mask-risc + jc: grouped power-sums / cross-power-sums terminals; ANOVA, Pearson, OLS from power sums; perturbation-sim angle_cov

.claude/board/entries/2026-09-29-grouped-cross-moments-fold.md

@@ -0,0 +1,19 @@
+# 2026-09-29 — Grouped cross moments: Pearson / covariance / OLS / R² over masks
+
+**Status:** MEASURED · DONE — ndarray `simd_masking_ops.rs`, `crates/lance-graph-mask-risc`, `crates/jc`
+
+## What landed
+- ndarray: `GroupCrossPowerSums { n, sum_x, sum_y: i64, sum_x2, sum_y2, sum_xy: i128 }` + `masked_group_cross_moments_i32{,_via,_pair}` over the same generic `group_walk`. `x_moments()`/`y_moments()` return the exact univariate `GroupPowerSums`; a test pins them equal to the univariate fold. Bounds: squares ∈ [0, 2^62], products ∈ [−2^62+2^31, 2^62]; the `i64` sums bind at 2^32 rows/group; the `i128` sums cannot overflow at any `u64` row count.
+- mask-risc: `Terminal::GroupCrossPowerSumsI32 { mask, key, x, y }` → `Out::CrossPowerSums`.
+- jc: `pearson_from_cross_power_sums`, `sample_covariance_from_cross_power_sums` (n−1), `simple_regression_from_cross_power_sums` (`SimpleRegression { slope, intercept }`), `r_squared_from_cross_power_sums` (`multiple_r_squared`'s k=1 contract). Shared tails extracted: `pearson_from_centered`, `sample_cov_tail`, `r_squared_tail`. Centred sums formed exactly in `i128`, checked.
+
+## Measured
+48 layouts (n ≤ 100k): |Δr| ≤ 1.3e-14, |ΔR²| ≤ 4.1e-14, relative Δslope ≤ 2.8e-14 vs the materialized path. Huge offset + tiny spread (x ≈ ±2·10⁹): fold path within 1e-15 of an independent exact reference. **Unlike ANOVA, the slice path does not fail here** — `pearson`/`multiple_r_squared` are two-pass. A naive f64 projection of the same moments returns NaN; that is what exact `i128` centring prevents (disable run: 2 tests fail).
+
+## Redundant guards (disable runs, kept as explicit contract)
+`pearson_from_cross_power_sums`'s `n < 2`, the regression's `cxx == 0`, and R²'s `cyy == 0` are each subsumed by the shared tail's zero/non-finite rejection. The first past-bound refusal fixture was vacuous (wrapped to negative centred squares); replaced with one that wraps to a plausible `r = 1`.
+
+## Open
+- Multi-membership in one pass: `lance-graph-report` lowers `CoordSpec::MaskSet` to one `Filter::Plane` pass per member tuple and per fold state. The single-group limit is in `GroupKeyAddr::group_of` (one `Option<usize>` per row), in the IR's `GroupKey`, and in report lowering. Seam: a word-level walker `for each 64-row word: sel & plane_m → fold hits into out[m]` over existing planes — same plane traffic as K passes, value-lane traffic of the union instead of the sum.
+
+**Renamed 2026-09-30 (rebase onto ndarray master):** upstream shipped `PowerSums { n, sum, sum_sq: u128 }` — the same record as `GroupPowerSums`. The duplicate was dropped: consumers use `PowerSums` / `CrossPowerSums` and `masked_group_(cross_)power_sums_i32*`; `checked_merge` moved onto the upstream types. Square sums are now `u128`; jc converts with `i128::try_from`, and a value past `i128::MAX` is refused as past the bound.

.claude/board/entries/2026-09-29-grouped-moments-fold-anova-over-masks.md

@@ -0,0 +1,18 @@
+# 2026-09-29 — One-way ANOVA over a population mask from grouped moments
+
+**Status:** MEASURED · DONE — ndarray `simd_masking_ops.rs`, `crates/lance-graph-mask-risc`, `crates/jc/src/stats.rs`
+
+## What landed
+- ndarray: `GroupPowerSums { n: u64, sum: i64, sum_sq: i128 }` and `masked_group_moments_i32{,_via,_pair}` — `(n, Σx, Σx²)` per group in one pass over the rows a mask selects, over the existing `group_walk` (now generic over its slot type). `checked_merge` is exact integer addition; exact up to `GROUP_MOMENTS_MAX_ROWS = 2^32` rows per group (the `i64` sum is the binding field).
+- mask-risc: `Terminal::GroupPowerSumsI32 { mask, key: GroupKey, val }` → `Out::PowerSums`. Own terminal, not a `GroupFold` member (a `GroupFold` slot is one seeded `i64`). Refuses planes past `MASKED_SUM_I32_MAX_ROWS`. Partial extents: admitted since #1323's extent wiring — per-extent sinks merge by `checked_merge` (pinned in `tests/extent.rs`).
+- jc: `anova_from_power_sums` / `eta_squared_from_power_sums`. F/p/η² policy is shared with `anova_one_way` / `eta_squared` via `anova_from_ss` / `eta_from_ss`. Sums of squares come from exact `i128` quantities; no two large floats are subtracted.
+
+## Measured
+Mask → fold → `anova_from_power_sums` vs materialized → `anova_one_way`, 42 non-degenerate layouts (n ≤ 100k, k ≤ 6, densities 20/128/256 of 256): max relative ΔF 4.2e-14, max |Δp| 4.4e-16, max |Δη²| 6.5e-16. Adversarial layout (group means ±2·10⁹, within spread 1): exact F = 3·10¹⁹; the moments path returns 3·10¹⁹, the slice path returns `None` (its `ss_t − ss_b` cancels the within-group SS to ≤ 0). A naive float `Q − S²/N` inside the moments path fails even the near-`i32::MAX` agreement test (ΔF ≈ 1e-3), so the exact forms are load-bearing.
+
+## Open
+- Multi-membership is NOT one pass: `GroupKey` resolves one group per row, and `lance-graph-report`'s `CoordSpec::MaskSet` executes one `Filter::Plane` program per member. Seam: a key address that walks the selected rows once and folds each row into every member plane that holds it.
+- Bivariate `(n, Σx, Σy, Σx², Σy², Σxy)`: same walk, a second value lane in the closure, a wider slot type; not built.
+- No early exit, and no `CausalEdge64` commit. ANOVA is non-monotone under future rows.
+
+**Renamed 2026-09-30 (rebase onto ndarray master):** upstream shipped `PowerSums { n, sum, sum_sq: u128 }` — the same record as `GroupPowerSums`. The duplicate was dropped: consumers use `PowerSums` / `CrossPowerSums` and `masked_group_(cross_)power_sums_i32*`; `checked_merge` moved onto the upstream types. Square sums are now `u128`; jc converts with `i128::try_from`, and a value past `i128::MAX` is refused as past the bound.

.claude/board/entries/2026-09-29-perturbation-sim-angle-covariance-via-cov-high-d.md

@@ -0,0 +1,31 @@
+# 2026-09-29 — perturbation-sim: Σθ = L⁺ Σp L⁺ through ndarray CovHighD::sandwich
+
+**STATUS:** MEASURED · **Scope:** `crates/perturbation-sim/src/angle_cov.rs` (feature `pillar`), ndarray `hpc::pillar::cov_high_d`
+
+## What landed
+- ndarray `4d4ee17`: `CovHighD::from_symmetric_fn` + public `get`. Before this a consumer
+  could not build a CovHighD from data without re-spelling the packed index.
+- lance-graph `bb6d044`: `angle_covariance::<N>(eig, Σp, rel_tol)`, the stochastic twin of
+  `pseudo_apply`. It is exact for DC flow in a fixed topology. It calls the Pillar-9 sandwich;
+  there is no local kernel.
+
+## Measured (6-bus ring + chords)
+- Against an f64 dense triple product: rel err 1.2e-7. This is the f32 floor.
+- Rank-1 Σp = ppᵀ against `pseudo_apply` outer product: 2.3e-7. That path never forms L⁺.
+- 40k-sample Monte Carlo of the deterministic solver: rel err 0.0009.
+- Disable runs, each red under exactly its own test:
+  - M := I;
+  - symmetry guard removed;
+  - N-mismatch guard removed.
+
+## Finding in ndarray
+Every existing `sandwich` test used M = I. Under a deliberately broken kernel (M·Σ·Σ),
+`sandwich_identity_is_identity` stayed green. The new dense non-identity test is the first
+that can see an index/transpose defect in the Pillar-9 kernel.
+
+## OPEN
+- CovHighD is const-generic N; `Grid::n` is runtime. The caller picks N and it is checked.
+  A runtime-sized sandwich is an ndarray change, not taken.
+- Line trips are a rank-1 update of L⁺ (LODF), not a sandwich. They are not wired.
+- Line-flow covariance has a non-symmetric rectangular Jacobian, which sandwich cannot
+  express. Per-line variance is a 2-sparse quadratic form over Σθ, not built.

.claude/board/entries/README.md

@@ -25,7 +25,7 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-208 entries, 2026-08-06 .. 2026-10-04.
+211 entries, 2026-08-06 .. 2026-10-04.
 
 | date | entry id | finding | file |
 |---|---|---|---|
@@ -61,6 +61,9 @@ exactly one of them, never both.
 | 2026-09-30 | `three-reference-sets-are-not-ordinal-aligned` |  | [2026-09-30-three-reference-sets-are-not-ordinal-aligned.md](2026-09-30-three-reference-sets-are-not-ordinal-aligned.md) |
 | 2026-09-30 | `deepnsm-v2-coverage-bands` |  | [2026-09-30-deepnsm-v2-coverage-bands.md](2026-09-30-deepnsm-v2-coverage-bands.md) |
 | 2026-09-30 | `cypher-mask-v2-is-a-replacement-not-a-phase` |  | [2026-09-30-cypher-mask-v2-is-a-replacement-not-a-phase.md](2026-09-30-cypher-mask-v2-is-a-replacement-not-a-phase.md) |
+| 2026-09-29 | `perturbation-sim-angle-covariance-via-cov-high-d` |  | [2026-09-29-perturbation-sim-angle-covariance-via-cov-high-d.md](2026-09-29-perturbation-sim-angle-covariance-via-cov-high-d.md) |
+| 2026-09-29 | `grouped-moments-fold-anova-over-masks` |  | [2026-09-29-grouped-moments-fold-anova-over-masks.md](2026-09-29-grouped-moments-fold-anova-over-masks.md) |
+| 2026-09-29 | `grouped-cross-moments-fold` |  | [2026-09-29-grouped-cross-moments-fold.md](2026-09-29-grouped-cross-moments-fold.md) |
 | 2026-09-29 | `deepnsm-v2-counted-pick-tag-deltas` |  | [2026-09-29-deepnsm-v2-counted-pick-tag-deltas.md](2026-09-29-deepnsm-v2-counted-pick-tag-deltas.md) |
 | 2026-09-26 | `deepnsm-v2-lexical-evidence-survives-routing` |  | [2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md](2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md) |
 | 2026-09-25 | `window-scheduling-and-two-level-ternlog` |  | [2026-09-25-window-scheduling-and-two-level-ternlog.md](2026-09-25-window-scheduling-and-two-level-ternlog.md) |

crates/jc/src/reliability.rs

@@ -110,6 +110,14 @@ pub fn pearson(x: &[f64], y: &[f64]) -> Option<f64> {
         sxx += dx * dx;
         syy += dy * dy;
     }
+    pearson_from_centered(sxy, sxx, syy)
+}
+
+/// Pearson's `r` from the centred co-moment and the two centred
+/// second moments — the ONE place its degeneracy policy lives, shared by
+/// [`pearson`] and `stats::pearson_from_cross_power_sums`. Any common positive
+/// scale on all three (e.g. `n·S` instead of `S`) cancels.
+pub(crate) fn pearson_from_centered(sxy: f64, sxx: f64, syy: f64) -> Option<f64> {
     let denom = (sxx * syy).sqrt();
     if denom == 0.0 || !denom.is_finite() {
         // `denom == 0` → at least one series is constant. `denom == ∞` → the