Skip to content

Changelog

[Unreleased]

Changed

  • Releases (PyPI, crates.io, plus a Forgejo release with all wheels attached) are now produced from Codeberg via .forgejo/workflows/builds.yml. The GitHub workflow at .github/workflows/builds.yml is reduced to a CI mirror that runs Linux lint + Python smoke tests on push/PR. The Codeberg Generic Packages registry is intentionally not used as a release mirror, and CI artifact retention is kept short (1 day for wheels/sdist, 3 days for the docs/web bundles), to stay under Codeberg's 1.5 GB per-repo storage quota.
  • PyPI publication switched from --trusted-publishing always (no Codeberg/Forgejo OIDC support yet) to a PYPI_TOKEN API token.

Removed

  • macOS x86_64 wheels are no longer produced. Apple Silicon (aarch64) wheels are still published for every supported Python. Intel macOS users should install from the published sdist or pin the last release that shipped Intel wheels (equiconc==0.4.1).

[0.4.1] - 2026-04-25

Fixed

  • The 0.4.0 PyPI publish was blocked by a stale version = "0.3.0" in pyproject.toml: maturin built every wheel as 0.3.0, which collided with the existing 0.3.0 record on PyPI. pyproject.toml now uses dynamic = ["version"] and lets maturin source the version from Cargo.toml, so a single bump there propagates to both the Rust crate and the Python wheel. The 0.4.0 Rust crate (already on crates.io) is unaffected; this is purely a re-publish to land the Python wheels of the same release.

[0.4.0] - 2026-04-25

Added

  • System::solve_with_progress(on_iter) / IterationStatus / SolveControl / EquilibriumError::Aborted. The new method invokes a callback once per outer trust-region iteration (linear and log paths both); the callback may return SolveControl::Abort to short-circuit with EquilibriumError::Aborted. The existing System::solve() is now a thin shim that supplies a no-op callback, so its behavior and performance are unchanged. Intended as the hook for live progress reporting from the web UI's worker, but useful anywhere a long sweep wants to surface convergence telemetry.

  • Browser-side equilibrium-concentration solver as a separate web/ crate (Leptos + Trunk). Builds to a single static-site bundle (trunk build --release → web/dist/); no backend, no JS code, all compute runs in WebAssembly via the unmodified equiconc solver. The page exposes the full SolverOptions surface tiered Basic / Advanced / Expert, includes equiconc-defaults and COFFEE-compatible preset buttons, a kcal/mol vs RT energy-units toggle (with the temperature input hiding itself when not consumed), drag-and-drop file loading for .cfe / .ocx / .con, baked-in testcases, sortable concentrations table with per-monomer share-of-mass and 100-row pagination for COFFEE-scale 50k-species systems, a horizontal bar chart of top species by absolute concentration, a per-monomer share-of-mass stacked bar chart, a compact live convergence chart (log₁₀ ‖∇‖ vs. iteration), and TSV / CSV / JSON-report exports. The solve runs in a dedicated Web Worker so the UI thread stays responsive; a Cancel button terminates the worker mid-iteration. New just web and just web-dev recipes; CI builds the dist artifact and on main / tag pushes deploys it to GitHub Pages alongside the docs (docs at the project URL root, web app at /app/).

  • pub mod equiconc::io with parse_cfe(text, n_mon) and parse_concentrations(text) parsers for NUPACK-style complex tables (.cfe / .ocx, with NUPACK header auto-detection) and one-value-per-line concentration files (.con). Accepts whitespace, ,, ;, and | as delimiters.
  • pub fn equiconc::water_molar_density(t_c) — molar density of liquid water from the Tanaka 2001 mass-density formula. Useful for converting between molarity and mole fraction in callers that mirror COFFEE's "scalarity" wrapper.
  • equiconc::Equilibrium::mass_balance_residual / mass_balance_residual_self and a free-function equiconc::mass_balance_residual for max_i |c0_i − Σ_j A_{ji} c_j|.

Changed

  • equiconc-coffee-cli now consumes equiconc::io::parse_cfe, equiconc::io::parse_concentrations, equiconc::water_molar_density, and equiconc::mass_balance_residual instead of carrying its own copies. No behavior change.

  • The repository is now a Cargo workspace; the published equiconc package is unchanged but the workspace also contains the equiconc-web crate (publish = false).

  • New simd Cargo feature (opt-in, off by default) that vectorizes the per-species element-wise hot loops in evaluate_into and evaluate_log_into via pulp runtime ISA dispatch (SSE2 / AVX2 / AVX-512 on x86, NEON on aarch64, scalar on wasm). Enable with cargo build --features simd (or --features python,simd for the Python wheel). Translates to ~9% end-to-end speedup on the COFFEE-large equiconc_linear testcases; preserves scalar performance on equiconc_log.

Numerical contract: linear-path kernels use a degree-12 Taylor polynomial after Cody-Waite range reduction (≤2 ulps vs libm); log-path kernels keep parallelism for everything around the exp call but route the exp itself through scalar libm per lane, because the trust-region step acceptance on g = ln f requires per-iteration progress measurable above 4·eps·|g| and the polynomial residual stalls the iteration on extremely stiff systems (COFFEE testcase 0 with sub-nanomolar c0 and ~54k species was the canonical failure). The full and candidate evaluators always agree on rounding so the trust-region ρ check stays consistent.

[0.3.0] - 2026-04-25

Added

  • Optional log-objective trust-region path, selected via SolverOptions::objective = SolverObjective::Log (Rust) or SolverOptions(objective="log") (Python). The default remains SolverObjective::Linear, so existing callers see no behavior change. The log path minimizes g(λ) = ln f(λ) rather than the linear dual f(λ); on stiff systems (very strong binding, asymmetric c⁰) it often converges in many fewer iterations because g compresses the exponential dynamic range of f. The log objective is non-convex (H_g can be indefinite away from the optimum); equiconc compensates with on-the-fly modified-Cholesky regularization of the model Hessian, so the dog-leg step always sees a PD matrix and predicted_reduction > 0 by construction. The implementation structurally avoids the three documented failure modes of COFFEE's log-Lagrangian solver (see coffee-bugs.md and coffee/docs/issue2-analysis.md):

  • Bug 1 (NaN from ∞ − ∞): evaluate_log_into computes the objective via log-sum-exp on the un-clamped log_q + Aᵀλ, never forming f and then taking its log. Steps that would push f ≤ 0 are rejected and the trust region shrinks.

  • Bug 2 (premature convergence at λ ≈ 0 under strong binding): the convergence test is on the primal mass-conservation residual |Ac − c⁰|_i < atol + rtol · c0_i for both objectives — never on the log-rescaled gradient ∇g = ∇f / f, which is the term COFFEE suppresses to floating-point underflow.
  • Bug 3 / coffee issue #2 (trust-region oscillation on indefinite Hessians): the model Hessian is regularized to PD before dog-leg sees it, and a defensive pred_reduction ≤ 0 → ρ = -1 sentinel catches any residual case where regularization saturates.

Validated against COFFEE on the existing tests/proptest_vs_coffee.rs cross-check (new prop_equiconc_log_matches_coffee) and against the linear path on tests/proptest_equiconc.rs (prop_log_matches_linear). Explicit reproducers for the three documented coffee failure cases now live in src/lib.rs (coffee_bug1_positive_dg_conformer_log, coffee_bug2_strong_binding_log, coffee_issue2_strong_dimer_log). - New optional binary equiconc-coffee-cli (gated behind the coffee-cli Cargo feature) that accepts the same NUPACK-style .ocx/.cfe + .con input files as COFFEE's coffee_cli and produces the same space-separated 2-decimal-scientific results payload. Hard-codes T = 37 °C, mole-fraction scaling, and the ΔG ≥ -230 kcal/mol clamp to match COFFEE's non-configurable defaults — producing byte-for-byte agreement with coffee_cli on the monomer free concentrations of ../coffee/testcases/{0,1,2} at the {:.2e} output precision, and on the full 8-species payload of testcase 2. Integration tests in tests/coffee_cli_compat.rs verify per-species agreement on a synthetic 2-monomer/1-dimer system and on all three COFFEE testcases (skipped if ../coffee/testcases/ is absent). Build with cargo build --release --features coffee-cli. - cargo-deny configuration (deny.toml) and a CI job that runs three checks against the runtime dependency tree: - licenses: fails on any SPDX expression outside the allow-list (MIT, Apache-2.0, Apache-2.0 WITH LLVM-exception, BSD-3-Clause, Unicode-3.0). - advisories: fails on any RustSec vulnerability, unmaintained crate, unsound advisory, or yanked version. - sources: fails if any runtime crate comes from anywhere other than the default crates.io registry.

Dev-dependencies (criterion, proptest, and the cgevans/coffee git dep used for cross-checks) are excluded via [graph] exclude-dev since they ship in neither the crate tarball nor the wheel. - CI lint job: cargo fmt --all --check, cargo clippy --all-features --all-targets -- -D warnings, and cargo doc --no-deps --all-features with RUSTDOCFLAGS=-D warnings. Catches formatting drift, clippy regressions, and broken intra-doc links ahead of a docs.rs publish. - CI cross-platform / cross-Python smoke tests (tests-matrix job): the existing tests job runs only on Linux + Python 3.12 because of the cargo-llvm-cov coverage instrumentation; the new matrix exercises the corners that ship in the PyPI wheel but were never otherwise tested — minimum supported Python (3.10), free-threaded Python (3.13t), macOS, and Windows. - Dependabot configuration (.github/dependabot.yml) for weekly Cargo, GitHub Actions, and uv (Python) dependency updates, with non-major bumps grouped per ecosystem to reduce PR churn.

Changed

  • cargo fmt --all sweep across src/, tests/, benches/, and examples/. No behavior change; lands ahead of the new cargo fmt --check CI gate so the gate starts green.
  • Clippy cleanup so cargo clippy --all-features --all-targets -- -D warnings passes. Mechanical fixes: derive Default on SolverObjective; collapse field_reassign_with_default patterns in tests into struct literals; rewrite !(f > 0.0) || !f.is_finite() as f <= 0.0 || !f.is_finite() (NaN-equivalent, clippy-clean); iter().copied().collect() → to_vec(); minor doc-comment re-indentation in benches; for i in 0..n index loop → enumerate; collapse nested if let into a Rust 2024 let-chain. Type aliases ComplexSpec / PyComplexSpec introduced for the Vec<(String, Vec<(String, usize)>, _)> builder fields. #[allow] annotations with one-line comments where the lint can't tell the code is intentional: NaN-safe !(a < b) rho ordering check in SolverOptions::validate; too_many_arguments on the evaluate_into / evaluate_log_into hot-path inner functions and on the pyo3-bound PySystem::complex method.
  • builds.yml now uses concurrency: cancel-in-progress keyed on ${{ github.ref }} for pull_request events, so superseded PR runs are cancelled. Tag and main pushes are unaffected.
  • Reverted the crate license field from "BSD-3-Clause AND Apache-2.0" back to "BSD-3-Clause". The dual declaration existed only because of the vendored COFFEE sources; with vendoring removed, the published crate contains no Apache-2.0-licensed code.
  • Comparative benchmarks, the proptest_vs_coffee cross-validation test, and the instrument_large / instrument_xl diagnostic examples now depend on the coffee crate as a pinned git dev-dependency (via the cgevans/coffee fork, which gates the polars-backed file-input path behind a feature) with default-features = false, instead of carrying a vendored copy under tests/coffee_vendor/. These are all dev-only consumers, so the git source doesn't affect the published crate. Because COFFEE still pins ndarray = 0.16 while equiconc is on 0.17, an aliased ndarray_coffee = { package = "ndarray", version = "0.16" } dev-dep supplies the array types COFFEE's API requires.
  • instrument_large / instrument_xl no longer report a COFFEE iteration count (upstream's Optimizer doesn't expose one); wall time only.
  • Bumped criterion dev-dependency from 0.5 to 0.8.

[0.2.0] - 2026-04-18

Changed

  • BREAKING: Renamed the Rust System builder type to SystemBuilder. The new System is a stateful solver handle that owns numerical inputs, work buffers, and the most recent solution; it supports in-place mutation for titration / parameter sweeps and re-solves with warm-started λ. The one-shot pattern is now SystemBuilder::new()…build()?.solve()? instead of System::new()…equilibrium()?.
  • BREAKING: Equilibrium is now a borrowed view (Equilibrium<'a>) into the owning System rather than an owned struct. Use eq.get(name), eq.at(idx), or indexing (eq["AB"], eq[idx]) for lookups. To keep results past a System mutation, copy the data out (e.g. eq.concentrations().to_owned()). The borrow checker now enforces "no stale reads": mutating accessors on System cannot fire while an Equilibrium view is alive.
  • Duplicate monomers in a complex composition now have their counts summed instead of raising an error.
  • BREAKING (Rust): bumped ndarray from 0.16 to 0.17. Downstream crates consuming equiconc's ArrayView{1,2} / Array{1,2}-valued API must upgrade ndarray in lockstep.

Added

  • SolverOptions struct exposing previously-hard-coded solver knobs: max_iterations, gradient tolerances (full + relaxed), trust-region parameters (initial / max δ, ρ thresholds, shrink / grow scale factors), stagnation threshold, and two numerical clamps (log_c_clamp, optional log_q_clamp). Every field has a default matching the previous constant, so SolverOptions::default() reproduces pre-configuration behavior bit-for-bit.
  • SystemBuilder::options / options_ref, System::options / options_mut / set_options, plus System::from_arrays_with_options and System::from_arrays_with_names_and_options for passing options directly alongside numerical inputs.
  • EquilibriumError::InvalidOptions variant, raised by SolverOptions::validate() on inconsistent combinations (non-positive tolerances, shrink_rho >= grow_rho, etc.) and surfaced by every constructor that accepts options.
  • Python equiconc.SolverOptions class with keyword-only constructor mirroring the Rust fields. Pass to System(options=opts).
  • System::from_arrays and System::from_arrays_with_names for constructing a solver directly from numerical inputs without going through the string-keyed builder. Temperature is not stored at this level — callers bake it into log_q.
  • System::c0_mut, System::log_q_mut, System::set_c0, System::set_log_q for in-place mutation in titration / parameter sweep workflows.
  • System::last_solution returning Option<Equilibrium<'_>>, None if any input has been modified since the last successful solve.
  • System::validate to re-run structural invariant checks after caller-driven mutation.
  • EquilibriumError::InvalidInputs variant, raised by from_arrays / from_arrays_with_names when shapes, the identity block, monomer log_q, or name tables are inconsistent.

Python bindings are unchanged.

[0.1.0] - 2026-03-10

  • Initial release