Skip to main content

gam_problem/
row_metric.rs

1//! `RowMetric` — the single provenance-carrying per-row inner product shared by
2//! the SAE-manifold **likelihood** (residual whitening) and the **gauge**
3//! (isometry pullback weight).
4//!
5//! # Why this exists
6//!
7//! The SAE-manifold machine historically carried *two* independent inner
8//! products:
9//!
10//! * the **likelihood** measured reconstruction residuals isotropically — a
11//!   single scalar dispersion `φ̂ = RSS / residual-dof`, the data-fit loop
12//!   summing the bare `½ rᵀr`; there was no per-row metric at all; and
13//! * the **gauge** carried its own per-row metric in
14//!   `IsometryPenalty.weight: WeightField` — a low-rank `W_n = U_n U_nᵀ`
15//!   pullback `g_n = J_nᵀ W_n J_n`, settable independently of anything the
16//!   likelihood saw.
17//!
18//! Nothing structurally forced "the metric the likelihood whitens by" to equal
19//! "the metric the gauge pulls back through". That is exactly the
20//! objective↔gradient-desync bug class wearing geometry clothing: a
21//! likelihood-metric ≠ gauge-metric state was *representable*.
22//!
23//! `RowMetric` collapses the two into one object. The likelihood whitens
24//! through it; the gauge `WeightField` is *constructed from* it. A
25//! divergent-metric state is therefore unrepresentable — there is only one
26//! per-row factor stack `U_n`, with one [`MetricProvenance`] tag.
27//!
28//! # Magic-by-default selector
29//!
30//! There is no flag. The provenance is chosen by whether per-row Fisher factors
31//! exist:
32//!
33//! * no factors supplied ⇒ [`MetricProvenance::Euclidean`]; `W_n = I_p`;
34//!   whitening is the identity, so `φ̂` and the data-fit loop are
35//!   **bit-for-bit** the prior isotropic path; and
36//! * per-row Fisher factors supplied ⇒ [`MetricProvenance::OutputFisher`]; the
37//!   residual is whitened by `U_nᵀ` and the gauge pulls back through the same
38//!   `U_n`.
39//!
40//! # Validation
41//!
42//! Every metric block is constructed **through**
43//! [`crate::normalize_fisher_rao_blocks`], which
44//! broadcasts and eigenvalue-validates PSD-ness. `RowMetric` does not
45//! reimplement that validation; it materializes `W_n = U_n U_nᵀ` (which is PSD
46//! by construction) and runs it through the shared normalizer as the
47//! single point of truth for "is this a valid precision metric".
48//!
49//! Any rank floor used to make a block invertible for an internal solve is
50//! **solver-only** (mirroring `RidgePolicy::solver_only`, #747): it never enters
51//! the residual the objective sums, so `δ` cannot bias the criterion.
52//!
53//! # Rung 1 — the behavioral metric *in the reconstruction loss* (nats currency)
54//!
55//! [`MetricProvenance::OutputFisher`] installs the output-Fisher inner product
56//! as a **gauge** metric only: it whitens *nothing* (`whitens_likelihood()` is
57//! `false`), by deliberate #980 contract, so reconstruction stays the isotropic
58//! `½‖r‖²`. That answers "what coordinate is canonical", not "what does a
59//! reconstruction error *cost*".
60//!
61//! [`MetricProvenance::BehavioralFisher`] is the opposite deliberate choice:
62//! the **same** low-rank output-Fisher factors, but installed as the
63//! reconstruction *likelihood weight*. Plain MSE prices a reconstruction error
64//! `e = x − x̂` by its Euclidean size; the model, however, reads the activation
65//! only through the rest of the network, so the behavioral cost of `e` is the
66//! KL between the clean and corrupted next-token distributions,
67//! `KL ≈ ½ eᵀ G(x) e` with `G = JᵀFJ` the network-Jacobian pullback of the
68//! output Fisher `F` (units: **nats**). Minimizing `(x−x̂)ᵀ G (x−x̂)` instead of
69//! `‖x−x̂‖²` is **generalized least squares**: for a *fixed* per-row `G` it is
70//! still a linear Gaussian model in the coefficients, so the entire
71//! REML/evidence/EDF/certificate stack survives verbatim — this is why the
72//! metric rides the identical `whitens_likelihood()` plumbing the
73//! [`MetricProvenance::WhitenedStructured`] noise model uses, and why the G=I
74//! limit reproduces the plain-MSE fit bit-for-bit (see the module tests).
75//!
76//! This is the principled form of Braun's end-to-end **KL + MSE** objective.
77//! Anchoring to the activation keeps it *reconstruction* (it does not collapse
78//! to "match the logits by any means" — the decoder still has to reproduce `x`),
79//! while pricing the residual in nats through `G`. The payoff is automatic
80//! selection for *mattering*: `G`'s null directions — activation structure the
81//! rest of the network cannot read — are penalized nothing, because
82//! `eᵀ G e = 0` there. MSE in a behaviorally-inert direction goes free, which is
83//! the correct behavior, not a bug: nothing downstream changes, so nothing
84//! should be paid.
85//!
86//! **The d×d `G` is never materialized.** `G` is sketched by `s` random probes,
87//! `vᵢ = Jᵀ F^{1/2} uᵢ` (`uᵢ` iid, `s ≈ 4…16`), computed by `s` backward passes
88//! per token at *harvest* time (the model-interaction boundary) and stored as
89//! the columns of the per-row factor `U_n = [v₁ … v_s] ∈ ℝ^{p×s}`. Then
90//! `G ≈ Σᵢ vᵢ vᵢᵀ = U_n U_nᵀ` and the criterion-facing
91//! `eᵀ G e ≈ Σᵢ (vᵢᵀ e)² = ‖U_nᵀ e‖²` is exactly what
92//! [`RowMetric::quad_form`] / [`RowMetric::whiten_residual_row`] already
93//! compute — zero train-time model cost, `O(p·s)` per row. See
94//! [`RowMetric::behavioral_fisher`] and the probe-packing helper
95//! [`pack_probe_factors`].
96
97use ndarray::{Array2, Array3, ArrayView1};
98use std::sync::Arc;
99
100use crate::normalize_fisher_rao_blocks;
101
102/// Per-observation behavioral-metric field `W_n ∈ ℝ^{p × p}`, stored in
103/// **low-rank factored form** `W_n = U_n U_n^T` with `U_n ∈ ℝ^{p × r_n}`.
104///
105/// The canonical coordinate is the one where one unit of motion in `t` is one
106/// unit of behavioral change in the output space, so the `W_n` weighting is
107/// load-bearing: the pullback metric is `g_n = J_n^T W_n J_n`. Storing as
108/// `U_n` lets every contraction in this module run in
109/// `(J^T U_n)(U_n^T J)` order, which is `O(p · r · d + r · d²)` per row — we
110/// **never** materialize the `p × p` `W_n`, which is essential when `p`
111/// (number of observation channels) is large but rank is small (e.g. one or
112/// two behavioral dimensions per latent observation).
113///
114/// `Identity` is the gauge-fix default and corresponds to `U_n = I_p` so the
115/// pullback reduces to the standard `J_n^T J_n`. `Factored` stores the
116/// per-row `U_n` blocks contiguously: every row's factor is `p × rank`, and
117/// rows may share the same rank (uniform-rank case) or vary if the field is
118/// data-driven. For the uniform-rank case the storage is
119/// `(n_obs, p * rank)` row-major.
120#[derive(Clone)]
121pub enum WeightField {
122    /// `W_n = I_p` for every `n`. Reduces to the bare pullback `J^T J`.
123    Identity,
124    /// Per-row low-rank factor `U_n ∈ ℝ^{p × rank}`. Storage layout: a
125    /// `(n_obs, p * rank)` row-major matrix where row `n` packs `U_n` in
126    /// column-major-within-row order `U_n[i, k] = u[n, i * rank + k]`.
127    Factored {
128        u: Arc<Array2<f64>>,
129        rank: usize,
130        p_out: usize,
131    },
132}
133
134impl std::fmt::Debug for WeightField {
135    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
136        match self {
137            WeightField::Identity => f.write_str("Identity"),
138            WeightField::Factored { u, rank, p_out } => f
139                .debug_struct("Factored")
140                .field("shape", &format_args!("{}×{}", u.nrows(), u.ncols()))
141                .field("rank", rank)
142                .field("p_out", p_out)
143                .finish(),
144        }
145    }
146}
147
148impl WeightField {
149    /// Apply `U_n^T J_n` for a specific row, given both the row's `J_n` flat
150    /// `(p * d)` slice and the row's `U_n` flat `(p * rank)` slice. Returns
151    /// the `(rank × d)` matrix and its row count.
152    pub fn project_jac_row_with_u(
153        u_row: &[f64],
154        jac_row: &[f64],
155        p: usize,
156        rank: usize,
157        d: usize,
158    ) -> Array2<f64> {
159        // M[k, a] = Σ_i U[i, k] · J[i, a].
160        let mut m = Array2::<f64>::zeros((rank, d));
161        for k in 0..rank {
162            for a in 0..d {
163                let mut s = 0.0;
164                for i in 0..p {
165                    s += u_row[i * rank + k] * jac_row[i * d + a];
166                }
167                m[[k, a]] = s;
168            }
169        }
170        m
171    }
172}
173
174/// Where the per-row metric came from — the provenance that makes
175/// "likelihood-metric ≠ gauge-metric" diagnosable instead of silent.
176///
177/// Object 4 (the gauge object) reads this to certify which inner product the
178/// fit actually used; #974 fills [`MetricProvenance::WhitenedStructured`] with a
179/// factor-analytic residual-covariance whitening.
180#[derive(Clone, Copy, PartialEq, Eq, Debug)]
181pub enum MetricProvenance {
182    /// `M_n = I_p` for every row. The likelihood is isotropic and the gauge
183    /// pullback reduces to the bare `J_nᵀ J_n`. This is the default and is
184    /// bit-for-bit the historical isotropic-`φ̂` path.
185    Euclidean,
186    /// `M_n = U_n U_nᵀ (+ solver-only δI)` from supplied per-row output-Fisher
187    /// factors `U_n ∈ ℝ^{p × rank}`. The canonical "one unit of latent motion ↦
188    /// one unit of behavioral change" metric: residuals are whitened in the
189    /// output-Fisher inner product and the gauge pulls back through the same
190    /// factors. The `rank` is carried in the provenance so a consumer (Object 4)
191    /// can certify the factor rank that produced the inner product.
192    OutputFisher { rank: usize },
193    /// `M_n = U_n U_nᵀ` from per-row output-Fisher factors that aggregate the
194    /// **downstream** influence of position `n` over future positions through
195    /// the KV path, rather than the same-position logits of
196    /// [`MetricProvenance::OutputFisher`] (#980, mechanism 2).
197    ///
198    /// The same-position pullback `∂logits_t/∂x_t` can be ≈ 0 for a feature
199    /// whose entire causal effect lands many tokens later (information carried
200    /// forward through attention); a gauge built on it is blind to exactly that
201    /// content. This provenance is the forward-looking alternative: each row's
202    /// factor `U_n` is the top-`rank` factorization of the aggregated output
203    /// Fisher `Σ_{t ≥ n} (∂logits_t/∂x_n)ᵀ F_t (∂logits_t/∂x_n)` over future
204    /// positions the residual stream at `n` reaches. It is provenance-generic:
205    /// it whitens nothing ([`Self::whitens_likelihood`] is `false`, like
206    /// [`MetricProvenance::OutputFisher`]) and drives the gauge / lens /
207    /// enrichment unchanged ([`Self::is_output_fisher_like`]). The lens/gauge
208    /// machinery consumes it identically; only the *scientific* reading
209    /// changes — dormant-feature detection becomes forward-looking (a feature
210    /// driving far-future tokens now registers behavioral coupling that the
211    /// same-position metric reported as ≈ 0).
212    OutputFisherDownstream { rank: usize },
213    /// **Rung 1** — the output-Fisher metric installed as the reconstruction
214    /// **likelihood weight** (generalized least squares in nats), not merely as
215    /// a gauge. `M_n = U_n U_nᵀ ≈ G_n = J_nᵀ F_n J_n` is the `s`-probe sketch of
216    /// the pulled-back output Fisher, with `U_n = [v₁ … v_s]`,
217    /// `vᵢ = J_nᵀ F_n^{1/2} uᵢ`, and `probes = s` the number of random probes
218    /// (the factor rank).
219    ///
220    /// This is the *only* [`RowMetric::is_output_fisher_like`]-adjacent
221    /// provenance for which [`RowMetric::whitens_likelihood`] is `true`: the
222    /// data-fit sums `½ eᵀ G_n e = ½ ‖U_nᵀ e‖²` (nats) instead of `½‖e‖²`. It is
223    /// distinct from [`Self::OutputFisher`] precisely because the choice to let
224    /// the metric enter the *loss* (rather than only the gauge) is deliberate and
225    /// must not be silently inherited by the #980 gauge / two-tier-harvest
226    /// contract — that contract relies on [`Self::OutputFisher`] whitening
227    /// nothing. Because `G_n` is a *fixed* per-row metric, the whitened problem
228    /// is again linear-Gaussian in the coefficients, so REML/evidence/EDF are
229    /// unchanged (the GLS-preserves-REML property, verified in the module tests
230    /// against the `G=I` plain-MSE limit).
231    BehavioralFisher { probes: usize },
232    /// Structured-residual whitening: `M_n = Σ_n^{-1}` from the **estimated**
233    /// factor-analytic residual covariance `Σ_n = Λ c(z_n) Λᵀ + D` (#974), with
234    /// `factor_rank` the selected factor count. Produced by
235    /// Structured-residual producers materialize this provenance when they fit
236    /// a residual-covariance whitening model;
237    /// the only provenance for which
238    /// [`whitens_likelihood`](RowMetric::whitens_likelihood) is `true`. It
239    /// carries the same low-rank factor layout as
240    /// [`MetricProvenance::OutputFisher`].
241    WhitenedStructured { factor_rank: usize },
242}
243
244/// The single per-row metric object. Holds one low-rank factor stack `U_n` (or
245/// none, for Euclidean) plus the validated PSD blocks, tagged with its
246/// [`MetricProvenance`].
247///
248/// `p` is the output dimensionality (residual / Jacobian-column dimension); the
249/// per-row factor `U_n ∈ ℝ^{p × rank}` so `W_n = U_n U_nᵀ ∈ ℝ^{p × p}` without
250/// ever being materialized as `p × p` in any hot path.
251#[derive(Clone, Debug)]
252pub struct RowMetric {
253    provenance: MetricProvenance,
254    n_rows: usize,
255    p: usize,
256    rank: usize,
257    /// `(n_rows, p * rank)` row-major: `U_n[i, k] = u[n, i * rank + k]`. `None`
258    /// for [`MetricProvenance::Euclidean`] (the identity factor is implicit).
259    factors: Option<Arc<Array2<f64>>>,
260    /// **Solver-only** Tikhonov floor `δ` added as `δ I_p` to make a
261    /// rank-deficient `U_n U_nᵀ` invertible for an *internal solve only*.
262    ///
263    /// Invariant (mirrors `RidgePolicy::solver_only`, #747): `δ` **never** enters
264    /// any quantity that feeds the evidence criterion. The criterion-facing
265    /// quad-form / whitening / fisher-mass methods all use the *un-floored*
266    /// `U_n U_nᵀ`; only [`Self::solve_floor`]-tagged solver helpers see `δ`. A
267    /// nonzero floor therefore cannot bias the objective the optimizer reports.
268    solver_delta: f64,
269    /// Per-row traces `tr(M_n)` of the criterion-facing (un-floored) metric.
270    ///
271    /// This is the only dense-block reduction any consumer reads (the #980
272    /// Fisher-mass row measure); the `(n_rows, p, p)` block stack itself is
273    /// validated **streamingly** at construction through
274    /// [`normalize_fisher_rao_blocks`] one row at a time and then dropped.
275    /// Retaining it was `n·p²·8` bytes — 13 GiB at `(n=2000, p=896)` and an
276    /// OOM at LLM-scale `p` — for a record nothing ever re-read. The solver
277    /// `δ` is deliberately *not* baked in here, so this is the
278    /// criterion-facing trace.
279    traces: ndarray::Array1<f64>,
280}
281
282impl RowMetric {
283    /// Euclidean metric: `W_n = I_p` for all `n`. Whitening is the identity, so
284    /// the likelihood residual path is bit-for-bit the prior isotropic `φ̂`.
285    ///
286    /// Constructed directly: the identity stack is PSD axiomatically, so
287    /// routing it through the dense normalizer would materialize and
288    /// spectrum-check `n` identity blocks (`n·p²` memory, `n·p³` flops) to
289    /// validate a tautology. `tr(I_p) = p` per row.
290    pub fn euclidean(n_rows: usize, p: usize) -> Result<Self, String> {
291        Ok(Self {
292            provenance: MetricProvenance::Euclidean,
293            n_rows,
294            p,
295            rank: p,
296            factors: None,
297            solver_delta: 0.0,
298            traces: ndarray::Array1::<f64>::from_elem(n_rows, p as f64),
299        })
300    }
301
302    /// Output-Fisher metric: per-row low-rank factors `U_n ∈ ℝ^{p × rank}`
303    /// supplied as a `(n_rows, p * rank)` row-major matrix (`U_n[i, k] =
304    /// u[n, i * rank + k]`). The induced `M_n = U_n U_nᵀ` is PSD by
305    /// construction; it is validated through [`normalize_fisher_rao_blocks`] so
306    /// the validation path is shared. No solver floor (`δ = 0`).
307    pub fn output_fisher(u: Arc<Array2<f64>>, p: usize, rank: usize) -> Result<Self, String> {
308        Self::from_factors(MetricProvenance::OutputFisher { rank }, u, p, rank, 0.0)
309    }
310
311    /// Downstream-influence output-Fisher metric: per-row factors `U_n ∈
312    /// ℝ^{p × rank}` whose `M_n = U_n U_nᵀ` is the aggregated output Fisher of
313    /// position `n` over the **future** positions it reaches through the KV path
314    /// ([`MetricProvenance::OutputFisherDownstream`], #980 mechanism 2). The
315    /// factor layout is identical to [`Self::output_fisher`]; only the
316    /// provenance tag (and hence the scientific reading) differs. Whitens
317    /// nothing, drives the gauge / lens / enrichment exactly as the
318    /// same-position metric does — the consuming machinery is provenance-generic
319    /// (see [`Self::is_output_fisher_like`]).
320    pub fn output_fisher_downstream(
321        u: Arc<Array2<f64>>,
322        p: usize,
323        rank: usize,
324    ) -> Result<Self, String> {
325        Self::from_factors(
326            MetricProvenance::OutputFisherDownstream { rank },
327            u,
328            p,
329            rank,
330            0.0,
331        )
332    }
333
334    /// **Rung 1** — the output-Fisher metric as a reconstruction *likelihood
335    /// weight* (GLS in nats): per-row `s`-probe factors `U_n ∈ ℝ^{p × probes}`
336    /// supplied as a `(n_rows, p * probes)` row-major matrix
337    /// (`U_n[i, k] = u[n, i * probes + k]`), so that column `k` is the probe
338    /// vector `v_k = J_nᵀ F_n^{1/2} u_k` and `M_n = U_n U_nᵀ ≈ G_n`. Unlike
339    /// [`Self::output_fisher`], the resulting metric returns
340    /// `whitens_likelihood() == true`: the data-fit prices reconstruction error
341    /// as `½ eᵀ G_n e`. Validated through [`normalize_fisher_rao_blocks`] like
342    /// every factored metric; no solver floor (`δ = 0`).
343    ///
344    /// See [`pack_probe_factors`] to build `u` from a natural `(n, p, s)` probe
345    /// stack emitted at harvest time.
346    pub fn behavioral_fisher(u: Arc<Array2<f64>>, p: usize, probes: usize) -> Result<Self, String> {
347        Self::from_factors(
348            MetricProvenance::BehavioralFisher { probes },
349            u,
350            p,
351            probes,
352            0.0,
353        )
354    }
355
356    /// Like [`Self::output_fisher`] but with a **solver-only** Tikhonov floor
357    /// `δ ≥ 0`. The floor is recorded for solver helpers only; every
358    /// criterion-facing method (`quad_form`, `whiten_residual`, `fisher_mass`)
359    /// ignores it (#747 discipline), so the evidence criterion is `δ`-free.
360    pub fn output_fisher_with_solver_floor(
361        u: Arc<Array2<f64>>,
362        p: usize,
363        rank: usize,
364        solver_delta: f64,
365    ) -> Result<Self, String> {
366        if !(solver_delta.is_finite() && solver_delta >= 0.0) {
367            return Err(format!(
368                "RowMetric::output_fisher_with_solver_floor: solver_delta must be finite and \
369                 non-negative; got {solver_delta}"
370            ));
371        }
372        Self::from_factors(
373            MetricProvenance::OutputFisher { rank },
374            u,
375            p,
376            rank,
377            solver_delta,
378        )
379    }
380
381    /// Structured-residual whitening from supplied per-row precision factors.
382    ///
383    /// `u` carries the per-row factor stack `U_n ∈ ℝ^{p × rank}` (row-major flat)
384    /// with `U_n U_nᵀ = M_n = Σ_n^{-1}` — the precision of the **estimated**
385    /// residual-covariance noise model. This is the low-level constructor; #974
386    /// producers that *fit* `Σ_n` (a low-rank factor + diagonal + smooth
387    /// activity-scale) assemble these factors and call through here. Because the
388    /// provenance is
389    /// [`MetricProvenance::WhitenedStructured`], [`Self::whitens_likelihood`] is
390    /// `true`: a metric built this way is the first that whitens the likelihood.
391    pub fn whitened_structured(u: Arc<Array2<f64>>, p: usize, rank: usize) -> Result<Self, String> {
392        Self::from_factors(
393            MetricProvenance::WhitenedStructured { factor_rank: rank },
394            u,
395            p,
396            rank,
397            0.0,
398        )
399    }
400
401    fn from_factors(
402        provenance: MetricProvenance,
403        u: Arc<Array2<f64>>,
404        p: usize,
405        rank: usize,
406        solver_delta: f64,
407    ) -> Result<Self, String> {
408        let n_rows = u.nrows();
409        if u.ncols() != p * rank {
410            return Err(format!(
411                "RowMetric::from_factors: factor matrix has {} cols; expected p*rank = {}*{} = {}",
412                u.ncols(),
413                p,
414                rank,
415                p * rank
416            ));
417        }
418        if !u.iter().all(|v| v.is_finite()) {
419            return Err("RowMetric::from_factors: factors must be finite".to_string());
420        }
421        // Materialize W_n = U_n U_nᵀ one row at a time (PSD by construction),
422        // validate each through the single shared normalizer rather than
423        // reimplementing the PSD check, record its trace, and drop the block.
424        // Streaming keeps construction O(p²) memory; the former whole-stack
425        // materialization retained `n·p²` doubles nothing ever re-read.
426        let mut traces = ndarray::Array1::<f64>::zeros(n_rows);
427        let mut full = Array3::<f64>::zeros((1, p, p));
428        for row in 0..n_rows {
429            for i in 0..p {
430                for j in 0..p {
431                    let mut acc = 0.0;
432                    for k in 0..rank {
433                        acc += u[[row, i * rank + k]] * u[[row, j * rank + k]];
434                    }
435                    full[[0, i, j]] = acc;
436                }
437            }
438            normalize_fisher_rao_blocks(full.view().into_dyn(), 1, p)
439                .map_err(|e| format!("RowMetric::from_factors: row {row}: {e}"))?;
440            let mut tr = 0.0_f64;
441            for i in 0..p {
442                tr += full[[0, i, i]];
443            }
444            traces[row] = tr;
445        }
446        Ok(Self {
447            provenance,
448            n_rows,
449            p,
450            rank,
451            factors: Some(u),
452            solver_delta,
453            traces,
454        })
455    }
456
457    /// Restrict the metric to the rows `rows` (an index subset or permutation),
458    /// preserving provenance, `p`, `rank`, and the solver floor. The
459    /// outer-criterion row subsample uses this to whiten the subsampled fit
460    /// through the SAME per-row metric the full-`N` fit uses, so the ρ search
461    /// ranks the delivered criterion (e.g. a #974 structured-whitening fit is not
462    /// silently searched unwhitened). Each gathered row's factor block is copied
463    /// verbatim, so the induced `M_n = U_n U_nᵀ` is bit-identical to the full
464    /// metric's on every selected row.
465    pub fn gather_rows(&self, rows: &[usize]) -> Result<Self, String> {
466        for (pos, &r) in rows.iter().enumerate() {
467            if r >= self.n_rows {
468                return Err(format!(
469                    "RowMetric::gather_rows: row index {r} at position {pos} is out of bounds \
470                     (n_rows = {})",
471                    self.n_rows
472                ));
473            }
474        }
475        match self.factors.as_ref() {
476            // Euclidean carries an implicit identity factor per row, so the subset
477            // is just a smaller identity stack — no factor storage to gather.
478            None => Self::euclidean(rows.len(), self.p),
479            Some(factors) => {
480                let cols = self.p * self.rank;
481                let mut sub = Array2::<f64>::zeros((rows.len(), cols));
482                for (pos, &r) in rows.iter().enumerate() {
483                    sub.row_mut(pos).assign(&factors.row(r));
484                }
485                // Re-runs the shared PSD normalizer on the subset (a subset of
486                // valid rows stays valid) and preserves the exact provenance and
487                // solver floor.
488                Self::from_factors(
489                    self.provenance,
490                    Arc::new(sub),
491                    self.p,
492                    self.rank,
493                    self.solver_delta,
494                )
495            }
496        }
497    }
498
499    /// The provenance tag (consumed by Object 4 to certify the inner product).
500    pub fn provenance(&self) -> MetricProvenance {
501        self.provenance
502    }
503
504    /// Whether this metric is allowed to **whiten the likelihood** (i.e. replace
505    /// the isotropic reconstruction data-fit `½ rᵀr` with the whitened
506    /// `½ rᵀ M_n r`).
507    ///
508    /// This is TRUE for two provenances, for two distinct reasons:
509    ///
510    /// * [`MetricProvenance::WhitenedStructured`] — a genuinely *estimated noise
511    ///   model* (a factor-analytic residual covariance, #974), for which
512    ///   whitening the likelihood is the statistically correct thing to do; and
513    /// * [`MetricProvenance::BehavioralFisher`] — the **Rung 1** deliberate
514    ///   choice to price reconstruction error in nats: the output-Fisher metric
515    ///   `G_n` installed *as the loss weight* (`½ eᵀ G_n e`), a generalized
516    ///   least-squares reconstruction. Because `G_n` is a fixed per-row metric
517    ///   the problem stays linear-Gaussian, so REML/evidence/EDF are preserved.
518    ///
519    /// It is FALSE for [`MetricProvenance::Euclidean`] (nothing to whiten by) and
520    /// for the *gauge-only* [`MetricProvenance::OutputFisher`] /
521    /// [`MetricProvenance::OutputFisherDownstream`]: there the output-Fisher
522    /// inner product is an **output-geometry gauge**, and whitening the
523    /// likelihood by it *implicitly* (without the caller electing GLS) would
524    /// silently replace the reconstruction loss with a Fisher pullback — the #980
525    /// failure mode, and the reason the two-tier harvest can withhold factors
526    /// from a row without changing its loss. `BehavioralFisher` is the *explicit*
527    /// election of that same arithmetic as the intended objective.
528    pub fn whitens_likelihood(&self) -> bool {
529        matches!(
530            self.provenance,
531            MetricProvenance::WhitenedStructured { .. } | MetricProvenance::BehavioralFisher { .. }
532        )
533    }
534
535    /// Whether this metric **drives the gauge** — i.e. the isometry-penalty
536    /// pullback weight is taken from it rather than the identity.
537    ///
538    /// TRUE for any non-[`MetricProvenance::Euclidean`] provenance: both
539    /// [`MetricProvenance::OutputFisher`] and
540    /// [`MetricProvenance::WhitenedStructured`] supply a non-identity per-row
541    /// inner product the gauge pulls back through. Euclidean reduces the gauge
542    /// pullback to the bare `J_nᵀ J_n`, so it does not drive the gauge.
543    pub fn drives_gauge(&self) -> bool {
544        !matches!(self.provenance, MetricProvenance::Euclidean)
545    }
546
547    /// Whether this metric is an **output-Fisher gauge** — either the
548    /// same-position [`MetricProvenance::OutputFisher`] or the downstream
549    /// [`MetricProvenance::OutputFisherDownstream`] (#980). The two share every
550    /// consumer behavior (Sym(F) separation under the gauge, two-lens coupling,
551    /// steering geometry, enrichment); they differ only in the *scientific*
552    /// reading of what behavioral coupling means (same-position vs
553    /// forward-looking). Consumers that gate on "is this an output-Fisher
554    /// pullback" should use this predicate rather than matching one variant, so
555    /// the downstream metric rides the identical path.
556    pub fn is_output_fisher_like(&self) -> bool {
557        matches!(
558            self.provenance,
559            MetricProvenance::OutputFisher { .. } | MetricProvenance::OutputFisherDownstream { .. }
560        )
561    }
562
563    /// Number of rows the metric is defined over.
564    pub fn n_rows(&self) -> usize {
565        self.n_rows
566    }
567
568    /// Output dimensionality `p` (residual / Jacobian-column dimension).
569    pub fn p_out(&self) -> usize {
570        self.p
571    }
572
573    /// The factor rank: the dimension of the whitened residual
574    /// [`Self::whiten_residual_row`] returns (and the column count of the per-row
575    /// factor `U_n ∈ ℝ^{p × rank}`). For [`MetricProvenance::Euclidean`] this is
576    /// `p` (the implicit identity factor), so a consumer that sizes a whitened
577    /// buffer by `metric_rank()` gets the right length in every provenance.
578    pub fn metric_rank(&self) -> usize {
579        self.rank
580    }
581
582    /// Per-row traces `tr(M_n)` of the criterion-facing (un-floored) metric —
583    /// the Fisher-mass reduction the #980 row measure consumes. The dense
584    /// `(n_rows, p, p)` stack is validated streamingly at construction and
585    /// never retained; consumers wanting an explicit `W_n` rebuild it from
586    /// [`Self::metric_rank`]-sized factors.
587    pub fn row_traces(&self) -> ndarray::ArrayView1<'_, f64> {
588        self.traces.view()
589    }
590
591    /// Whiten a single `p`-dimensional residual row `r` into the coordinates
592    /// whose squared Euclidean norm equals `rᵀ W_n r`.
593    ///
594    /// * Euclidean: returns `r` unchanged (`‖r‖² = rᵀ I r`), so the likelihood
595    ///   reproduces the isotropic `½ rᵀr` data-fit bit-for-bit.
596    /// * Factored: returns `U_nᵀ r ∈ ℝ^{rank}`, with
597    ///   `‖U_nᵀ r‖² = rᵀ U_n U_nᵀ r = rᵀ W_n r`.
598    ///
599    /// This is the load-bearing identity that lets the data-fit loop sum
600    /// `0.5 * Σ whitened²` and recover exactly `rᵀ W_n r` whatever the
601    /// provenance.
602    pub fn whiten_residual_row(&self, row: usize, r: ArrayView1<'_, f64>) -> Vec<f64> {
603        match &self.factors {
604            None => r.iter().copied().collect(),
605            Some(u) => {
606                let mut out = vec![0.0_f64; self.rank];
607                for k in 0..self.rank {
608                    let mut acc = 0.0;
609                    for i in 0..self.p {
610                        acc += u[[row, i * self.rank + k]] * r[i];
611                    }
612                    out[k] = acc;
613                }
614                out
615            }
616        }
617    }
618
619    /// The factor entry `U_n[i, k]` for one row (`i ∈ [0, p)`, `k ∈ [0, rank)`).
620    /// For [`MetricProvenance::Euclidean`] the implicit factor is `I_p`, so this
621    /// returns `1.0` when `i == k` and `0.0` otherwise — letting a consumer that
622    /// whitens a Jacobian via `factor_entry` produce the identity whitening
623    /// without a provenance branch. Reads the **un-floored** factors (criterion
624    /// face, #747).
625    #[inline]
626    pub fn factor_entry(&self, row: usize, i: usize, k: usize) -> f64 {
627        match &self.factors {
628            None => {
629                if i == k {
630                    1.0
631                } else {
632                    0.0
633                }
634            }
635            Some(u) => u[[row, i * self.rank + k]],
636        }
637    }
638
639    /// Apply the full per-row metric `M_n x = U_n (U_nᵀ x) ∈ ℝ^p` for one
640    /// `p`-vector `x`, formed factored (`rank` flops in, `p` flops out) — never
641    /// materializing `M_n` as `p × p`. Euclidean returns `x` unchanged
642    /// (`M_n = I_p`). This is the p-space metric-applied vector the SAE β-tier
643    /// data-fit gradient contracts (β lives in p-output space, so its gradient
644    /// needs `M_n r_n`, not the rank-space whitened residual `U_nᵀ r_n`). Uses the
645    /// **un-floored** factors (criterion face, `δ`-free, #747 invariant).
646    pub fn apply_metric_row(&self, row: usize, x: ArrayView1<'_, f64>) -> Vec<f64> {
647        match &self.factors {
648            None => x.iter().copied().collect(),
649            Some(u) => {
650                // w = U_nᵀ x ∈ ℝ^{rank}.
651                let mut w = vec![0.0_f64; self.rank];
652                for k in 0..self.rank {
653                    let mut acc = 0.0;
654                    for i in 0..self.p {
655                        acc += u[[row, i * self.rank + k]] * x[i];
656                    }
657                    w[k] = acc;
658                }
659                // out = U_n w ∈ ℝ^p.
660                let mut out = vec![0.0_f64; self.p];
661                for i in 0..self.p {
662                    let mut acc = 0.0;
663                    for k in 0..self.rank {
664                        acc += u[[row, i * self.rank + k]] * w[k];
665                    }
666                    out[i] = acc;
667                }
668                out
669            }
670        }
671    }
672
673    /// Pullback metric `g_n = J_nᵀ W_n J_n` for one row, formed as
674    /// `(J_nᵀ U_n)(U_nᵀ J_n)` — never materializing the `p × p` `W_n`.
675    ///
676    /// `j_row` is the row's Jacobian `J_n ∈ ℝ^{p × d}` flattened row-major
677    /// (`J_n[i, a] = j_row[i * d + a]`). Returns the `d × d` `g_n`.
678    pub fn pullback(&self, row: usize, j_row: &[f64], d: usize) -> Array2<f64> {
679        match &self.factors {
680            None => {
681                // W_n = I_p ⇒ g_n = J_nᵀ J_n.
682                let mut g = Array2::<f64>::zeros((d, d));
683                for a in 0..d {
684                    for b in a..d {
685                        let mut acc = 0.0;
686                        for i in 0..self.p {
687                            acc += j_row[i * d + a] * j_row[i * d + b];
688                        }
689                        g[[a, b]] = acc;
690                        g[[b, a]] = acc;
691                    }
692                }
693                g
694            }
695            Some(u) => {
696                // M_n = U_nᵀ J_n ∈ ℝ^{rank × d}; g_n = M_nᵀ M_n.
697                let mut m = Array2::<f64>::zeros((self.rank, d));
698                for k in 0..self.rank {
699                    for a in 0..d {
700                        let mut acc = 0.0;
701                        for i in 0..self.p {
702                            acc += u[[row, i * self.rank + k]] * j_row[i * d + a];
703                        }
704                        m[[k, a]] = acc;
705                    }
706                }
707                let mut g = Array2::<f64>::zeros((d, d));
708                for a in 0..d {
709                    for b in a..d {
710                        let mut acc = 0.0;
711                        for k in 0..self.rank {
712                            acc += m[[k, a]] * m[[k, b]];
713                        }
714                        g[[a, b]] = acc;
715                        g[[b, a]] = acc;
716                    }
717                }
718                g
719            }
720        }
721    }
722
723    /// Quadratic form `r_nᵀ M_n r_n` for one row's residual `r_n ∈ ℝ^p`, formed
724    /// **factored** as `‖U_nᵀ r_n‖²` — never materializing the `p × p` `M_n`.
725    ///
726    /// This is the criterion-facing squared residual the likelihood sums; it uses
727    /// the **un-floored** `U_n U_nᵀ`, so the solver `δ` does not enter it
728    /// (#747 invariant). Euclidean provenance returns the bit-identical `‖r_n‖²`.
729    #[inline]
730    pub fn quad_form(&self, row: usize, r: ArrayView1<'_, f64>) -> f64 {
731        match &self.factors {
732            None => r.iter().map(|&v| v * v).sum(),
733            Some(_) => self
734                .whiten_residual_row(row, r)
735                .iter()
736                .map(|&w| w * w)
737                .sum(),
738        }
739    }
740
741    /// Whiten a per-row Jacobian `J_n ∈ ℝ^{p × d}` (row-major flat,
742    /// `J_n[i, a] = j_row[i * d + a]`) into `M_n = U_nᵀ J_n ∈ ℝ^{rank × d}` so
743    /// that `M_nᵀ M_n = J_nᵀ (U_n U_nᵀ) J_n = J_nᵀ W_n J_n` is the pullback
744    /// **without** any `p × p` intermediate. Euclidean returns `J_n` reshaped to
745    /// `(p, d)` (the identity whitening). Solver `δ` is not applied (criterion
746    /// face).
747    pub fn whiten_jacobian(&self, row: usize, j_row: &[f64], d: usize) -> Array2<f64> {
748        match &self.factors {
749            None => {
750                let mut out = Array2::<f64>::zeros((self.p, d));
751                for i in 0..self.p {
752                    for a in 0..d {
753                        out[[i, a]] = j_row[i * d + a];
754                    }
755                }
756                out
757            }
758            Some(u) => {
759                let mut m = Array2::<f64>::zeros((self.rank, d));
760                for k in 0..self.rank {
761                    for a in 0..d {
762                        let mut acc = 0.0;
763                        for i in 0..self.p {
764                            acc += u[[row, i * self.rank + k]] * j_row[i * d + a];
765                        }
766                        m[[k, a]] = acc;
767                    }
768                }
769                m
770            }
771        }
772    }
773
774    /// Fisher mass of a per-row output vector `x_n ∈ ℝ^p`: the scalar
775    /// `x_nᵀ M_n x_n` (alias of [`Self::quad_form`] read as an information mass
776    /// rather than a residual square). Factored, never `p × p`, `δ`-free.
777    #[inline]
778    pub fn fisher_mass(&self, row: usize, x: ArrayView1<'_, f64>) -> f64 {
779        self.quad_form(row, x)
780    }
781
782    /// The **solver-only** Tikhonov floor `δ` (#747). Returned for internal
783    /// solver helpers that need `U_n U_nᵀ + δ I` to be invertible; by contract
784    /// no caller may fold this into a criterion-facing quantity. Always `0` for
785    /// Euclidean and for factored metrics built without an explicit floor.
786    pub fn solver_floor(&self) -> f64 {
787        self.solver_delta
788    }
789
790    /// The gauge view of this metric: the
791    /// [`crate::WeightField`] the isometry penalty pulls back through.
792    ///
793    /// This is the **single** way an `IsometryPenalty` acquires a non-identity
794    /// gauge metric — the independent `WeightField` setter has been removed — so
795    /// the gauge metric is, by construction, the same object the likelihood
796    /// whitens with.
797    pub fn to_weight_field(&self) -> crate::WeightField {
798        use crate::WeightField;
799        match &self.factors {
800            None => WeightField::Identity,
801            Some(u) => WeightField::Factored {
802                u: Arc::clone(u),
803                rank: self.rank,
804                p_out: self.p,
805            },
806        }
807    }
808}
809
810/// Pack a harvest-emitted probe stack into the row-major factor layout
811/// [`RowMetric::behavioral_fisher`] expects.
812///
813/// The harvest boundary (the model-interaction side) emits, per token, `s`
814/// probe vectors `vₖ = J_nᵀ F_n^{1/2} uₖ ∈ ℝ^p` — the natural shape is
815/// `probes[n, i, k] = (vₖ)ᵢ`, an `(n_rows, p, probes)` stack. This assembles the
816/// `(n_rows, p · probes)` row-major matrix `u[n, i·probes + k] = probes[n, i, k]`
817/// that the constructor consumes so that column `k` of the per-row factor `U_n`
818/// is exactly probe `vₖ` and `M_n = U_n U_nᵀ = Σₖ vₖ vₖᵀ ≈ G_n`.
819///
820/// This is a pure repack of the standard C-order flattening; it exists so the
821/// harvest → metric seam is a single named, validated Rust surface rather than
822/// an ad-hoc reshape at each call site. Errors on non-finite entries so the
823/// failure is caught here rather than deep in [`normalize_fisher_rao_blocks`].
824pub fn pack_probe_factors(probes: ndarray::ArrayView3<'_, f64>) -> Result<Array2<f64>, String> {
825    let (n_rows, p, s) = probes.dim();
826    if s == 0 {
827        return Err("pack_probe_factors: need at least one probe (s == 0)".to_string());
828    }
829    if !probes.iter().all(|v| v.is_finite()) {
830        return Err("pack_probe_factors: probe entries must be finite".to_string());
831    }
832    let mut u = Array2::<f64>::zeros((n_rows, p * s));
833    for n in 0..n_rows {
834        for i in 0..p {
835            for k in 0..s {
836                u[[n, i * s + k]] = probes[[n, i, k]];
837            }
838        }
839    }
840    Ok(u)
841}
842
843#[cfg(test)]
844mod tests {
845    use super::*;
846    use ndarray::array;
847
848    // ── RowMetric::euclidean ──────────────────────────────────────────────────
849
850    #[test]
851    fn euclidean_metric_has_correct_dimensions() {
852        let m = RowMetric::euclidean(5, 3).unwrap();
853        assert_eq!(m.n_rows(), 5);
854        assert_eq!(m.p_out(), 3);
855        assert_eq!(m.metric_rank(), 3);
856    }
857
858    #[test]
859    fn euclidean_metric_traces_equal_p() {
860        let p = 4_usize;
861        let m = RowMetric::euclidean(3, p).unwrap();
862        for tr in m.row_traces().iter() {
863            assert!((*tr - p as f64).abs() < 1e-14, "trace {tr} != p={p}");
864        }
865    }
866
867    #[test]
868    fn euclidean_provenance_is_euclidean() {
869        let m = RowMetric::euclidean(1, 2).unwrap();
870        assert_eq!(m.provenance(), MetricProvenance::Euclidean);
871    }
872
873    #[test]
874    fn euclidean_does_not_whiten_likelihood() {
875        let m = RowMetric::euclidean(1, 2).unwrap();
876        assert!(!m.whitens_likelihood());
877    }
878
879    #[test]
880    fn euclidean_does_not_drive_gauge() {
881        let m = RowMetric::euclidean(1, 2).unwrap();
882        assert!(!m.drives_gauge());
883    }
884
885    #[test]
886    fn euclidean_is_not_output_fisher_like() {
887        let m = RowMetric::euclidean(1, 2).unwrap();
888        assert!(!m.is_output_fisher_like());
889    }
890
891    #[test]
892    fn euclidean_solver_floor_is_zero() {
893        let m = RowMetric::euclidean(1, 2).unwrap();
894        assert_eq!(m.solver_floor(), 0.0);
895    }
896
897    #[test]
898    fn euclidean_to_weight_field_is_identity() {
899        let m = RowMetric::euclidean(1, 2).unwrap();
900        assert!(matches!(m.to_weight_field(), WeightField::Identity));
901    }
902
903    #[test]
904    fn euclidean_whiten_residual_is_passthrough() {
905        let m = RowMetric::euclidean(1, 3).unwrap();
906        let r = array![1.0_f64, 2.0, 3.0];
907        let w = m.whiten_residual_row(0, r.view());
908        assert_eq!(w, vec![1.0, 2.0, 3.0]);
909    }
910
911    #[test]
912    fn euclidean_factor_entry_is_identity() {
913        let m = RowMetric::euclidean(1, 3).unwrap();
914        assert_eq!(m.factor_entry(0, 0, 0), 1.0);
915        assert_eq!(m.factor_entry(0, 1, 1), 1.0);
916        assert_eq!(m.factor_entry(0, 2, 2), 1.0);
917        assert_eq!(m.factor_entry(0, 0, 1), 0.0);
918        assert_eq!(m.factor_entry(0, 1, 0), 0.0);
919    }
920
921    #[test]
922    fn euclidean_quad_form_is_squared_norm() {
923        let m = RowMetric::euclidean(1, 3).unwrap();
924        let r = array![1.0_f64, 2.0, 2.0];
925        assert!((m.quad_form(0, r.view()) - 9.0).abs() < 1e-14);
926    }
927
928    // ── MetricProvenance predicates ───────────────────────────────────────────
929
930    #[test]
931    fn output_fisher_drives_gauge_but_not_likelihood() {
932        let u = Arc::new(array![[1.0_f64]]);
933        let m = RowMetric::output_fisher(u, 1, 1).unwrap();
934        assert!(m.drives_gauge());
935        assert!(!m.whitens_likelihood());
936        assert!(m.is_output_fisher_like());
937    }
938
939    #[test]
940    fn whitened_structured_whitens_likelihood_and_drives_gauge() {
941        let u = Arc::new(array![[1.0_f64]]);
942        let m = RowMetric::whitened_structured(u, 1, 1).unwrap();
943        assert!(m.whitens_likelihood());
944        assert!(m.drives_gauge());
945        assert!(!m.is_output_fisher_like());
946    }
947
948    #[test]
949    fn behavioral_fisher_whitens_likelihood_and_drives_gauge() {
950        // The Rung-1 deliberate GLS metric: unlike the gauge-only OutputFisher,
951        // it whitens the reconstruction likelihood.
952        let u = Arc::new(array![[1.0_f64, 0.5]]); // p=1, probes=2
953        let m = RowMetric::behavioral_fisher(u, 1, 2).unwrap();
954        assert!(m.whitens_likelihood());
955        assert!(m.drives_gauge());
956        assert_eq!(
957            m.provenance(),
958            MetricProvenance::BehavioralFisher { probes: 2 }
959        );
960        assert_eq!(m.metric_rank(), 2);
961    }
962
963    #[test]
964    fn behavioral_fisher_quad_form_is_probe_sum() {
965        // p=2, s=2 probes v1=(1,0), v2=(0,2) → G = diag(1,4);
966        // e=(3,1) → eᵀGe = 9·1 + 1·4 = 13 = Σ (vᵢᵀe)² = 3² + 2² = 13.
967        // Column-major-within-row layout U[i,k]=u[i*probes+k]:
968        //   U[0,0]=1 U[0,1]=0  U[1,0]=0 U[1,1]=2
969        let u = Arc::new(array![[1.0_f64, 0.0, 0.0, 2.0]]);
970        let m = RowMetric::behavioral_fisher(u, 2, 2).unwrap();
971        let e = array![3.0_f64, 1.0];
972        assert!((m.quad_form(0, e.view()) - 13.0).abs() < 1e-12);
973    }
974
975    #[test]
976    fn behavioral_fisher_g_identity_reproduces_euclidean_quad_form() {
977        // GLS with G=I must reduce to plain MSE. Identity probes (s=p, U=I_p)
978        // ⇒ M_n = I ⇒ quad_form == ‖e‖², matching Euclidean bit-for-bit, and
979        // metric_rank == p so the whitened residual-dof accounting is unchanged.
980        let p = 3;
981        let mut u = Array2::<f64>::zeros((1, p * p));
982        for i in 0..p {
983            u[[0, i * p + i]] = 1.0;
984        }
985        let bf = RowMetric::behavioral_fisher(Arc::new(u), p, p).unwrap();
986        let euc = RowMetric::euclidean(1, p).unwrap();
987        let e = array![1.5_f64, -2.0, 0.25];
988        assert_eq!(bf.metric_rank(), euc.metric_rank());
989        assert!((bf.quad_form(0, e.view()) - euc.quad_form(0, e.view())).abs() < 1e-14);
990        // and whitened residual is the residual itself (identity whitening)
991        assert_eq!(bf.whiten_residual_row(0, e.view()), vec![1.5, -2.0, 0.25]);
992    }
993
994    #[test]
995    fn pack_probe_factors_matches_manual_layout() {
996        use ndarray::Array3;
997        // n=1, p=2, s=2: probes[0,i,k] = v_k[i]; v0=(1,3), v1=(2,4)
998        let mut probes = Array3::<f64>::zeros((1, 2, 2));
999        probes[[0, 0, 0]] = 1.0; // v0[0]
1000        probes[[0, 1, 0]] = 3.0; // v0[1]
1001        probes[[0, 0, 1]] = 2.0; // v1[0]
1002        probes[[0, 1, 1]] = 4.0; // v1[1]
1003        let u = pack_probe_factors(probes.view()).unwrap();
1004        // Layout U[i,k] = u[i*s + k]: [v0[0],v1[0], v0[1],v1[1]] = [1,2,3,4]
1005        assert_eq!(u.as_slice().unwrap(), &[1.0, 2.0, 3.0, 4.0]);
1006        // Round-trips into a valid metric whose G = v0 v0ᵀ + v1 v1ᵀ.
1007        let m = RowMetric::behavioral_fisher(Arc::new(u), 2, 2).unwrap();
1008        // e=(1,0): eᵀGe = v0[0]²+v1[0]² = 1+4 = 5.
1009        let e = array![1.0_f64, 0.0];
1010        assert!((m.quad_form(0, e.view()) - 5.0).abs() < 1e-12);
1011    }
1012
1013    #[test]
1014    fn pack_probe_factors_rejects_zero_probes() {
1015        use ndarray::Array3;
1016        let probes = Array3::<f64>::zeros((2, 3, 0));
1017        assert!(pack_probe_factors(probes.view()).is_err());
1018    }
1019
1020    #[test]
1021    fn output_fisher_downstream_is_output_fisher_like() {
1022        let u = Arc::new(array![[1.0_f64]]);
1023        let m = RowMetric::output_fisher_downstream(u, 1, 1).unwrap();
1024        assert!(m.is_output_fisher_like());
1025        assert!(m.drives_gauge());
1026    }
1027
1028    // ── WeightField::project_jac_row_with_u ──────────────────────────────────
1029
1030    #[test]
1031    fn project_jac_with_identity_returns_jac() {
1032        // p=2, rank=2, d=2; U=I_2, J=[[1,2],[3,4]] → M = U^T J = J
1033        let u_row = [1.0_f64, 0.0, 0.0, 1.0]; // U[i,k]=u[i*rank+k], I_2
1034        let j_row = [1.0_f64, 2.0, 3.0, 4.0]; // J[i,a]=j[i*d+a]
1035        let m = WeightField::project_jac_row_with_u(&u_row, &j_row, 2, 2, 2);
1036        assert!((m[[0, 0]] - 1.0).abs() < 1e-14);
1037        assert!((m[[0, 1]] - 2.0).abs() < 1e-14);
1038        assert!((m[[1, 0]] - 3.0).abs() < 1e-14);
1039        assert!((m[[1, 1]] - 4.0).abs() < 1e-14);
1040    }
1041
1042    #[test]
1043    fn project_jac_with_zeros_returns_zero_matrix() {
1044        let u_row = [0.0_f64, 0.0];
1045        let j_row = [1.0_f64, 2.0];
1046        let m = WeightField::project_jac_row_with_u(&u_row, &j_row, 2, 1, 1);
1047        assert_eq!(m[[0, 0]], 0.0);
1048    }
1049}