        method="nmc-v3.1" runs the v3.1-r3 estimator unchanged (no goal, magnetic and density only)."""
        if method == "nmc-v3.1":
            if goal is not None:
                raise ValueError("goal needs method='v4'")
            extra = self._eig_indicator_names()[1]
            if extra:
                raise ValueError(f"method='nmc-v3.1' predates per-draw indicators beyond the mechanisms ({extra}); "
                                 "use method='v4'")
            return self._eig_v31(candidates, n_outer, n_inner_per_mechanism or 6000, seed, target)
        n_inner_per_mechanism = n_inner_per_mechanism or 4000
        if method != "v4":
            raise ValueError("method must be 'v4' (default) or 'nmc-v3.1' (the v3.1-r3 estimator, for comparison)")
        if not candidates:
            raise ValueError("eig needs at least one candidate")
        goal = self._eig_goal_parse(goal) if goal is not None else None
        out = [self._eig_one(cand, n_outer, n_inner_per_mechanism, np.random.default_rng(seed + 7919 * (ci + 1)),
                             target, goal) for ci, cand in enumerate(candidates)]
        names = list(out[0]["eig_indicators"])
        keys = [self._eig_key(n) + "_nats" for n in names] + ["eig_all_parameters_nats", "ln_Hci_sd_reduction_target"]
        if goal is not None:
            keys += ["eig_goal_event_nats", "goal_p_success_sd_reduction"]
        orderings = {k: [d["name"] for d in sorted(out, key=lambda d, k=k: -d[k])] for k in keys}
        out.sort(key=lambda d: (-d["eig_magnetic_mechanism_nats"], -d["eig_all_parameters_nats"]))
        for i, d in enumerate(out):
            d["rank"] = i + 1
        return self._stamp({
            "ranking": out, "orderings": orderings, "indicators": names,
            "ranking_rule": "primary: EIG on the magnetic mechanism indicator; tie-break: EIG on all parameters; see "
                            "orderings for every metric",
            "all_converged": bool(all(d["eig_indicators"][n]["lower_bound_stable"] for d in out for n in names)),
            "converged_rule": "value = the nested estimate H(prior) - E[H(posterior)], clipped into [lower bound, "
                              "upper], when the bracket is closed, "
                              "else max(lower bound, 0); lower_bound_stable (= converged, and all_converged over "
                              "every indicator and candidate): the two independent halves agree within 2 combined SE; "
                              "bracket_closed: upper - lower <= max(0.05 nats, 2 SE), i.e. the value is the EIG itself; "
                              "rank differences smaller than about 3 lower-bound SE are not resolved",
            "method": EIG_METHOD_V4 + ": simulated experiments from the stored posterior draws (two halves by stored-"
                      "draw parity); exact likelihoods of each simulated experiment under an inner set that excludes "
                      "its own stored draws (Gaussian measurement models in log space, Student-t(5) for density, "
                      "per-body specimen effects integrated; new-preparation effects, degas state, H0 discrepancy "
                      "direction and conditional indicators shared across routes within a draw); per indicator a "
                      "Barber-Agakov lower bound with the pooled posterior (exact nested posterior where its effective "
                      "sample size allows, cross-fitted classifier elsewhere) and the outer-draw-included nested "
                      "estimate as the upper end"})

    @staticmethod
    def _eig_key(name):
        return f"eig_{name}_mechanism" if name in ("magnetic", "density") else f"eig_{name}"

    def _eig_indicator_names(self):
        """(base, conditional) indicators. Base: every positive-weight label has stored parameter arrays (magnetic,
        density); eig assigns them. Conditional: labels drawn per magnetic draw by self._assign_indicator."""
        base, cond = [], []
        for name, wd in self.w.items():
            labs = [l for l, v in wd.items() if v > 0]
            (base if labs and all(l in self.post for l in labs) else cond).append(name)
        if "magnetic" not in base or "density" not in base:
            raise ValueError("eig needs the magnetic and density indicators with stored draws")
        base.sort(key=lambda n: (n != "magnetic", n != "density", n))
        return base, cond

    def _eig_labels(self, name):
        return [l for l, v in self.w[name].items() if v > 0]

    def _eig_draws(self, n, rng, balanced):
        """n joint draws. Base indicators: labels in equal shares (balanced, the inner set) or in proportion to the
        weights (the simulated experiments), each carrying the log importance weight log(w / share); stored indices by
        systematic resampling on the per-draw weights, restricted to one parity ("fold") per joint draw so the two
        halves share no stored draw. Conditional indicators: self._assign_indicator(name, mlab, midx, rng)."""
        base, cond = self._eig_indicator_names()
        fold = rng.permutation(np.arange(n) % 2)
        D = {"n": n, "fold": fold, "lab": {}, "idx": {}, "logw_own": {}, "logw": np.zeros(n)}
        for name in base:
            labs = self._eig_labels(name)
            w = np.array([self.w[name][l] for l in labs], float)
            w = w / w.sum()
            cnt = _alloc(w, n, balanced)
            code = np.repeat(np.arange(len(labs)), cnt)[rng.permutation(n)]
            lw = np.log(w)[code] - np.log(np.maximum(cnt, 1) / n)[code]
            idx = np.zeros(n, int)
            for k, l in enumerate(labs):
                K = len(next(iter(self.post[l].values())))
                for f in (0, 1):
                    pos = np.where((code == k) & (fold == f))[0]
                    if not len(pos):
                        continue
                    cand = np.arange(f, K, 2) if K > 1 else np.arange(K)
                    p = np.asarray(self.dw[l], float)[cand]
                    p = p if p.sum() > 0 else np.ones(len(cand))
                    idx[pos] = cand[_systematic(p, len(pos), rng)][rng.permutation(len(pos))]
            D["lab"][name], D["idx"][name], D["logw_own"][name] = np.array(labs)[code], idx, lw
            D["logw"] = D["logw"] + lw
        for name in cond:
            D["lab"][name] = np.asarray(self._assign_indicator(name, D["lab"]["magnetic"], D["idx"]["magnetic"], rng)
                                        ).astype(str)
            D["logw_own"][name] = np.zeros(n)
        return D

    @staticmethod
    def _eig_cat(A, B):
        cat = lambda a, b: np.concatenate([a, b])
        return {"n": A["n"] + B["n"], "fold": cat(A["fold"], B["fold"]), "logw": cat(A["logw"], B["logw"]),
                "lab": {k: cat(A["lab"][k], B["lab"][k]) for k in A["lab"]},
                "idx": {k: cat(A["idx"][k], B["idx"][k]) for k in A["idx"]},
                "logw_own": {k: cat(A["logw_own"][k], B["logw_own"][k]) for k in A["logw_own"]}}

    def _eig_measurements(self, cand):
        meas = cand.get("measurements") if isinstance(cand, dict) else None
        if not meas:
            raise ValueError("each candidate needs a non-empty 'measurements' list")
        out = []
        for mm in meas:
            q = list(mm.get("quantities", ["Hci", "Br", "sq"]))
            bad = [x for x in q if x not in EIG_QUANTITIES]
            if bad or not q:
                raise ValueError(f"unknown or empty quantities {bad}; allowed: {list(EIG_QUANTITIES)}")
            reps = mm.get("replicates", 1)
            if not (_is_number(reps) and int(reps) == reps and reps >= 1):
                raise ValueError("replicates must be a positive integer")
            r = self._route(mm["route"])
            if abs(float(r["T_measure_C"]) - 20.0) > 1e-9 and set(q) & {"Hci", "Br", "sq", "Hci_ratio_T"}:
                raise ValueError("eig models Hci, Br and Hk/Hci measured at 20 C: a route with T_measure_C != 20 is "
                                 "not supported for these quantities; measure the hot loop with Hci_ratio_T and T2_C")
            out.append(dict(mm, quantities=q, replicates=int(reps)))
        return out

    def _eig_latents(self, meas, extra_routes, D, rng):
        """Latent rows for the draws in D, one row per distinct (route, rho_observed); shared within a draw across
        routes: new-preparation effects, the degas state of each preparation, H0's discrepancy direction and the
        conditional indicators. Hci_ratio_T hooks are per measurement (as in v3.1)."""
        n = D["n"]
        mlab, midx = D["lab"]["magnetic"], D["idx"]["magnetic"]
        dlab, didx = D["lab"]["density"], D["idx"]["density"]
        cond = self._eig_indicator_names()[1]
        kw = {"indicators": {k: D["lab"][k] for k in cond}} if cond else {}
        z_hull = rng.standard_normal(n)
        prep_z, degas, eps, prep_dc, prep_fd = {}, {}, {}, {}, {}
        extra = None
        rows, rkey, mrow, xrow, ratio, rres = [], {}, [], [], {}, []
        items = [(mm["route"], mm.get("rho_observed")) for mm in meas] + [(rt, None) for rt in extra_routes]
        for j, (route, ro) in enumerate(items):
            r = self._route(route)
            key = json.dumps([r, ro], sort_keys=True, default=str)
            if key not in rkey:
                dc = json.dumps(r["degas_completeness"], sort_keys=True, default=str)
                if prep_dc.setdefault(r["prep"], dc) != dc:
                    raise ValueError(f"routes sharing preparation {r['prep']!r} must share degas_completeness")
                _check_prep_feedstock(prep_fd, r)
                r["_degas_expected"] = self._expected_state(r)
                A = self._arrays(r)
                if r["prep"] not in prep_z:
                    prep_z[r["prep"]] = rng.standard_normal((n, 1))
                    u = rng.random(n)
                    if not self._completeness_is_default(r) and extra is None:
                        extra = rng.spawn(1)[0]
                    degas[r["prep"]], eps[r["prep"]] = self._degas_state(r, u, extra)
                mu, sig = self._density_mean(A, r, dlab, didx, prep_z[r["prep"]])
                rho = np.full(n, (ro / RHO_TD if ro > 1.5 else ro)) if ro else np.clip(mu, 0.3, 1.0)
                lat = self._magnetic_latent(A, r, mlab, midx, rho, degas[r["prep"]], z_hull, eps=eps[r["prep"]], **kw)
                rkey[key] = len(rows)
                rows.append(dict(mu_rho=mu, sig_rho=sig, rho_known=ro is not None, **lat))
                rres.append(r)
            (mrow if j < len(meas) else xrow).append(rkey[key])
        for j, mm in enumerate(meas):
            if "Hci_ratio_T" in mm["quantities"]:
                if extra is None:
                    extra = rng.spawn(1)[0]
                ratio[j] = self._ln_hci_ratio_T(rows[mrow[j]], float(mm.get("T2_C", 120.0)), extra)
        return {"rows": rows, "routes": rres, "mrow": mrow, "xrow": xrow, "ratio": ratio, "kw": kw,
                "nuis": {"z_hull": z_hull, "prep_z": prep_z, "degas": degas, "eps": eps}}

    @staticmethod
    def _eig_obs(meas, L):
        """Observation models, one entry per measured quantity per body (arrays over all draws):
        kind "g": Gaussian in log space (mean mu, sd), optionally conditional on the same body's Hci (pair, cs);
        kind "t": Student-t(5) density. col groups observations of one latent for the summaries."""
        obs = []
        for j, mm in enumerate(meas):
            rk = L["mrow"][j]
            R = L["rows"][rk]
            q, reps = set(mm["quantities"]), mm["replicates"]
            if "density" in q and not R["rho_known"]:
                obs.append(dict(kind="t", mu=R["mu_rho"], sd=R["sig_rho"], pair=-1, cs=None, col=(rk, "density")))
            hci = -1
            for qn, key, sd in (("Hci", "lh", lambda: np.sqrt(R["sig_spec"] ** 2 + R["sig_piece"] ** 2 / reps)),
                                ("Br", "lbr", lambda: R["sig_Br"] / math.sqrt(reps)),
                                ("sq", "sq", lambda: R["sig_sq"] / math.sqrt(reps)),
                                ("grain", "lgrain", lambda: np.full(len(R["lh"]), 0.15)),
                                ("H_ppm", "lh_res", lambda: np.full(len(R["lh"]), H_MEASUREMENT_LOG_SD))):
                if qn in q:
                    obs.append(dict(kind="g", mu=R[key], sd=sd(), pair=-1, cs=None, col=(rk, key)))
                    if qn == "Hci":
                        hci = len(obs) - 1
            if "Hci_ratio_T" in q:  # same body at two temperatures (v3.1-r3 joint model with its Hci)
                r0, g = L["ratio"][j]
                sd = np.sqrt(2.0 * R["sig_piece"] ** 2 / reps + (g * R["sig_spec"]) ** 2)
                if hci >= 0:
                    v1 = R["sig_spec"] ** 2 + R["sig_piece"] ** 2 / reps
                    cov = g * R["sig_spec"] ** 2 - R["sig_piece"] ** 2 / reps
                    obs.append(dict(kind="g", mu=r0, sd=np.sqrt(np.maximum(sd ** 2 - cov ** 2 / v1, 1e-12)),
                                    pair=hci, cs=cov / v1, col=("ratio", j)))
                else:
                    obs.append(dict(kind="g", mu=r0, sd=sd, pair=-1, cs=None, col=("ratio", j)))
        return obs

    @staticmethod
    def _eig_simulate(obs, sel, rng):
        n = sel.stop - sel.start
        Y = np.empty((n, len(obs)))
        for t, ob in enumerate(obs):
            mu, sd = ob["mu"][sel], ob["sd"][sel]
            if ob["kind"] == "t":
                Y[:, t] = mu + sd * rng.standard_t(5.0, n)
            elif ob["pair"] >= 0:
                p = ob["pair"]
                Y[:, t] = mu + ob["cs"][sel] * (Y[:, p] - obs[p]["mu"][sel]) + sd * rng.standard_normal(n)
            else:
                Y[:, t] = mu + sd * rng.standard_normal(n)
        return Y

    @staticmethod
    def _eig_loglik_own(obs, sel, Y):
        """Log-likelihood of each simulated outcome under its own draw (the rows of Y align with sel)."""
        ll = np.zeros(sel.stop - sel.start)
        for t, ob in enumerate(obs):
            mu, sd = ob["mu"][sel], ob["sd"][sel]
            if ob["kind"] == "t":
                ll += _t_logpdf(Y[:, t], mu, sd)
            else:
                m = mu + (ob["cs"][sel] * (Y[:, ob["pair"]] - obs[ob["pair"]]["mu"][sel]) if ob["pair"] >= 0 else 0.0)
                ll += _n_logpdf(Y[:, t], m, sd)
        return ll

    @staticmethod
    def _eig_summaries(obs, Y):
        """Classifier inputs: per latent, the mean of its body observations; where >= 2 bodies, the log sd."""
        cols = {}
        for t, ob in enumerate(obs):
            cols.setdefault(ob["col"], []).append(t)
        M = [Y[:, ts].mean(1) for ts in cols.values()]
        V = [np.log(Y[:, ts].std(1, ddof=1) + 1e-3) for ts in cols.values() if len(ts) >= 2]
        return np.stack(M + V, 1), len(M)

    @staticmethod
    def _eig_gauss_terms(obs, sp, center):
        """Gaussian log densities of outcomes under the draws in sp as feat(Y) @ C + c0, where feat(Y) =
        [Y - c, (Y - c)^2, cross products of conditional pairs] over the Gaussian observations (one matrix product)."""
        gi = [t for t, ob in enumerate(obs) if ob["kind"] == "g"]
        pos = {t: k for k, t in enumerate(gi)}
        pairs = [t for t in gi if obs[t]["pair"] >= 0]
        G, n = len(gi), sp.stop - sp.start
        C, c0 = np.zeros((2 * G + len(pairs), n)), np.zeros(n)
        for k, t in enumerate(gi):
            ob = obs[t]
            w = 1.0 / ob["sd"][sp] ** 2
            kap = ob["mu"][sp] - center[t]
            if ob["pair"] >= 0:
                p, cs = ob["pair"], ob["cs"][sp]
                kap = kap - cs * (obs[p]["mu"][sp] - center[p])
                C[G + pos[p]] += -0.5 * cs * cs * w
                C[2 * G + pairs.index(t)] += cs * w
                C[pos[p]] += -cs * kap * w
            C[k] += kap * w
            C[G + k] += -0.5 * w
            c0 += -0.5 * kap * kap * w - np.log(ob["sd"][sp]) - 0.5 * math.log(2 * math.pi)

        def feat(Y):
            Yc = Y[:, gi] - center[gi]
            cr = [((Y[:, t] - center[t]) * (Y[:, obs[t]["pair"]] - center[obs[t]["pair"]]))[:, None] for t in pairs]
            return np.concatenate([Yc, Yc ** 2] + cr, 1)
        return feat, C, c0

    def _eig_nmc(self, obs, YO, sp, keysO, keysP, logwP, H, X, chunk_o=400, chunk_p=4000):
        """Exact likelihood of every simulated outcome (rows of YO) under every inner draw (sp), leaving out inner
        draws that share a stored draw with the outcome's own draw (keys). Returns log sums of w f per label column of
        H, of (w f)^2 per column, of w f over all inner draws, and posterior means of the columns of X."""
        no, npool = YO.shape[0], sp.stop - sp.start
        center = YO.mean(0)
        feat, C, c0 = self._eig_gauss_terms(obs, sp, center)
        F = feat(YO)
        ti = [t for t, ob in enumerate(obs) if ob["kind"] == "t"]
        tmu = [obs[t]["mu"][sp] for t in ti]
        tinv = [1.0 / obs[t]["sd"][sp] for t in ti]
        tc = math.lgamma(3.0) - math.lgamma(2.5) - 0.5 * math.log(5.0 * math.pi)
        c0 = c0 + sum(tc - np.log(obs[t]["sd"][sp]) for t in ti)
        L, nx = H.shape[1], X.shape[1]
        out = {"logS": np.empty((no, L)), "logS2": np.empty((no, L)), "logSall": np.empty(no),
               "mean_x": np.empty((no, nx))}
        # the left-out entries (copies of the outcome's own stored draws) are kept apart: adding them back gives the
        # v3.1-style estimate in which the generating stored draw is part of the inner set ("stored-draw" estimate)
        mo, mp, mv = [], [], []
        with np.errstate(divide="ignore", invalid="ignore", over="ignore"):
            for o0 in range(0, no, chunk_o):
                os_ = slice(o0, min(no, o0 + chunk_o))
                m = os_.stop - os_.start
                M = np.full(m, -np.inf)
                S, S2, SX, Sall = np.zeros((m, L)), np.zeros((m, L)), np.zeros((m, nx)), np.zeros(m)
                for p0 in range(0, npool, chunk_p):
                    ps = slice(p0, min(npool, p0 + chunk_p))
                    ll = F[os_] @ C[:, ps] + c0[ps]
                    for k, t in enumerate(ti):
                        z = (YO[os_, t][:, None] - tmu[k][ps][None]) * tinv[k][ps][None]
                        z *= z
                        z *= 0.2
                        ll -= 3.0 * np.log1p(z)
                    mask = np.zeros(ll.shape, bool)
                    for kO, kP in zip(keysO, keysP):
                        mask |= kO[os_][:, None] == kP[ps][None, :]
                    r_, c_ = np.nonzero(mask)
                    if len(r_):
                        mo.append(r_ + o0)
                        mp.append(c_ + p0)
                        mv.append(ll[r_, c_] + logwP[ps][c_])
                    ll[mask] = -np.inf
                    ll += logwP[ps][None]
                    newM = np.maximum(M, ll.max(1))
                    shift = np.where(np.isfinite(newM), newM, 0.0)
                    sc = np.where(np.isfinite(M), np.exp(M - shift), 0.0)
                    E = np.exp(ll - shift[:, None])
                    S = S * sc[:, None] + E @ H[ps]
                    S2 = S2 * (sc * sc)[:, None] + (E * E) @ H[ps]
                    SX = SX * sc[:, None] + E @ X[ps]
                    Sall = Sall * sc + E.sum(1)
                    M = newM
                out["logS"][os_] = np.log(S) + M[:, None]
                out["logS2"][os_] = np.log(S2) + 2 * M[:, None]
                out["logSall"][os_] = np.log(Sall) + M
                out["mean_x"][os_] = SX / Sall[:, None]
        logSm, logSm_all = np.full((no, L), -np.inf), np.full(no, -np.inf)
        if mo:
            mo, mp, mv = np.concatenate(mo), np.concatenate(mp), np.concatenate(mv)
            np.logaddexp.at(logSm_all, mo, mv)
            Hb = H[mp] > 0
            for j in range(L):
                sel = Hb[:, j]
                np.logaddexp.at(logSm[:, j], mo[sel], mv[sel])
        out["logS_copies"], out["logSall_copies"] = logSm, logSm_all
        return out

    def _eig_structural_zero(self, name, meas, extra_routes, rng_seed):
        """True when no measured quantity depends on the indicator: latents recomputed for 64 draws with only this
        indicator's labels (and, for a base indicator, its stored draw) changed are identical."""
        D = self._eig_draws(64, np.random.default_rng(rng_seed), balanced=True)
        labs = self._eig_labels(name)
        if len(labs) < 2:
            return True
        D2 = {**D, "lab": dict(D["lab"]), "idx": dict(D["idx"])}
        new = np.array([labs[(labs.index(l) + 1) % len(labs)] for l in D["lab"][name]])
        D2["lab"][name] = new
        if name in D["idx"]:
            r = np.random.default_rng(rng_seed + 1)
            D2["idx"][name] = np.array([r.integers(len(next(iter(self.post[l].values())))) for l in new])
        o1 = self._eig_obs(meas, self._eig_latents(meas, extra_routes, D, np.random.default_rng(rng_seed + 2)))
        o2 = self._eig_obs(meas, self._eig_latents(meas, extra_routes, D2, np.random.default_rng(rng_seed + 2)))
        return all(np.array_equal(a["mu"], b["mu"]) and np.array_equal(a["sd"], b["sd"]) for a, b in zip(o1, o2))

    def _eig_one(self, cand, n_outer, n_per, rng, target, goal=None):
        meas = self._eig_measurements(cand)
        tgt = target or meas[0]["route"]
        base, cond = self._eig_indicator_names()
        names = base + cond
        n_mag = len(self._eig_labels("magnetic"))
        N, n_outer = max(4, int(n_per)) * n_mag, max(40, int(n_outer))
        extra_routes = [tgt] + ([goal["route"]] if goal is not None else [])
        zero = {nm: self._eig_structural_zero(nm, meas, extra_routes, int(rng.integers(2 ** 31))) for nm in names}
        DP, DO = self._eig_draws(N, rng, balanced=True), self._eig_draws(n_outer, rng, balanced=False)
        D = self._eig_cat(DP, DO)
        L = self._eig_latents(meas, extra_routes, D, rng)
        obs = self._eig_obs(meas, L)
        sp, so = slice(0, N), slice(N, N + n_outer)
        YO, YP = self._eig_simulate(obs, so, rng), self._eig_simulate(obs, sp, rng)
        own = self._eig_loglik_own(obs, so, YO)
        # label columns of every indicator, leave-one-out keys of every stored draw, posterior-mean columns
        cols, H = {}, []
        for nm in names:
            labs = self._eig_labels(nm)
            start = sum(h.shape[1] for h in H)
            H.append((D["lab"][nm][sp][:, None] == np.array(labs)[None, :]).astype(float))
            cols[nm] = list(range(start, start + len(labs)))
        H = np.concatenate(H, 1)
        keys = []
        for nm in base:
            labs = self._eig_labels(nm)
            code = np.array([labs.index(l) for l in D["lab"][nm]], np.int64)
            keys.append(code * 10 ** 7 + D["idx"][nm])
        T = L["rows"][L["xrow"][0]]
        X = [T["lh"][sp]]
        gp = None
        if goal is not None:
            kb = np.concatenate([np.full(N, GOAL_BODIES_POOL), np.full(n_outer, GOAL_BODIES_OUTER)])
            gp = self._eig_goal_success(goal, L, D, kb, rng)
            X.append(gp["p"][sp])
        X = np.stack(X, 1)
        nm_ = self._eig_nmc(obs, YO, sp, [k[so] for k in keys], [k[sp] for k in keys], D["logw"][sp], H, X)
        ess = np.exp(2 * nm_["logSall"] - _lse(nm_["logS2"][:, cols["magnetic"]], 1))
        lo_, hi_ = (math.log(v) for v in EIG_ESS_POOL)
        lam = np.clip((np.log(np.maximum(ess, 1e-300)) - lo_) / (hi_ - lo_), 0.0, 1.0)
        (ZP, nq), (ZO, _) = self._eig_summaries(obs, YP), self._eig_summaries(obs, YO)
        fP, fO = DP["fold"], DO["fold"]
        res = {"name": cand.get("name", ""), "measurements": cand["measurements"], "cost": cand.get("cost"),
               "n_outer": n_outer, "n_inner_per_magnetic_mechanism": int(N // n_mag), "n_inner": N,
               "eig_indicators": {}}
        ar = np.arange(n_outer)
        wpool = np.exp(DP["logw"] - DP["logw"].max())
        for nm in names:
            labs = self._eig_labels(nm)
            w = np.array([self.w[nm][l] for l in labs], float)
            if nm in cond:  # marginal of a conditional indicator as the draws generate it (sum over draws of p(label))
                w = np.array([wpool[DP["lab"][nm] == l].sum() for l in labs]) + 1e-12
            w = w / w.sum()
            HM = float(-(w * np.log(w)).sum())
            y = np.array([labs.index(l) for l in DO["lab"][nm]])
            r = {"labels": labs, "prior_weights": dict(zip(labs, w.tolist())), "prior_entropy_nats": HM}
            if zero[nm]:
                r.update(stored_draw_bracket=[0.0, 0.0], stored_draw_se=0.0, stored_draw_closed=True, stored_draw_nats=0.0)
                r.update(lower_bound_source="structural zero", lower_pooled=0.0)
                r.update(nats=0.0, value_se=0.0, value_kind="structural zero", nested_estimate=0.0,
                         nested_estimate_se=0.0, lower=0.0, upper=0.0, se=0.0, lower_bound_halves=[0.0, 0.0],
                         lower_bound_stable=True,
                         bracket_closed=True, converged=True, structural_zero=True, lower_nested=0.0,
                         lower_classifier=0.0,
                         expected_posterior_weights=dict(zip(labs, w.tolist())),
                         p_identify_map=float(np.mean(np.argmax(w) == y)), p_identify_map_bracket=[
                             float(np.mean(np.argmax(w) == y))] * 2, n_nonfinite_outer=0)
                res["eig_indicators"][nm] = r
                continue
            lS = nm_["logS"][:, cols[nm]]
            with np.errstate(divide="ignore", invalid="ignore"):
                q_nmc = np.exp(lS - nm_["logSall"][:, None])
            q_nmc = np.where(np.isfinite(q_nmc), q_nmc, 0.0)
            q_clf = np.empty((n_outer, len(labs)))
            sw = np.exp(DP["logw"] - DP["logw_own"][nm])
            yP = np.array([labs.index(l) for l in DP["lab"][nm]])
            for f in (0, 1):
                tr, ev = fP == f, fO != f
                if not ev.any():
                    continue
                Tt = np.eye(len(labs))[yP[tr]]
                clf = _Softmax(ZP[tr], Tt, sw[tr], nq)
                prior_tr = np.maximum((Tt * sw[tr][:, None]).sum(0) / sw[tr].sum(), 1e-12)
                lp = clf.log_prob(ZO[ev]) + np.log(w)[None] - np.log(prior_tr)[None]
                q_clf[ev] = np.exp(lp - _lse(lp, 1)[:, None])
            qp = (1 - EIG_EPS) * (lam[:, None] * q_nmc + (1 - lam[:, None]) * q_clf) + EIG_EPS * w[None]
            qn = (1 - EIG_EPS) * q_nmc + EIG_EPS * w[None]
            qc = (1 - EIG_EPS) * q_clf + EIG_EPS * w[None]
            # three valid Barber-Agakov bounds on the same simulated experiments (pooled; exact nested posterior with
            # the generating stored draw left out, i.e. v3.1's exclude-own estimate made leave-out; classifier); the
            # lower end is the largest of them
            cands = {"pooled": qp, "nested": qn, "classifier": qc}
            lqs = {k: np.log(v[ar, y]) for k, v in cands.items()}
            bounds = {k: HM + float(v.mean()) for k, v in lqs.items()}
            src = max(bounds, key=bounds.get)
            q, lq, lower = cands[src], lqs[src], bounds[src]
            se = float(lq.std(ddof=1) / math.sqrt(n_outer))
            # upper: generating draw included in its own stratum (log p(y | m) with own term, v3.1 style)
            logp = lS - np.log(N * w)[None]
            logp[ar, y] = np.logaddexp(lS[ar, y], own) - np.log(N * w[y] + 1)
            lpi = logp + np.log(w)[None]
            q_in = np.exp(lpi - _lse(lpi, 1)[:, None])
            upper = HM + float(np.mean(np.log(np.maximum(q_in[ar, y], 1e-300))))
            # stored-draw estimate (v3.1 style): inner set including copies of the generating stored draw
            lSf = np.logaddexp(lS, nm_["logS_copies"][:, cols[nm]])
            lq_sd = np.clip(lSf[ar, y] - np.logaddexp(nm_["logSall"], nm_["logSall_copies"]), -50.0, 0.0)
            sd_lower, sd_se = HM + float(lq_sd.mean()), float(lq_sd.std(ddof=1) / math.sqrt(n_outer))
            sd_closed = bool(abs(upper - sd_lower) <= max(0.05, 2 * sd_se))
            halves, stable = _eig_halves(lq, fO, HM)
            closed = bool(abs(upper - lower) <= max(0.05, 2 * se))
            # closed bracket: the posterior is accurate, so report the nested estimate in its lower-variance form
            # H(prior) - E[H(q(. | y))] (the generating label averaged out); otherwise the lower bound
            ent = -(qn * np.log(qn)).sum(1)  # the exact nested posterior (generating stored draw left out)
            nested_val, nested_se = HM - float(ent.mean()), float(ent.std(ddof=1) / math.sqrt(n_outer))
            clipped = bool(closed and not (lower <= nested_val <= upper))
            r.update(nats=min(max(nested_val, lower), upper) if closed else max(lower, 0.0),
                     value_se=nested_se if closed else se,
                     value_kind=("nested estimate (bracket closed)" + (", clipped to the bracket" if clipped else ""))
                     if closed else "lower bound", value_clipped_to_bracket=clipped,
                     nested_estimate=nested_val, nested_estimate_se=nested_se,
                     lower=lower, upper=upper, se=se, lower_bound_source=src, lower_bound_halves=halves,
                     lower_bound_stable=stable, bracket_closed=closed,
                     converged=stable, structural_zero=False,
                     stored_draw_bracket=[sd_lower, upper], stored_draw_se=sd_se, stored_draw_closed=sd_closed,
                     stored_draw_nats=0.5 * (max(sd_lower, 0.0) + upper) if sd_closed else None,
                     lower_pooled=bounds["pooled"], lower_nested=bounds["nested"],
                     lower_classifier=bounds["classifier"],
                     expected_posterior_weights=dict(zip(labs, q.mean(0).tolist())),
                     p_identify_map=float(np.mean(np.argmax(q, 1) == y)),
                     p_identify_map_bracket=[float(np.mean(np.argmax(q, 1) == y)),
                                             float(np.mean(np.argmax(q_in, 1) == y))],
                     n_nonfinite_outer=int((~np.isfinite(lq)).sum()))
            res["eig_indicators"][nm] = r
        for nm, r in res["eig_indicators"].items():  # flat keys (v3.1 names for magnetic and density)
            k = self._eig_key(nm)
            res[k + "_nats"], res[k + "_bracket"] = r["nats"], [r["lower"], r["upper"]]
            res[k + "_se"], res[k + "_converged"] = r["value_se"], r["converged"]
            res[k + "_lower_bound"], res[k + "_lower_bound_se"] = r["lower"], r["se"]
            res[k + "_lower_bound_stable"], res[k + "_bracket_closed"] = r["lower_bound_stable"], r["bracket_closed"]
            res[k + "_value_se"] = r["value_se"]
            if nm in ("magnetic", "density"):
                res[f"n_nonfinite_outer_{nm}"] = r["n_nonfinite_outer"]
                res[f"expected_posterior_{nm}_weights"] = r["expected_posterior_weights"]
                res[f"p_identify_{nm}_mechanism_map"] = r["p_identify_map"]
                res[f"p_identify_{nm}_mechanism_bracket"] = r["p_identify_map_bracket"]
        # all parameters: prior-contrastive bound with the generating draw included (a lower bound)
        c = own - (np.logaddexp(own, nm_["logSall"]) - math.log(N + 1))
        res["eig_all_parameters_nats"] = float(c.mean())
        res["eig_all_parameters_se"] = float(c.std(ddof=1) / math.sqrt(n_outer))
        # target route: expected posterior sd of the route-level ln Hci (an upper bound: held-out prediction error)
        wP = np.exp(DP["logw"] - DP["logw"].max())
        wP /= wP.sum()
        tP, tO = T["lh"][sp], T["lh"][so]
        spec_var = float(wP @ T["sig_spec"][sp] ** 2)
        pred = np.empty(n_outer)
        for f in (0, 1):
            tr, ev = fP == f, fO != f
            if ev.any():
                pred[ev] = _ridge(ZP[tr], tP[tr], wP[tr], ZO[ev], nq)
        mean_nmc = np.where(np.isfinite(nm_["mean_x"][:, 0]), nm_["mean_x"][:, 0], pred)
        pred = lam * mean_nmc + (1 - lam) * pred
        res["target_route"] = tgt
        res["ln_Hci_sd_target_now"] = math.sqrt(float(wP @ tP ** 2 - (wP @ tP) ** 2) + spec_var)
        res["ln_Hci_sd_target_expected_after"] = math.sqrt(float(np.mean((tO - pred) ** 2)) + spec_var)
        res["ln_Hci_sd_reduction_target"] = res["ln_Hci_sd_target_now"] - res["ln_Hci_sd_target_expected_after"]
        res["ess_inner_median"] = float(np.median(ess))
        res["ess_inner_q10"] = float(np.quantile(ess, 0.10))
        res["pooling_weight_nested_mean"] = float(lam.mean())
        if goal is not None:
            res["goal"] = self._eig_goal_summary(goal, gp, nm_["mean_x"][:, 1], lam, ZP, ZO, nq, fP, fO, DP, sp, so)
            res["eig_goal_event_nats"] = res["goal"]["eig_event_nats"]
            res["goal_p_success_sd_reduction"] = (res["goal"]["p_success_epistemic_sd_now"]
                                                  - res["goal"]["p_success_epistemic_sd_expected_after"])
        return res

    # ---------------------------------------------------------------- goal-oriented target (v4)
    def _eig_goal_parse(self, goal):
        if not isinstance(goal, dict) or "route" not in goal or "spec" not in goal:
            raise ValueError('goal needs {"route": {...}, "spec": {...}, "rho_observed": optional}')
        bad = sorted(set(goal) - {"route", "spec", "rho_observed"})
        if bad:
            raise ValueError(f"unknown goal keys {bad}; allowed: route, spec, rho_observed")
        spec = dict(_eig_spec_check(dict(goal["spec"])))
        route = dict(goal["route"])
        if "T_operating_C" in spec:
            route["T_measure_C"] = float(spec["T_operating_C"])
        self._route(route)
        ro = goal.get("rho_observed")
        if ro is not None and not (_is_number(ro) and ro > 0):
            raise ValueError("goal rho_observed must be a positive number (g/cm3 or relative)")
        return {"route": route, "spec": spec, "rho_observed": ro}

    def _eig_goal_success(self, goal, L, D, kb, rng):
        """p(goal | draw): the fraction of kb[i] simulated bodies made by the goal route that meet the spec, scored as
        in p_meets_spec (measured values: body density scatter, specimen, piece, Br, squareness and curve-shape scatter;
        the temperature scenario at T_operating_C), with the draw's parameters and its shared latents (preparation
        effect, degas state, H0 direction, conditional indicators) held fixed. This is the aleatoric success
        probability given the draw; its spread over draws is the epistemic uncertainty."""
        xi = L["xrow"][-1]
        R, r = L["rows"][xi], dict(L["routes"][xi])
        A = self._arrays(r)
        rep = np.repeat(np.arange(D["n"]), kb)
        nb = len(rep)
        if goal["rho_observed"] is not None:
            ro = float(goal["rho_observed"])
            rho = (ro / RHO_TD if ro > 1.5 else ro) + 0.005 * rng.standard_normal(nb)
        else:
            rho = np.clip(R["mu_rho"][rep] + R["sig_rho"][rep] * rng.standard_t(5.0, nb), 0.3, 1.02)
        nu = L["nuis"]
        eps = nu["eps"][r["prep"]]
        kw = {"indicators": {k: v[rep] for k, v in L["kw"]["indicators"].items()}} if L["kw"] else {}
        u = R["sig_spec"][rep] * rng.standard_normal(nb)
        lat = self._magnetic_latent(A, r, D["lab"]["magnetic"][rep], D["idx"]["magnetic"][rep], rho,
                                    nu["degas"][r["prep"]][rep], nu["z_hull"][rep], u_spec=u,
                                    eps=None if eps is None else eps[rep], **kw)
        hci = np.exp(lat["lh"]) * np.exp(lat["sig_piece"] * rng.standard_normal(nb))
        br = np.exp(lat["lbr"] + lat["sig_Br"] * rng.standard_normal(nb))
        sq = 1.0 / (1.0 + np.exp(-(lat["sq"] + lat["sig_sq"] * rng.standard_normal(nb))))
        hci_T, br_T = self._temperature(lat, hci, br, float(r["T_measure_C"]), rng)
        bh = self._bh(br_T, hci_T, sq, rng)
        vals = {"Hci_kOe_min": hci_T, "BHmax_MGOe_min": bh, "Br_kGs_min": br_T, "density_rel_min": rho}
        ok = np.ones(nb, bool)
        for k, v in vals.items():
            if k in goal["spec"]:
                ok &= np.asarray(v) >= float(goal["spec"][k])
        return {"p": np.bincount(rep, weights=ok.astype(float), minlength=D["n"]) / np.asarray(kb, float),
                "kb": np.asarray(kb)}

    def _eig_goal_summary(self, goal, gp, p_nmc, lam, ZP, ZO, nq, fP, fO, DP, sp, so):
        """Information about reaching the goal. G = 1 if the next body made by the goal route meets the spec.
        theta = a joint draw (parameters and shared latents). p(theta) = P(G = 1 | theta) is aleatoric (body-to-body
        and measurement scatter); its spread over the posterior is epistemic and is what an experiment can reduce."""
        pP, pO = gp["p"][sp], gp["p"][so]
        kP, kO = float(GOAL_BODIES_POOL), float(GOAL_BODIES_OUTER)
        wP = np.exp(DP["logw"] - DP["logw"].max())
        wP /= wP.sum()
        pbar = float(wP @ pP)
        res = {"route": self._route(goal["route"]), "spec": goal["spec"], "rho_observed": goal["rho_observed"],
               "p_success_now": pbar,
               "p_success_now_se": float(math.sqrt(max(wP @ (pP - pbar) ** 2, 1e-30) / max(1.0 / (wP @ wP), 1.0)))}
        var_now = max(float(wP @ (pP - pbar) ** 2 - wP @ (pP * (1 - pP)) / (kP - 1)), 0.0)
        H_event = float(_bern_entropy(pbar))
        h_ale = _bern_entropy(pO) + np.where((pO > 0) & (pO < 1), 1.0 / (2 * kO), 0.0)
        aleatoric = float(min(np.mean(np.minimum(h_ale, math.log(2))), H_event))
        if pbar <= 1e-12 or pbar >= 1 - 1e-12:
            res.update(no_draw_meets_spec=bool(pbar <= 1e-12), all_draws_meet_spec=bool(pbar >= 1 - 1e-12),
                       event_entropy_nats=0.0, aleatoric_entropy_nats=0.0, epistemic_entropy_nats=0.0,
                       eig_event_nats=0.0, eig_event_bracket=[0.0, 0.0], eig_event_se=0.0, eig_event_converged=True,
                       eig_event_lower_bound_stable=True, eig_event_bracket_closed=True,
                       p_success_epistemic_sd_now=0.0, p_success_epistemic_sd_expected_after=0.0,
                       expected_posterior_p_success_quantiles={f"q{int(100 * q):02d}": pbar for q in GOAL_QUANTILES})
            return res
        pred = np.empty(len(pO))
        for f in (0, 1):
            tr, ev = fP == f, fO != f
            if ev.any():
                Tt = np.stack([1 - pP[tr], pP[tr]], 1)
                clf = _Softmax(ZP[tr], Tt, wP[tr], nq)
                pred[ev] = np.exp(clf.log_prob(ZO[ev])[:, 1])
        p_nmc = np.where(np.isfinite(p_nmc), p_nmc, pred)
        m = (1 - EIG_EPS) * (lam * p_nmc + (1 - lam) * pred) + EIG_EPS * pbar
        m = np.clip(m, 1e-12, 1 - 1e-12)
        c = pO * np.log(m) + (1 - pO) * np.log1p(-m)
        lower = H_event + float(c.mean())
        se = float(c.std(ddof=1) / math.sqrt(len(c)))
        upper = max(H_event - aleatoric, 0.0)
        halves, stable = _eig_halves(c, fO, H_event)
        after = max(float(np.mean((pO - m) ** 2) - np.mean(pO * (1 - pO)) / (kO - 1)), 0.0)
        res.update(no_draw_meets_spec=False, all_draws_meet_spec=False,
                   event_entropy_nats=H_event, aleatoric_entropy_nats=aleatoric,
                   epistemic_entropy_nats=upper,
                   eig_event_nats=max(lower, 0.0), eig_event_bracket=[lower, upper], eig_event_se=se,
                   eig_event_lower_bound_halves=halves, eig_event_lower_bound_stable=stable,
                   eig_event_converged=stable, eig_event_bracket_closed=bool(upper - lower <= max(0.05, 2 * se)),
                   p_success_epistemic_sd_now=math.sqrt(var_now),
                   p_success_epistemic_sd_expected_after=min(math.sqrt(after), math.sqrt(var_now)),
                   expected_posterior_p_success_quantiles={f"q{int(round(100 * q)):02d}": float(np.quantile(m, q))
                                                           for q in GOAL_QUANTILES},
                   expected_posterior_p_success_mean=float(m.mean()),
                   bodies_per_draw={"inner": int(kP), "outer": int(kO)},
                   definitions="G = 1 if the next body made by the goal route meets every spec key (scored on measured "
                               "values as p_meets_spec). theta = a joint posterior draw (parameters and shared latents); "
                               "p(theta) = P(G = 1 | theta). event_entropy = H(E[p]); aleatoric = E[H(p(theta))] "
                               "(scatter no experiment removes); epistemic = event - aleatoric = I(G; theta), the most "
                               "any experiment can gain about G. eig_event: I(G; y) bracketed by a Barber-Agakov lower "
                               "bound and I(G; theta). p_success_epistemic_sd: sd of p(theta) now and the expected "
                               "posterior sd after the experiment (an upper bound: held-out prediction error). "
                               "expected_posterior_p_success_quantiles: the distribution over simulated outcomes of the "
                               "updated success probability (approximate; shrunk toward the mean where the approximation "
                               "is weak)")
        return res

    # ---------------------------------------------------------------- conditioning on new data
    def condition(self, observations, k_latent=32, seed=12345):
        """Importance-reweight the stored draws by new observations; returns (updated model, info).
        observations: list of {"route": {...}, "rho_observed": optional, "Hci_kOe", "Br_kGs", "Hk_kOe",
        "density_rel", "grain_um" (each optional)}; each observation is one new body (its own specimen effect).
        Shared latents are integrated numerically with k_latent samples per draw: new-preparation effects (shared by
        observations naming the same prep), the degas state per preparation (route["degas"] occurrence probability
        times route["degas_completeness"]) and H0's discrepancy direction. Mechanism weights are updated by each mechanism's marginal likelihood."""
        rng = np.random.default_rng(seed)
        logw = {m: np.zeros(len(v["sig_rho" if m.startswith("D") else "sig_spec"])) for m, v in self.post.items()}
        obs = [dict(o, route=self._route(o["route"])) for o in observations]
        dc_by_prep = {}
        for o in obs:
            key = json.dumps(o["route"]["degas_completeness"], sort_keys=True)
            if dc_by_prep.setdefault(o["route"]["prep"], key) != key:
                raise ValueError(f"observations sharing preparation {o['route']['prep']!r} must share degas_completeness")
        fd_by_prep = {}
        for o in obs:
            _check_prep_feedstock(fd_by_prep, o["route"], "observations")
        dens_obs = [o for o in obs if "density_rel" in o]
        mag_obs = [o for o in obs if any(k in o for k in ("Hci_kOe", "Br_kGs", "Hk_kOe", "grain_um"))]
        preps = sorted({o["route"]["prep"] for o in obs})
        K = k_latent
        if dens_obs:
            for m in self._mechs("density"):
                Kd = len(self.post[m]["lnk"])
                zs = {p: rng.standard_normal((Kd, K)) for p in preps}
                acc = np.zeros((Kd, K))
                p = {k: v[:, None] for k, v in self.post[m].items()}
                for o in dens_obs:
                    y = float(o["density_rel"])
                    y = y / RHO_TD if y > 1.5 else y
                    A = self._arrays(o["route"])
                    if o["route"]["prep"] in self.preps:
                        off = p["tau_prep"] * p[f"z_prep_{self.preps.index(o['route']['prep'])}"]
                        mu = self.fw.density(p, m, A, lnk_offset=off, xp=np)[:, 0]
                        acc += _t_logpdf(y, mu, p["sig_rho"][:, 0])[:, None]
                    else:
                        for k in range(K):
                            off = p["tau_prep"] * zs[o["route"]["prep"]][:, k:k + 1]
                            mu = self.fw.density(p, m, A, lnk_offset=off, xp=np)[:, 0]
                            acc[:, k] += _t_logpdf(y, mu, p["sig_rho"][:, 0])
                logw[m] += _lme(acc, 1)
        new_ltp = {}
        if mag_obs:
            for o in mag_obs:  # backlog item 2: no silent 0.95 TD default for an unknown density
                if o.get("rho_observed", o.get("density_rel")) is None:
                    raise ValueError("a magnetic observation needs the body's measured density: pass rho_observed "
                                     "(g/cm3 or relative) or density_rel")
            for m in self._mechs("magnetic"):
                Km = len(self.post[m]["sig_spec"])
                idx = np.arange(Km)
                lab = np.array([m] * Km)
                levels = LT_LEVELS if m in self.ltp else ("load",)
                acc = {L: np.zeros((Km, K)) for L in levels}
                zh = rng.standard_normal((Km, K)) if m == "H0" else np.zeros((Km, K))
                dz = {p: rng.random((Km, K)) for p in preps}
                nondefault = [o for o in mag_obs if not self._completeness_is_default(o["route"])
                              and o["route"]["degas_completeness"] != "complete"]
                dstate, deps = {}, {}
                if nondefault:
                    extra = rng.spawn(1)[0]
                    for pr in preps:
                        ro = next((o["route"] for o in mag_obs if o["route"]["prep"] == pr), None)
                        if ro is not None and not self._completeness_is_default(ro) \
                                and ro["degas_completeness"] != "complete":
                            dstate[pr] = self._sample_completeness(ro["degas_completeness"], (Km, K), extra)
                            deps[pr] = extra.uniform(0.0, 1.0, (Km, K))
                for o in mag_obs:
                    r = o["route"]
                    A = self._arrays(r)
                    rho_v = o.get("rho_observed", o.get("density_rel"))
                    rho_v = float(rho_v) / RHO_TD if float(rho_v) > 1.5 else float(rho_v)
                    rho = np.full(Km, rho_v)
                    shared = m == "H0" or 0.0 < float(r["degas"]) < 1.0 or r["prep"] in dstate
                    r = dict(r, _degas_expected=self._expected_state(r))
                    for k in (range(K) if shared else [0]):
                        degas_ind = (dz[r["prep"]][:, k] < float(r["degas"])).astype(float)
                        e = None
                        if r["prep"] in dstate:
                            degas_ind = degas_ind * dstate[r["prep"]][:, k]
                            e = deps[r["prep"]][:, k]
                        for L in levels:
                            ind = {"load_timing": np.full(Km, L)} if m in self.ltp else None
                            lat = self._magnetic_latent(A, r, lab, idx, rho, degas_ind, zh[:, k], eps=e, indicators=ind)
                            ll = np.zeros(Km)
                            if "Hci_kOe" in o:
                                ll += _n_logpdf(math.log(o["Hci_kOe"]), lat["lh"],
                                                np.sqrt(lat["sig_spec"] ** 2 + lat["sig_piece"] ** 2))
                            if "Br_kGs" in o:
                                ll += _n_logpdf(math.log(o["Br_kGs"]), lat["lbr"], lat["sig_Br"])
                            if "Hk_kOe" in o and "Hci_kOe" in o and 0 < o["Hk_kOe"] < o["Hci_kOe"]:
                                ratio = o["Hk_kOe"] / o["Hci_kOe"]
                                ll += _n_logpdf(math.log(ratio / (1 - ratio)), lat["sq"], lat["sig_sq"])
                            if "grain_um" in o:
                                ll += _n_logpdf(math.log(o["grain_um"]), lat["lgrain"], 0.15)
                            if shared:
                                acc[L][:, k] += ll
                            else:
                                acc[L] += ll[:, None]
                lik = np.stack([_lme(acc[L], 1) for L in levels], 1)          # (Km, n_levels) log p(y_new | theta, LT)
                if m in self.ltp:
                    lp = np.log(np.maximum(self.ltp[m], 1e-300)) + lik
                    tot = _lse(lp, 1)
                    new_ltp[m] = np.exp(lp - tot[:, None])
                    logw[m] += tot
                else:
                    logw[m] += lik[:, 0]
        new_dw, newW, ess = {}, {}, {}
        for kind in ("density", "magnetic"):
            lm = {}
            for m in self._mechs(kind):
                base = np.log(self.dw[m] / self.dw[m].sum())
                tot = _lse(base + logw[m])
                lm[m] = tot
                wv = np.exp(base + logw[m] - tot)
                new_dw[m] = wv
                ess[m] = float(1.0 / np.sum(wv ** 2))
            z = _lse(np.array([math.log(max(self.w[kind][m], 1e-300)) + lm[m] for m in self._mechs(kind)]))
            newW[kind] = {m: math.exp(math.log(max(self.w[kind][m], 1e-300)) + lm[m] - z) for m in self._mechs(kind)}
        ltp = {m: new_ltp.get(m, v) for m, v in self.ltp.items()}
        if ltp:
            tot = np.zeros(len(LT_LEVELS))
            for m in self._mechs("magnetic"):
                tot += newW["magnetic"][m] * (new_dw[m] @ ltp[m])
            newW["load_timing"] = {L: float(x) for L, x in zip(LT_LEVELS, tot / tot.sum())}
        for k in self.w:
            newW.setdefault(k, dict(self.w[k]))
        clone = object.__new__(PhysicsModel)
        clone.__dict__.update(self.__dict__)
        clone.dw, clone.w, clone.ltp = new_dw, newW, ltp
        info = {"mechanism_weights": newW, "ess_per_mechanism": ess,
                "warning": "ESS below ~200 for a mechanism that still carries weight means the stored draws no longer "
                           "represent its posterior; refit owner-side (NumPyro) instead. Shared latents of a new "
                           "preparation are integrated for the reweighting but redrawn in later predictions."}
        clone.conditioning = {"observations": observations, **info}
        return clone, info


_DEFAULT = None


def _default():
    global _DEFAULT
    if _DEFAULT is None:
        _DEFAULT = load()
    return _DEFAULT


def predict(route, **kw):
    return _default().predict(route, **kw)


def p_meets_spec(route, spec, **kw):
    return _default().p_meets_spec(route, spec, **kw)


def eig(candidates, **kw):
    return _default().eig(candidates, **kw)


def forecast_record(route, **kw):
    return _default().forecast_record(route, **kw)