Feat sprs transition

.github/dependabot.yml

@@ -0,0 +1,19 @@
+version: 2
+updates:
+  - package-ecosystem: cargo
+    directory: "/"
+    schedule:
+      interval: monthly
+    open-pull-requests-limit: 5
+    groups:
+      # One PR per month for routine bumps; anything that needs thought will fail CI
+      # on its own rather than being buried in a batch of five.
+      minor-and-patch:
+        update-types:
+          - minor
+          - patch
+
+  - package-ecosystem: github-actions
+    directory: "/"
+    schedule:
+      interval: monthly

.github/workflows/ci.yml

@@ -0,0 +1,61 @@
+name: CI
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: clippy, rustfmt
+      # Not `--all-targets`: that pulls in the bench targets, and criterion's
+      # smoke-run of them in a debug build takes minutes.
+      - name: Test
+        run: cargo test --lib --tests --examples
+      # `las2` is off by default, so the default run never compiles that module.
+      - name: Test (all features)
+        run: cargo test --all-features --lib --tests --examples
+      - name: Doctests
+        run: cargo test --all-features --doc
+      - name: Build benches
+        run: cargo build --all-features --benches
+      - name: Clippy
+        run: cargo clippy --all-features --all-targets -- -D warnings
+      - name: Format
+        run: cargo fmt --check
+
+  # `rust-version` in Cargo.toml is a promise; without this job it is only a comment.
+  msrv:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@1.88.0
+      # Dev-dependencies routinely need a newer compiler than the library does, so the
+      # promise is checked against the library and its public API, not the test suite.
+      - name: Build at the declared MSRV
+        run: cargo build --all-features
+
+  # Plain cargo-audit rather than rustsec/audit-check: that action reports through the
+  # Checks API, which needs `checks: write` and is unavailable to fork PRs entirely. It
+  # degrades to printing the report and does not reliably fail the job.
+  audit:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+      - name: Install cargo-audit
+        run: cargo install cargo-audit --locked
+      # Exits non-zero on vulnerabilities. Unsoundness warnings are printed but do not
+      # fail the build — they are usually in transitive crates we cannot bump.
+      - name: Audit
+        run: cargo audit

.github/workflows/publish.yml

@@ -1,11 +1,8 @@
 name: Publish to crates.io
 
 on:
-  push:
-    branches:
-      - master
-    tags:
-      - 'v*'
+  release:
+    types: [published]
 
 jobs:
   publish:

.gitignore

@@ -1,3 +1,10 @@
 /target
+
+# Editor state
 .idea
-.idea/
+.idea/
+.vscode/
+*.swp
+
+# macOS
+.DS_Store

CHANGELOG.md

@@ -0,0 +1,60 @@
+# Changelog
+
+## 2.0.0
+
+A rewrite: different sparse backend, new solvers, and a handful of 1.x results that were
+just wrong.
+
+### Breaking
+
+- `sprs` replaces `nalgebra-sparse`. `SvdMat<T, I = u32, Iptr = u64>` is `sprs::CsMatI`,
+  generic over both index types — 12 bytes per non-zero for `f64` instead of 16.
+  `nalgebra` is now dev-only.
+- `u` is always `m × d` (numpy layout). 1.x returned `d × m` from Lanczos and `m × d`
+  from randomized.
+- `SMat` is split into `SparseMat` (shape, matvec) and `SparseMatDense` (block products,
+  means, centering).
+- `MaskedCSRMatrix` is now `MaskedCsMat`, and masks rows as well as columns.
+- Solvers take config builders (`IrlbaConfig`, `RandomizedConfig`) instead of positional
+  arguments.
+- `randomized` returns `single_svdlib::Result`, not `anyhow::Result`.
+- `svd` and `svd_dim` use IRLBA, not LAS2.
+- LAS2 is deprecated and behind the `las2` feature, off by default.
+- IRLBA errors instead of returning an unconverged result. `allow_unconverged()` opts out.
+- MSRV 1.88.
+
+### Added
+
+- `irlba` — restarted Lanczos bidiagonalization. The default solver.
+- `dense::tsqr` — tall-skinny QR.
+- `dense::small_svd` — one-sided Jacobi, replacing a dependency that lost accuracy on
+  badly scaled columns.
+- Block Krylov sketching in `randomized`.
+- `explained_variance`, `explained_variance_ratio`, `total_variance` and `scores` on
+  `SvdRec`.
+- `MaskedCsMat::to_sparse` — copy a view into an owned sparse submatrix.
+- `SvdRec::converged` and `max_residual`.
+- `Diagnostics::matvecs`.
+- Property tests, an edge-case suite, and benchmarks.
+
+### Fixed
+
+All checked against the published 1.0.9.
+
+- `randomized_svd` panicked for every built-in matrix type — four of five `SMat` methods
+  were `todo!()`.
+- Mean centering computed the product of the sums instead of the sum of the products.
+- `seed: None` meant seed 0, so unseeded sketches were all identical.
+- Masked matrices used the unmasked matrix for small inputs.
+- NaN or infinity in the input could hang forever.
+- All-zero matrices reported a non-zero singular value.
+- Rank-deficient matrices could fail on an absolute breakdown test.
+- `min(rows, cols) == 1` was rejected instead of solved directly.
+- The centered Frobenius norm cancelled on columns with a large offset — 51% off at an
+  `f32` offset of 1000. Now summed per entry.
+
+### Known issues
+
+LAS2 is still wrong: 18–100% error against LAPACK. One cause is fixed (`imtqlb` hoisted
+its shift out of the loop), at least one remains in `ritvec`. Inherited from 1.x, not the
+port. Its accuracy tests are `#[ignore]`d. Hence deprecated and gated.