pm3-rs-python 0.1.2__tar.gz → 0.2.5__tar.gz

pm3_rs_python-0.2.5/CHANGELOG.md

@@ -0,0 +1,971 @@
+# Changelog
+
+## 0.2.5
+
+A broad oracle, and the three defects it found.
+
+0.2.4 was validated against a dozen hand-picked molecules. This release compares
+against **189, spanning every element PM3 is parameterized for**, and the answer
+to "does pm3-rs reproduce MOPAC" is now a measurement rather than a belief: all
+189 agree on heat of formation, dipole, net atomic charges and ionization
+potential. Getting there took fixing two real bugs and correcting one assumption
+about what the oracle was.
+
+### Added
+
+- **`tests/data/mopac_oracle.tsv`** — MOPAC v23.2.5's own answers for 189
+  molecules, generated by `tools/oracle/build_oracle_set.py` and read by
+  `tests/mopac_oracle.rs`. Geometries are MOPAC-optimized (so each fixture is a
+  PM3 stationary point) except where there is nothing to optimize — a noble-gas
+  pair has no bound minimum, a Sparkle no valence orbitals — and those rows say so
+  in a `relaxed` column. Coverage is the whole parameter set: main group, the
+  noble gases, the heavy s-block, Zn/Cd/Hg, the fifteen La–Lu Sparkles and MOPAC's
+  `Cb`/`+`/`-`. Francium is the one absentee, because neither code has parameters
+  for it and both refuse it.
+
+  **Agreement: 189 of 189**, worst heat of formation `7.2e-4 kcal/mol`, worst
+  dipole `1.1e-3 D`, worst charge `1.6e-4 e`.
+
+- **`Pm3Options::stability`** ([`ScfStability`]) — an SCF stability search. A
+  converged SCF is one fixed point of an operator that has several, and nothing in
+  a convergence test says it is the ground state. When the frontier gap is small
+  enough that the orbital ordering was in doubt, the molecule is re-solved from
+  deliberately different starting points and the lowest solution kept.
+
+- **`Pm3Options::guess`** ([`ScfGuess`]) — the starting density is now selectable
+  (`Sad`, `Core`, `SymmetryBroken`), because it decides *which* solution is
+  reached and not merely how fast.
+
+- **`Pm3Options::mmok`** — MOPAC's `MMOK` molecular-mechanics amide correction,
+  `K·sin²(O=C–N–H)` per amide hydrogen with `K = 7.1853 kcal/mol`. **Off by
+  default**, unlike MOPAC. Available from Python and the CLI as a `+mmok` suffix
+  on the method (`method="pm3+mmok"`, `--method pm3-d3h4+mmok`). Written over the
+  crate's dual numbers, so its gradient and Hessian contributions come from the
+  same expression and an optimization with it on minimizes what it reports.
+
+### Fixed
+
+- **The SCF converged to excited solutions on seven molecules**, by up to 279
+  kcal/mol, and reported them as converged. `2-butyne` (+279), `ketene` (+204),
+  `C₂F₄` (+166), `C₂Cl₄` (+157), the `O₂` triplet (+135), `NH₂•` (+37) and the
+  methyl radical (+111) each reached a self-consistent state that obeyed the
+  aufbau principle among its own eigenvalues and was not the ground state. Every
+  one of them now agrees with MOPAC.
+
+  Three separate causes, which is why no single change fixed them: DIIS stepping
+  over the fixed point (`2-butyne`, `ketene`), the SAD guess being in the wrong
+  basin (`C₂Cl₄`), and the restricted iteration having no path to the solution at
+  all (`C₂F₄` — reachable only through the unrestricted loop, which converges
+  spin-pure there and so returns a restricted state). The search covers all three
+  by trying them and keeping the lowest.
+
+  It fires on 6 of the 189 and costs nothing on the rest: the frontier gap
+  separates the two populations cleanly (every wrong solution below 4 eV, every
+  first-try-correct one above 5.7), so the trigger is a gap threshold sitting in
+  the empty band between. The whole 189-molecule set solves in under a second.
+
+- **The unrestricted path started from core-Hamiltonian orbitals**, which is the
+  wrong guess. From SAD, the `O₂` triplet and the amino radical land on MOPAC's
+  answer instead of 135 and 37 kcal/mol above it, and the methoxy radical takes a
+  few dozen iterations instead of about 1600.
+
+- **`Pm3Result` gained `scf_paths_tried` and `scf_improvement_ev`**, so a caller
+  can see when the search changed the answer rather than having to infer it.
+
+### Changed
+
+- **MOPAC's default is not PM3, and the oracle now says so.** MOPAC applies `MMOK`
+  unless given `NOMM`, adding a classical amide torsion term after the SCF. It
+  changes no orbital and no total energy, which is what makes it easy to mistake
+  for a discrepancy: acetamide's density, dipole and every eigenvalue agreed with
+  ours to six figures while its heat of formation was 0.66 kcal/mol away. Every
+  row of the oracle is now generated with `NOMM`, so the file is plain PM3.
+
+  Both the form and the constant of the correction were measured against MOPAC
+  rather than copied, by scanning formamide's torsion with `MMOK` and `NOMM` and
+  subtracting; the fit is exact to the digits MOPAC prints.
+
+## 0.2.4
+
+Most of what was broken in 0.2.3 was not the physics but the **surface**: things
+the library could do that no caller could ask for, and one assumption that was
+true at Γ and quietly false everywhere else.
+
+### Fixed
+
+- **A meshed `D(q)` was wrong at `q = 0`** — by 300%, anisotropic on a cubic
+  crystal, and with the *sign* reversed, so a stable structure reported an
+  imaginary optical mode. One substitution made in two places: `P(T) = P(0)`,
+  which **is** Γ sampling (one k-point cannot tell images apart) and is false on a
+  mesh, where `P(0)` is the Brillouin-zone average while `P(T)` decays with `T`.
+  `phased_skeleton` asserted it for the skeleton — the dominant half — and
+  `bare_blocks` asserted it again for the response's exchange term. NaCl `3×3×3`
+  now gives an isotropic `D(0)` agreeing with a central difference of the analytic
+  forces to `1e-4`. Γ sampling is unchanged bit for bit: there the two expressions
+  are the same number.
+
+  The clue had been recorded a release earlier and not used — the error converged
+  smoothly with mesh size, **0% at `1³`** rising to 77% at `7³`. Zero at `1³` is
+  the tell: a `1×1×1` mesh *is* Γ.
+
+- **`native.optimize()` returned no `energy_ev`, and `relax()` no `iterations`
+  or heat of formation.** Reading the docs for one and applying them to the other
+  ended a structure optimization with a `KeyError` from the call that had done the
+  most work. All three optimizers now answer to the same names, held by a
+  table-driven contract test over every native function.
+
+- **`pm3-rs optimize big.xyz --dc 4.8` optimized nothing** and exited 0: the
+  divide-and-conquer branch returned before the command was read. The same held
+  for `frequencies`, `hessian`, `gradient`, `molden`, `ir` and `charges`.
+
+- **Translations and rotations are removed by projection**, not by being small.
+  The rigid-body subspace is discovered from the geometry — 2 rotations for a
+  linear molecule, 0 for an atom — and those modes are **exactly** `0.00 cm⁻¹`.
+  Three different thresholds (`< 50`, `< 100`, `< 300 cm⁻¹`) for the same quantity
+  are gone, as is `SOFT_MODE_FLOOR`.
+
+- **The Γ acoustic sum rule was applied row-wise**, which is not symmetric, and
+  the eigensolver reads one triangle — so the rule that had just been imposed was
+  partly undone and `PeriodicPhonons::hessian` came back non-symmetric. It is a
+  symmetric projection now, and `acoustic_residual_cm` reports the **pre**-projection
+  residual, which is the number worth seeing rather than the one the correction
+  just flattened.
+
+- **`force_tol` was converted the wrong way** in the new CLI and partitioned
+  paths — a gradient scales the opposite way to a length, so the factor is
+  `BOHR_TO_ANGSTROM`. The tolerance was 3.57× looser than asked for, the optimizer
+  stopped early, and it still reported `converged: true`. Now
+  `constants::force_tol_to_au` and `stress_tol_to_au`, with a test that pins the
+  direction.
+
+- **A stalled periodic SCF had no recourse on the Γ path and one rung on the
+  k-point path.** Damping now goes ahead of smearing on both, because its guarantee
+  is stronger: it changes the path and not the equations, and both loops measure
+  convergence against the *undamped* step, so a converged damped run solves what was
+  asked. Silicon on a 3×3×3 mesh is rescued (−145.536995 eV); diamond still fails,
+  because its residual alternates rather than decays and no mixing fraction reaches
+  a symmetry-breaking limit cycle. There is deliberately no third rung: a level
+  shift converges diamond to a *different solution*, which is worse than not
+  converging, so it stays in the diagnosis for the caller to choose.
+
+- **A periodic response returned an incomplete answer for a partially occupied
+  cell, silently.** The occupation factors are the metallic
+  `(f_n − f_m)/(ε_n − ε_m)` form, which makes it look handled; what is missing is
+  the Fermi-level shift, and the sum skips the near-degenerate pairs that carry the
+  Fermi-surface term. It is refused now — keyed on whether the occupations came out
+  **integral**, not on whether smearing was used, so a gapped cell under a small
+  smearing still runs. Same `T·S` certificate the SCF rescue uses.
+
+- Seventeen lines across three files had been committed after a PowerShell cp932
+  round-trip. In `pbc/mod.rs` the casualties were the electrostatics derivation:
+  `√(r² + a)` had lost its root and bracket and `[ W_KO(r) − W_point(r) ]` its
+  minus sign, so a formula read as noise rather than as wrong.
+
+### Added
+
+- **Orbital energies, coefficients and occupations** (`native.orbitals`,
+  `pm3-rs orbitals`, `--coefficients`). Coefficients come with `ao_labels`, and
+  the frontier is taken across both spin channels — a radical's β LUMO sits below
+  its α one.
+- **Phonon eigenvectors.** They were computed and discarded. `modes` (real) at Γ,
+  `modes_real`/`modes_imag` at a wavevector, through `native.phonons` and
+  `PM3.get_phonons`.
+- **`Pm3Options::cphf_max_iter`** and a `cphf_max_iter=` argument on the six
+  Python entry points that run a coupled-perturbed solve. It was two hard-coded
+  `400`s, so a stiff response could only be rescued by editing the crate.
+- **Divide-and-conquer geometry optimization** from the CLI (`--dc` with
+  `optimize`), `native.divide_and_conquer_optimize` and `PM3`.
+- **Molden `[STO]`** (`--sto`, `basis="sto"`) beside the default `[GTO]`, and
+  `--output` to say where the file goes.
+- **`--pbc` in the spelling the reader has in mind**: `1,0,1`, `101`, `xz`,
+  `x,z`, `true,false,true`, `none`.
+- **The periodic optimizer's controls**: `--relax-cell`, `--fixed-cell`,
+  `--max-steps`, `--force-tol`, `--stress-tol`, `--pressure`. `write_xyz` now
+  carries the cell as extended-XYZ `Lattice="..."`, which was silent data loss the
+  moment variable-cell relaxation existed.
+- `--kpts` and `--q` on the CLI; `native.divide_and_conquer_forces`; periodic
+  `hessian` from Python.
+
+### Performance
+
+- **Periodic runs are about 10× faster.** A 48-atom Γ single point went from
+  17.7 s to 1.82 s; the Python test suite from 396 s to 72 s. The numbers are
+  unchanged.
+
+  None of the three optimizations that had been planned were the problem. Measured
+  first, as `examples/born_profile.rs` established: the serial k-point loop was
+  31–45% of a *meshed* run and nothing of a Γ one, the serial Fock build was
+  **0.4%** of an iteration at every size from 6 to 72 atoms, and an optimization
+  step cost 1.6 gradients rather than the up-to-30 the line search was feared to be
+  paying. It was the Ewald sum, at 89–98% of every iteration — and inside that,
+  90.5 s of real space against 2.5 s of reciprocal, so the guess *within* the guess
+  was wrong too.
+
+  The real-space sum evaluated `erfc` and `exp` for every pair inside the cutoff,
+  **twice per iteration**, at distances that do not move while the charges do. The
+  kernel is geometry and now sits in `EwaldContext` beside the phase table. Real
+  space fell 32×. `examples/periodic_profile.rs` and `PM3_GAMMA_PROFILE=1` are the
+  measurement.
+
+- **A Γ-point Hessian is 2.3× faster, and its Ewald phase 210×.** 48 atoms: 51.3 s
+  to 22.8 s, with the Ewald second derivative going from 31.60 s to 0.15 s. Numbers
+  unchanged.
+
+  Profiled first, again, and again it was not the assumption: the coupled-perturbed
+  response — the phase everyone expects to dominate — was 38%, the skeleton rounded
+  to zero, and the Ewald term was 62%. Its inner sum turned out to be a **structure
+  factor written out longhand**: what the site-pair loop accumulates for an atom
+  pair is `Re[S_A S_B*]`, so `O(N_G · N_sites²)` becomes `N_sites` to build the
+  factors plus `N_atoms²` to combine them — and PM3 puts four to five multipole
+  sites on every heavy atom. `examples/hessian_profile.rs` and
+  `PM3_HESSIAN_PROFILE=1` are the measurement. The Hessian is now 96% response,
+  which is where the next look starts.
+
+### Licensing
+
+- **The wheel shipped `LICENSE` alone.** `_native.pyd` has the MOPAC-derived PM3
+  parameter tables compiled into it, so `pip install` distributed Apache-2.0
+  material without the licence text or the attribution. It now carries
+  `THIRD_PARTY_NOTICES.md`, the Apache text and every `NOTICE`.
+- **Nothing recorded the ~60 Rust crates statically linked into the binaries.**
+  `faer`, `rayon`, `pyo3` and their transitive closure are compiled into
+  everything shipped, and MIT asks for the copyright notice to travel with a
+  substantial portion. `third_party/rust-crates/` now carries the index and 524 KB
+  of verbatim texts, generated by `tools/collect_rust_notices.py`.
+
+  Two things the graph turned up: `faer` declares MIT and ships **no MIT text**
+  (what it ships is its own upstream Eigen/LAPACK/SuiteSparse attributions), and
+  `unicode-ident` is `(MIT OR Apache-2.0) **AND** Unicode-3.0` — the only entry
+  where choosing a licence does not settle the question.
+- `tests/attribution.rs` holds all of it to the tree: every linked crate is in the
+  notice, the two distributions agree, no upstream source is tracked, and no file
+  has been through a codepage round-trip.
+
+### Testing
+
+- **Every CLI command against every flag**, mechanically — the matrix that found
+  `--dc` swallowing seven commands. Plus a second layer running each command
+  through the installed console script, where only `energy` had ever been
+  exercised across the Python↔Rust boundary.
+- **All fifteen La–Lu Sparkles against MOPAC** at `1e-5 kcal/mol`, re-run through
+  the executable rather than compared to a stored JSON.
+- `tests/pbc_cubic_identities.rs` has no `#[ignore]`s left, and neither does the
+  suite.
+
+### Known limits
+
+- PM3's transition metals are **Zn, Cd and Hg** and nothing else. `TiO₂`, `SrTiO₃`,
+  `LiCoO₂` and ferrocene are outside the model, not unimplemented. Those three are
+  enough for real organometallic chemistry: `examples/phonon_structures.rs` and
+  `examples/organometallic_bands.rs` cover `Zn(CH₃)₂` and `Hg(CH₃)₂` — linear
+  dialkyls with genuine `M–C` σ bonds — and the `Cd(CN)₂` framework.
+
+- **Phonons must be evaluated at PM3's own geometry, not at the experimental one**,
+  and for ionic solids the two are far apart. A Hessian off a stationary point is
+  not a phonon spectrum, so the experimental lattice constant is the wrong place to
+  ask. At the lattice constant PM3 actually prefers, MgO's spectrum is clean — but
+  that equilibrium sits 12.7% *above* the measured value, and the highest mode comes
+  out at 745 cm⁻¹ against 401 measured.
+
+  For NaCl the energy is still falling at the compressive edge of the range Γ
+  sampling can hold, so PM3 has no bound rocksalt minimum there at all. Neither is
+  a defect in the second-derivative code; PM3 was fitted to molecular heats of
+  formation and never saw a Madelung lattice.
+
+- Where PM3 *is* on firm ground is the molecular chemistry it was fitted to.
+  Dimethylzinc relaxes to a Zn–C bond of 1.924 Å against 1.930 Å measured, and its
+  symmetric Zn–C stretch comes out at 599 cm⁻¹ against ≈615.
+
+- A periodic SCF on some cubic cells with an even Monkhorst–Pack mesh still
+  converges to a mesh-dependent solution; see `docs/pbc.md`.
+
+## 0.2.3
+
+One performance fix, found by running real crystals rather than by reading code,
+and two reporting defects it turned up on the way.
+
+### Faster
+
+- **Born charges and Γ-point phonons are ~2.4× faster**, and the numbers are
+  unchanged. The long-range term of the bare perturbation ran two lattice sums
+  over the site pairs *inside* the per-degree-of-freedom loop — `6N` sums where
+  two suffice. The displacements do not depend on the Cartesian axis at all, so
+  each was recomputed three times per atom for nothing, and across atoms the
+  sets are disjoint slices of one site-by-site table. It is now built once
+  (`pbc::dfpt::LongRangeKernels`).
+
+  At 12 atoms a Born-charge run went from 62.2 s to 26.3 s; the Rust suite's
+  periodic tests roughly halved.
+
+  Worth recording how it was found, because three earlier attempts missed it.
+  Replacing the `3N` coupled-perturbed solves with three by the interchange
+  theorem changed the wall clock by **nothing**; hoisting the neighbour list out
+  of the bare perturbations changed **nothing**; parallelising the three field
+  solves changed **nothing**. Timing the pieces put 85% of the run in one
+  untimed call. The lesson is in `examples/born_profile.rs`, which decomposes
+  the cost with public entry points that differ in one term each.
+
+### Fixed
+
+- **`Pm3Error::ScfNotConverged` always reported `error: NaN`.** It was
+  hardcoded at all four sites — the Γ path, the k-point path and both molecular
+  paths — so the field meant to say how far off a run was never did, and `NaN`
+  reads as a numerical blow-up when the truth was usually a slow tail. Diamond's
+  conventional cell reports `3.854e-4` against a `1e-7` tolerance; it was not
+  diverging at all.
+- **A failed Γ-point SCF never mentioned the Γ margin**, which is usually the
+  cause. Below it one k-point cannot represent the cell and no amount of damping
+  helps. The message now gives the margin and says to use a k-mesh or a larger
+  supercell. Measured on NaCl's conventional cell, where the Γ answer and the
+  4×4×4 one differ by **421 eV**.
+
+### Housekeeping
+
+- `cargo fmt` across the tree: 0.2.2 shipped 26 files that rustfmt disagreed
+  with, all of them new or edited in that release.
+
+## 0.2.2
+
+The response properties a dynamical matrix was already most of the way to, and a
+Berry phase to check them from outside. The largest fix is none of those: a
+coupled-perturbed solver that had been returning unconverged responses under a
+converged label, which every analytic Hessian and infrared intensity this crate
+has produced was built on.
+
+### Added
+
+- **Born effective charges** (`pbc::born`, `pm3_rs.born_charges`, CLI `born`,
+  `PM3.get_born_charges`), reported with the acoustic sum-rule residual rather
+  than with the rule imposed.
+- **Polarizability and the dielectric tensors** (`pbc::dielectric`,
+  `pm3_rs.dielectric`, CLI `dielectric`, `PM3.get_dielectric`): `ε∞` for a
+  fully periodic cell, `α` in every dimensionality, and — with `include_ionic`
+  / `--static` — the static tensor `ε₀`. `skipped_modes` says how much of the
+  ionic sum is missing, since at an unrelaxed geometry the answer is incomplete
+  rather than merely odd.
+- **LO–TO splitting** (`pbc::lo_to`, `phonons(lo_to_direction=)`,
+  `--lo-to`). The coefficient was not transcribed: it was fixed against this
+  crate's own finite-`q` DFPT in the `q → 0` limit, where the ratio converges to
+  1 as `q²` (0.56 → 0.89 → 0.973).
+- **Supercell force constants and phonon dispersion** (`pbc::phonon`,
+  `ForceConstants`, `pm3_rs.phonon_bands`, CLI `phonon-bands`,
+  `PM3.get_phonon_bands`). One Hessian buys the whole band structure. `Φ(T)`
+  lives in a sorted vector, not a hash map: the Fourier sum's accumulation order
+  is otherwise run-dependent, which in a near-degenerate mode is not a
+  last-digit effect.
+- **Berry-phase polarization** (`pbc::berry`, `pm3_rs.berry_polarization`, CLI
+  `berry`, `PM3.get_berry_polarization`), as an independent check on the above
+  rather than as a feature in itself.
+- **A finite electric field along a periodic direction** (`pbc::finite_field`,
+  `pm3_rs.finite_field`, CLI `finite-field`, `PM3.get_finite_field`), by the
+  Nunes–Gonze electric enthalpy `F = E − Ω 𝓔·P`. `𝓔·R` is not lattice-periodic
+  along a periodic axis, so `H − 𝓔·R` has no ground state there; a field
+  orthogonal to every lattice vector still goes through `Pm3Options::field`.
+  Reproduces the CPHF polarizability to 1 part in 10⁴. Restricted, gapped, 3D
+  only, and there is no force.
+- **`DfptOptions` / `LongRange` / `DfptResult`**, `force_constants_at_q`,
+  `frequencies_at_q`, and the response density (`PhononResponse`) everything
+  above contracts.
+- **The Mermin electronic free energy**: `KpointResult::entropy_ts_ev` and
+  `free_energy_ev`, and ASE's `free_energy`, so
+  `get_potential_energy(force_consistent=True)` works. Documented in four places
+  as *not* a Gibbs energy — no zero-point energy, no vibrational partition
+  function, no `pV`, no nuclear entropy.
+- **`divide_and_conquer_forces`**, `periodic_hessian`, `magnetization`,
+  `smearing_ev`, CLI `--reference`, and the mode vectors and masses
+  `frequencies` and `dynamical_matrix` had been withholding.
+
+### Fixed
+
+- **CPHF returned unconverged responses as converged.** All four paths ran to an
+  iteration limit and returned `Ok`. Adding the check failed eight existing
+  tests, which exposed the cause: `hessian.rs`'s own DIIS divided `B_ij` by each
+  residual's magnitude, replacing the constraint `Σc = 1` with a different one.
+  The correct implementation was already in `scf.rs`. Water's response went from
+  160 passes to **3** (residual `8.1e-17`); the Rust suite from 38.7 s to 31.4 s.
+- **`D(q)` diverged as `1/q²`** in a polar cell — the acoustic sum rule read
+  220.1 at `q = 0.0125`, and water in a 10 Bohr cell put its lowest mode at
+  −4705 cm⁻¹. The `G = 0` term of the bare long-range perturbation was scaled by
+  the displaced atom's own charge rather than the cell's, so `Σ_a Q_a = 0` never
+  cancelled it. Fixed by excluding the macroscopic field from the
+  self-consistent response and restoring it analytically; the term restored is
+  exactly `pbc::lo_to`. No existing DFPT test ran below `q = 0.2`.
+- **`dynamical_matrix_on_mesh` built its skeleton from the Γ density**, so
+  passing a mesh corrected the response and left the dominant term wrong.
+- **ASE ran two SCF calculations per step and stitched them together**, taking
+  energy and forces from an unsmeared one and charges from a smeared one — on a
+  metal, different states. Now one call; the Python suite went 99 s → 51 s.
+- **UHF CPHF had no DIIS**, so the hardest case was the one without acceleration.
+- **Open-shell HOMO/LUMO ignored the β spectrum.**
+- **`stress_tol` crossed the Python boundary without its unit conversion**,
+  leaving the periodic optimizer 6.75× looser than documented and returning
+  relaxed cells with `converged=True`.
+- **`long_range_cutoff` was silently ignored by periodic divide-and-conquer**,
+  and `--kpts` by three CLI paths that cannot sample a mesh. Both now refuse.
+- **ASE's result cache was not invalidated by attribute changes**, so setting
+  `atoms.calc.charge = 1` returned the neutral energy.
+- **`vibrational_analysis` divided by zero mass** for MOPAC's sparkles.
+
+### Fixed in the new code, by cross-checks rather than by review
+
+- **The field operator was half its correct size.** `M = i λ (W₊ − W₋) C†` is
+  one-sided — `M|v⟩ = 0` for a virtual `v` — so it holds the whole
+  virtual-occupied block and none of the occupied-virtual one. The conventional
+  `½(M + M†)` therefore halves the block a linear response is made of. Measured
+  at a ratio of `0.5001` against the CPHF polarizability; `M + M†` gives
+  `1.0003`.
+- **The finite field computed polarization only along the axes the field
+  touched.** Those are not the axes that carry polarization: a zero field came
+  back with an electronic polarization of exactly zero. `resolved` now reports
+  which axes the mesh could see.
+- **The CLI handed `run_finite_field` a negated field.** `--field` is stored
+  with the sign that makes `E = E₀ + μ·F` hold for the molecular `−𝓔·r`
+  coupling. Nothing failed — the loop converged and `∂P/∂𝓔` was backwards.
+
+### Wiring
+
+A layer-by-layer audit found four features present in some layers and not
+others, the same shape of gap as the "DFPT is Rust-only" report:
+`berry_polarization` was Rust-only, supercell dispersion was CLI-only, the
+finite field reached neither ASE nor the CLI, and LO–TO was unreachable from the
+CLI. All four are now in every layer.
+
+### Measured, and not fixed
+
+- **The correction Hessian is still `O(N⁴)`.** Bounding the D3/H4 sums at the
+  radii the periodic path uses was tried and reverted: it cost `2.5e-4` eV at
+  375 atoms and bought nothing, the log-log slope staying at **1.959**. The sums
+  are `for i { for j { if r > cutoff { continue } } }`, so a radius bounds the
+  range and not the work — and 30 Bohr encloses ~2000 atoms of liquid water,
+  more than any system measured. See `examples/correction_cutoff.rs` and
+  `examples/correction_scaling.rs`.
+- **Divide-and-conquer is at slope 1.16**, against an acceptance criterion of
+  1.15, over 360 to 2880 atoms — 86× over full diagonalization at 2880 atoms, at
+  28 µeV per atom, a figure flat in system size.
+- **A cliff at the Γ-point margin.** Crossing `DEFAULT_SHORT_RANGE_CUTOFF`
+  (14 Bohr) steps a water box's energy by **35 eV** between 7.4080 and 7.4085 Å.
+  Documented behaviour rather than a defect — below the margin the SCF converges
+  cleanly to a well-defined wrong answer — but the size of the step is not
+  obvious from that sentence, and a test fixture of this crate's own was sitting
+  on the wrong side of it. See `examples/cell_continuity.rs`.
+
+### Two Berry-phase conventions worth writing down
+
+Both were derived against this crate's own gauge rather than taken from a
+published form, because getting either wrong returns a plausible number rather
+than an error.
+
+- **No closure factor on the last link.** `bloch_fock` carries the phase on the
+  lattice translation alone, so `H(k + G) = H(k)` exactly and the coefficients at
+  `k₀ + G` *are* those at `k₀`. An extra `e^{−iG·τ}` put fluorine's Born charge
+  at `+21.8 e` against a true `−0.33`.
+- **The sign of the electronic term.** With `e^{−ib·τ}` in the overlap, the
+  textbook `−(e/Ω) φ a` counts that sign twice. Fixed by the single-orbital
+  limit; getting it wrong gives `+14.4`.
+
+With both right, the Berry and CPHF Born charges differ by `0.147 e` — and
+removing the intra-atomic `dd` moment the phase omits (`PM3_BORN_NO_DD=1`)
+collapses that to `1.9e-4`. The two formalisms agree on everything except one
+identified physical term.
+
+## 0.2.1
+
+Phonons everywhere the Γ point already worked, an external electric field, and
+the wavefunction outputs that go with it. Three defects that had shipped in
+0.2.0 are fixed; all of them were invisible at `q = 0`, which is where every
+test that could have caught them ran.
+
+### Added
+
+- **Analytic stress for a slab.** Every periodic dimensionality now reports one:
+  a chain its axis, a slab its two in-plane components, a crystal all nine, with
+  exact zeros in the non-periodic directions. `relax` relaxes the cell vectors
+  that exist and leaves a slab's vacuum thickness alone; only an isolated cell
+  refuses, having no strain rather than an underived derivative.
+- **A uniform external electric field** (`Pm3Options::field`), molecular only —
+  energy, analytic gradient and analytic Hessian. Validated against MOPAC's own
+  `FIELD=` keyword to all eight digits MOPAC prints. Refused under periodic
+  boundary conditions, where `−f·r` is not lattice-periodic.
+- **A shared dipole operator** (`pm3_rs::dipole`). The reported dipole and the
+  field coupling are the same matrix, which is what makes `μ = −∂E/∂F` hold by
+  construction rather than by coincidence.
+- **Molden wavefunction output** (`pm3_rs::molden`, the CLI's `molden` command,
+  `write_molden` in both Python layers). The coefficients are the raw ZDO ones,
+  which is MOPAC's own `VECTORS`/`GRAPHF` convention and makes the comparison
+  against it direct; the Slater functions are expanded in Gaussians derived at
+  run time by a regularized linear least squares rather than transcribed from a
+  table, to a measured overlap deficit of `2e-5` at worst.
+- **β orbitals on `Pm3Result`** (`mo_energies_beta`, `mo_coeff_beta`, `n_beta`),
+  which an unrestricted Molden file needs and which the UHF Hessian was
+  re-diagonalizing to recover.
+- **Infrared intensities** (`pm3_rs::ir`): the dipole-derivative tensor
+  `∂μ/∂R` from **three** coupled-perturbed solves rather than `3N`, by the
+  interchange theorem, with the nuclear term the electronic part alone would
+  miss; and from it the per-mode spectrum in km/mol, translations and rotations
+  projected out of the mass-weighted modes. RHF and UHF, both finite-difference
+  checked. The km/mol conversion is computed from its SI inputs rather than
+  transcribed, and checked against the published 974.88.
+- **Phonons for a spin-polarized cell.** The response is written over a list of
+  spin channels rather than as a pair of code paths: a closed shell is one
+  channel holding the total density at half exchange strength with two electrons
+  per state, an open shell two channels each holding its own at full strength
+  with one. Those are the same number when the spins are equal, which keeps the
+  restricted path at its old cost — one diagonalization per k-point, not two —
+  and pins it, every closed-shell value being unchanged. Validated against
+  central differences of the periodic UHF force, which shares none of the
+  response machinery.
+- **The classical corrections at finite `q`.** `D(q)` carries the D3/H4/X terms
+  instead of refusing them. Their contribution is the bilinear form
+  `Σ_{T,T'} e^{−iq·T} e^{+iq·T'} ∂²E_cell/∂x_{(κ,T)}∂x_{(κ',T')}`, obtained by
+  scaling each cluster entry's displacement by its own weight and taking a
+  forward-mode second derivative of the whole cluster energy — never of a pair,
+  so D3's coordination-number coupling and H4's donor–hydrogen–acceptor triples
+  come along. Four real evaluations per entry, `Dual2` being real.
+- **Every new feature reaches every API.** The external field, the phonons at a
+  wavevector, the band structure and the variable-cell relaxation are now
+  callable from `pm3_rs`, `pm3_rs.native`, the ASE calculator and (for the
+  field) the CLI's `--field`; `divide_and_conquer` gained the ASE accessor it
+  never had. The Python layers take the field in **volts per Angstrom**, the unit
+  MOPAC's own `FIELD=` keyword uses. `dipole`, `ir`, `molden` and `pbc::dfpt`
+  are re-exported from the crate root and pinned by `tests/api_surface.rs`,
+  which covered none of them.
+- **The divide-and-conquer SCF's memory is linear in the system.** The eight
+  `nao × nao` matrices the loop held — two densities, two Fock matrices, four
+  workspaces — live on the sparsity pattern now. Every one of them was already
+  read back only through that pattern, and the subsystem gather looks exactly
+  where some subsystem holds both orbitals, so nothing outside it was ever
+  consulted: no number changes, and `a_reaching_buffer_reproduces_the_full_result_exactly`
+  still holds to the bit. The Fock build writes into the pattern directly instead
+  of filling an array to have most of it ignored. Measured on a chain of waters,
+  8 → 128 molecules: dense grows 0.14 → 36.0 MiB, a log-log slope of exactly
+  2.00; on the pattern, 0.141 → 5.06 MiB, slope **1.06**.
+- **`pm3_rs::dipole` and `pbc::dfpt` reach Python.** `dipole(...)` returns the
+  operator, the centre of mass it is taken about, and the `3 × 3N` derivative
+  tensor — three coupled-perturbed solves, not `3N`, and no Hessian, which is
+  what separates it from `ir_spectrum`. `dynamical_matrix(...)` returns `D(q)`
+  itself rather than only the frequencies it is diagonalized to, with the
+  Hermitian defect it was assembled at. Both reach `pm3_rs`, `pm3_rs.native` and
+  the ASE calculator.
+- **The heavy ASE accessors are cached.** They were already lazy — nothing but
+  energy, forces, charges and dipole is computed in a `calculate` cycle — but
+  each recomputed on every ask, so a caller who wanted frequencies and then an
+  infrared spectrum paid for two Hessians and got no warning. Each now keeps its
+  most recent result against the geometry and parameters it was computed at, and
+  the phonon and dynamical-matrix caches are keyed on the wavevector and mesh as
+  well, so a dispersion sweep neither reuses the wrong answer nor accumulates
+  every matrix it built.
+- **A multipole tree for the isolated far field.** The point-multipole sum past
+  the 80 Bohr handover was quadratic in atoms; it now runs over an octree whose
+  nodes carry the combined moments of everything beneath them, so a distant group
+  is one term rather than many. The translation up the tree is exact — a node's
+  moments *are* its clouds', shifted — so the only error is the expansion's own
+  truncation, which a test checks against the definition rather than against the
+  recursion that produced it.
+
+  A node is accepted only when `d − s > MULTIPOLE_RADIUS`, so by the triangle
+  inequality every cloud inside it was already far by the pairwise rule and **the
+  near field is exactly the set it was**. The second condition is an *absolute*
+  error bound, not a Barnes–Hut opening angle: the pairwise rule it has to match
+  is absolute, and a fixed angle is far looser where it matters — `θ = 0.3` moved
+  the energy by 1.5 meV against a 50 µeV tolerance. Measured by counting terms
+  rather than by a clock, since this machine runs other work: over a block
+  growing 125 → 1000 clouds the far-field count grows with a log-log slope of
+  **1.58** against the pairwise sum's exact 2.00, doing 18% of the pairwise work
+  at the larger size.
+- **An independent check on the phased lattice sum.** `ewald_phased` shipped in
+  0.2.0 with four tests, all of which compare it against itself.
+  `ewald_reference::direct_phased_sum` compares it against a direct sum over
+  whole cells that shares none of its algebra — value, gradient and Hessian, to
+  better than `1e-6` at a generic interior wavevector.
+
+- **Phonons off Γ in one and two dimensions.** The phased lattice sum, the phased
+  Parry slab and the phased direct chain now all exist, so a slab or a wire has a
+  dynamical matrix at any wavevector in its periodic subspace — a component along
+  a non-periodic axis is refused rather than quietly summed. `slab_kernel` gained
+  its second `z` derivative (and its first unit test), and `ewald_atom_hessian`
+  covers 1D and 2D through the phased sum at `q = 0` instead of a second
+  implementation.
+
+### Fixed
+
+- **`D(q)` was not Hermitian**, by an amount that was exactly zero at `q = 0` and
+  grew linearly with `q` — a tenth of an eV/Bohr² out of twenty-five by
+  `q = (0.2, 0, 0)`. The first-order density's contribution to the multipole
+  charges read only the lower triangle and doubled it. That is the same number as
+  the symmetrized form for a ground-state density, which is real and symmetric,
+  and a different number for a complex first-order one; it also stopped the
+  charge map being the adjoint of the potential map, which is what Hermiticity
+  rests on. Every identity test in the file ran on the rigid-ion matrix, so none
+  of them could see it.
+- **The response solve diverged at a general wavevector and said nothing.** It
+  reached `1e30` in its two hundred passes and returned what it was holding. Two
+  changes: it now refuses rather than returns, and it extrapolates over a history
+  instead of substituting the equation into itself. The response is linear, so
+  the plain iteration converges only where the spectral radius of `χ₀K` is below
+  one, and at a general `q` it is not; damping only rescales that eigenvalue.
+  Extrapolation solves the linear system on the Krylov subspace the history
+  spans, which does not care. Every wavevector probed now converges to `1e-10`.
+- **The response summed the time-reversal-irreducible mesh.** The ground state may
+  — `P(−k) = P(k)*` — and the response may not: time reversal maps the coupled
+  pair `(k, k+q)` onto a pair at `−q`, so doubling the irreducible weights is the
+  wrong sum wherever `q ≠ −q`. The response now regenerates the full mesh, while
+  still sharing the potential the reduced SCF converged.
+- **The external field was half-wired into three paths.** The open-shell
+  skeleton (`skeleton_fock_ov_spin`) never carried the field's first derivative,
+  so a UHF Hessian in a field came back symmetric, with six near-zero modes, and
+  short by the cross term `Tr[(∂P/∂R)(∂H'/∂R)]` — measured at `3e-3 eV/Bohr²`
+  for the methyl radical, sixteen times the finite-difference tolerance. The
+  divide-and-conquer energy omitted the field's nuclear half `−Σ_A Z_A R_A·f`,
+  and its gradient omitted `−q_A f`. `analytic_gradient` dropped the field
+  entirely. The screened long-range path and the periodic path now refuse a
+  field rather than ignore one — including on an isolated cell, where the field
+  is well-defined but that machinery still does not carry it.
+- **Every periodic image was given the reference cell's coordination-number
+  response.** `cn[image] = cn[parent]` is exact at `Γ`, where all copies move
+  alike, and wrong under a phased displacement, where the image's neighbours
+  move by different amounts than its parent's do. It was therefore invisible to
+  every test that existed: at `q = 0` it is not an approximation. An image is
+  now given its own coordination number wherever the cluster is wide enough
+  around it for that number to be right — a triangle-inequality test against the
+  cluster radius — and the parent's beyond. Restoring the blanket copy fails the
+  supercell folding test by four thousand times its tolerance.
+- **The phased reciprocal sum carried a conjugated phase**, `e^{+i(G+q)·d}` where
+  Poisson summation gives `e^{−i(G+q)·d}`, which made `D(q)` wrong in its
+  imaginary parts at a generic interior `q` in 3D. Every existing test was
+  structurally blind to it: the oracle contraction is identically real over `±d`
+  pairs, the folding test's `q = b/2` makes the shifted set negation-symmetric,
+  and the internal identities conjugate both halves together. Pinned now by the
+  α-independence of `Im Φ_q` and by `Φ_q(d+T₀) = e^{−iq·T₀} Φ_q(d)`.
+- **The shipped Python is now ASCII.** Sixteen docstrings could not be printed on
+  a legacy-codepage console: the package imported and computed correctly, and
+  `help()` on it raised `UnicodeEncodeError` on Japanese, Chinese or Korean
+  Windows. Installing was never affected — verified by building the sdist and
+  installing it under a forced `cp932` locale — but a documented API surface was
+  unusable. Two tests keep it that way.
+- **The divide-and-conquer radii disagreed between layers by a factor of 1.9.**
+  `DcOptions` defaults to 6.0 and 9.0 *Bohr*; the Python signature took Ångström
+  and defaulted to 6.0 and 9.0, so a caller who omitted the argument silently got
+  subsystems nearly twice the intended size. The CLI and `pm3_rs.native` were
+  right; the extension's own defaults are now the same radii.
+- **`long_range_cutoff` was unreachable from `import pm3_rs`** — the whole
+  linear-scaling path existed and could not be switched on.
+- **`phonons` alone had no `reference` argument**, so an open-shell cell could
+  not be asked for. ASE's forwarding is now by keyword, since adding the argument
+  immediately exposed a positional call handing `method` to `reference`.
+- The type stub declared `optimize(max_iter, gtol)`, arguments that never
+  existed. A test now compares every stub signature against the extension, and
+  another asserts `pm3_rs.native` can pass every argument the extension accepts.
+
+## 0.2.0
+
+Periodic boundary conditions. Molecular results are unchanged except for one
+bug fix, noted below.
+
+### Added — periodic boundary conditions
+
+- `Cell` on `Molecule`, covering 1D chains, 2D slabs, and 3D crystals. The
+  non-periodic directions never enter the measure, the reciprocal basis, or the
+  stress, so a slab's vacuum thickness cannot affect a result.
+- Γ-point periodic SCF (`pbc::gamma::run_gamma`), RHF and UHF, with analytic
+  forces and analytic stress (`pbc::gradient::periodic_gradient`) and
+  fixed- or variable-cell relaxation (`pbc::optimize::relax`).
+- Ewald electrostatics with the dimension dependence confined to the reciprocal
+  sum: 3D tinfoil, exact 2D Parry slab, a cell-grouped 1D direct sum, and a
+  plain owner-excluded pair sum at zero dimensions. Charged cells are supported
+  in every dimensionality — a uniform neutralizing background in 3D and 2D, a
+  neutralizing line charge in 1D — and the neutralizer's potential enters the
+  Fock matrix rather than only correcting the energy afterwards.
+- D3, H4, X, and the simple hydrogen-bond correction are lattice summed,
+  including the D3 coordination number over images.
+- `PeriodicResult::gamma_margin` reports the Γ-point validity condition — see
+  `docs/pbc.md`. A cell narrower than the exchange cutoff converges cleanly to
+  an answer that is wrong by tens of eV, and nothing else reveals it.
+- **k-point sampling** (`pbc::kscf::run_kpoints`), RHF and UHF, neutral and
+  charged: Γ-centred Monkhorst–Pack meshes with exact time-reversal reduction,
+  a global Fermi level with optional Fermi–Dirac smearing, fixed or free
+  magnetization, and band structures along a path. Sampling more than one `k`
+  is what lets `P(0, T)` decay with `T`, which lifts the Γ-point cell-width
+  condition above.
+- New `cmatrix` module: complex Hermitian matrices and eigensolver. ZDO makes
+  `S(k) = I`, so no generalized eigenproblem is needed.
+- **Γ-point analytic Hessian and phonons** (`pbc::hessian`), with the Ewald
+  second derivative (`pbc::ewald_hessian`) and the periodic CPHF response. The
+  acoustic sum rule holds before enforcement, so enforcement cleans up rounding
+  rather than hiding a misplaced term.
+- **Divide and conquer** (`dc`), molecular and Γ-point periodic, RHF and UHF:
+  Yang–Lee partitioning with a single global chemical potential, plus forces and
+  stress from the partitioned density.
+- **Zero dimensions as a member of the same family** (`Cell::isolated`). An
+  isolated system is the case with no images, where the lattice sum is a plain
+  owner-excluded pair sum. Running a molecule through the periodic path gives it
+  the crystal's `O(N)` near field — neighbour-list pair tables plus a
+  point-charge model outside the cutoff — instead of a dense `O(N²)` pair cache.
+  It reproduces molecular PM3 to 3 µeV per atom with the default switch, and to
+  `1e-8` eV total with the switch pushed past the molecule; the error is
+  intensive, staying put from 12 to 288 atoms rather than accumulating.
+- **A linear-scaling near field for molecular divide and conquer**
+  (`DcOptions::long_range_cutoff`, `None` by default). Same split at zero
+  dimensions. Measured at 960 atoms: 7.0× faster than full diagonalization
+  against 3.2× for the dense path, with the log-log slope down from 2.03 to
+  1.27. Left off by default because the switch costs a measured 28 µeV per atom,
+  about a hundred times the divide-and-conquer truncation at the default buffer.
+- **A multipole far field for isolated systems.** Beyond 80 Bohr two atoms now
+  interact through their net charge, dipole and second moment rather than through
+  all six hundred and twenty-five of their auxiliary site pairs, with the
+  potential carried back out to the sites by a Taylor expansion of matching
+  order. The self-consistent field also stops computing site gradients it never
+  reads. Together these took the Fock build at 960 atoms from 2.52 s to 1.22 s
+  and the whole run from 4.9 s to 3.0 s, with no measurable change in the answer
+  — the divide-and-conquer error stays at 28.2 µeV/atom, and `pbc::ewald`'s own
+  test puts the collapse at 8 µeV over forty-eight atoms.
+  `PM3_DC_PROFILE=1` prints the loop's split into Fock build, subsystem solves
+  and density assembly.
+- **Analytic stress in 1D.** 1D is summed directly rather than through
+  reciprocal space, so its virial is the ordinary pair virial and comes free
+  with the gradients. Only the axial component exists; the projection onto the
+  strains a cell actually has now happens once, for every dimensionality, rather
+  than being left to each sum.
+- **Charged 1D cells**, at Γ and at any k mesh, through a uniform neutralizing
+  line charge. The subtraction is in terms of the physical extent summed rather
+  than the image count — `H_N + ln(L/L₀)` with `L₀` a fixed reference — which is
+  what makes it size-consistent. `Q² H_N / L` alone gives each cell a stable,
+  plausible energy while a chain and its own doubled cell disagree by eV.
+- **Sparkles, point atoms and `d` shells in the periodic path.** An atom with no
+  orbitals has no electronic multipoles, so its realization is the nucleus it
+  already has; a `d` element's realization comes from MNDO-d's own multipole
+  table, which has only `l ≤ 2` and so is carried in full rather than truncated.
+  (PM3 itself parameterizes all forty-two of its elements on an s/p basis, so
+  the `d` path is unreachable through the standard tables and its test builds
+  its own subject.)
+- **DFPT at arbitrary `q`** (`pbc::dfpt`), with the phased lattice sum it needs
+  (`pbc::phased`): the dynamical matrix and phonon frequencies at any
+  wavevector, from the primitive cell, at a cost that does not depend on `q`.
+  The electrons' response is included, and it is most of the answer — on water
+  it turns a rigid-ion force constant of 5.96 eV/Bohr² into 0.56.
+  `rigid_ion_dynamical_matrix` gives the fixed-density part alone.
+
+  Validated at two levels. `D(0)` against `pbc::hessian::periodic_hessian`, an
+  independent implementation with no phases in it, itself checked against finite
+  differences; and `D(q)` by folding, a doubled cell's Γ-point force constants
+  holding the primitive cell's `D(0)` and `D(zone boundary)` between them. Each
+  constituent is separately checked against something built differently, which
+  is what made three real bugs findable rather than merely visible: a lattice
+  sum that skipped same-atom images, a phased kernel with no Ewald self term,
+  and an exchange handed a total density where it wanted a spin one.
+
+  The response samples one k-point, `Γ`, paired with `q` — the sampling the
+  ground state used, carrying the same `gamma_margin` condition. 3D, closed
+  shell, plain PM3.
+- New `docs/pbc.md` and `docs/divide-and-conquer.md`.
+
+### Added — interfaces
+
+- Python: `periodic_single_point`, `periodic_forces`, `phonons` and
+  `divide_and_conquer`, all taking a cell in Ångström.
+- ASE: the calculator switches to the periodic path from `atoms.pbc`, adds
+  `stress` to `implemented_properties` (6-component Voigt, eV/ų) and gains
+  `get_phonons`. A molecule or a slab raises on `get_stress()` rather than
+  returning zeros, because zeros would be a claim rather than an absence.
+- CLI: `--cell`, `--pbc`, `--kpts`, `--dc`, `--dc-core`, and the `stress`,
+  `phonons` and `bands` subcommands. A Γ-point run that violates the validity
+  condition prints a warning naming it.
+- Packaging: PEP 639 licence metadata, classifiers, project URLs, keywords,
+  a `py.typed` marker and `_native.pyi` stubs, and GitHub Actions workflows for
+  CI and for wheels published through PyPI Trusted Publishing.
+- **`pip install` puts the `pm3-rs` command on your path.** The CLI moved out of
+  `src/bin` and into the library, taking its argument vector rather than reading
+  the process environment, so a PyO3 wrapper can hand it `sys.argv`. `pip` and
+  `cargo install` give the identical interface rather than two that drift.
+  Verified by installing the source distribution into a clean virtualenv,
+  compiling from scratch, and running the installed command.
+
+### Fixed
+
+- **XYZ files with a byte-order mark.** Notepad and PowerShell's
+  `-Encoding utf8` both put one in front of the atom count, which produced
+  `invalid XYZ atom count: 3` on a file whose first line was visibly `3`. A
+  leading mark is now stripped; a second one is still an error, because that is a
+  malformed file rather than a Windows editor.
+- **Charged-system dipole origin.** The dipole was referenced to the coordinate
+  origin rather than to the centre of mass, so for any system with a net charge
+  it depended on where the molecule sat: translating NH₄⁺ by 5 Å moved its
+  dipole from 0 to 24.02 D. It now matches MOPAC's `dipole.F90` (centre of mass,
+  with `+`/`−` point atoms carrying zero mass) for all 60 oracle cases to ~1e-6 D.
+  Neutral systems are unaffected, the dipole being origin-independent there.
+- `erfc` was computed as `1 − erf`, which cancels catastrophically: 1.16e-8
+  relative error at `x = 3.9`. It now uses a continued fraction above `x = 2`,
+  giving 3.9e-14 against scipy over 157 points. (New code only; no molecular
+  result used `erfc`.)
+- The DIIS coefficient solve normalizes its Gram matrix before the pivot test.
+  The threshold was absolute while `⟨E_i,E_j⟩ ~ ‖E‖²` shrinks quadratically, so
+  a converging run could be handed a matrix of ~1e-10 entries and get back
+  wildly amplified coefficients — DIIS converging nicely and then walking back
+  out into a limit cycle. The coefficients are invariant under the scaling, so
+  no converged result changes.
+
+- **The Γ-point periodic energy belonged to no state.** It was reported as
+  `½(P·H + P·F)` with `F` the *extrapolated* Fock — a combination of history
+  matrices — paired with the freshly diagonalized density, while the density
+  actually returned was the damped and extrapolated one. Three different
+  objects. The convergence test could not see it: a stable set of DIIS weights
+  makes a wrong energy stop moving as convincingly as a right one. Worst case
+  measured, 1.03 eV on a chain of 92 water molecules; usually far below
+  tolerance, which is why it survived. The k-point path was checked for the same
+  defect and does not have it — its accelerator extrapolates the density, not
+  the Fock.
+- **The `B` auxiliary integrals lost seven digits just above `|x| = 0.5`.**
+  MOPAC's closed-form recursion multiplies its cancellation by `k/|x|` at every
+  step, and its power series stops at `0.5`; at `x = −0.51`, `B₉` was off by a
+  relative `5e-7` against the integral itself. Below `1e-6` the `x → 0` branch
+  returned constants, so a differentiating scalar got a zero derivative where
+  `dB₁/dx = −2/3`. Both go away by summing the defining series directly out to
+  `|x| = 3` — absolutely convergent, every term independent of every other,
+  exact at `x = 0` including its derivative — and the `x → 0` special case
+  disappears rather than being widened. Every frozen MOPAC value is unchanged.
+- **The Slater overlap went NaN beyond about 500 Bohr.** `A_k` carries
+  `e^{−r(ζ_a+ζ_b)/2}` and `B_k` carries `e^{+r|ζ_a−ζ_b|/2}`; the overlap is their
+  product, which is tiny, but the two factors separately are `0` and `∞`, and
+  `B` overflows first. For an O–H pair that happens at 501 Bohr, and `0 · ∞` then
+  propagated silently through `H_core`. This affects **any** calculation
+  containing a 500 Bohr separation, not only a periodic or partitioned one; it
+  surfaced here because divide-and-conquer was the first thing run on a system
+  that large. The overlap is now cut where its own decaying factor is below
+  `1e-130`, so no overlap that could matter is affected and every frozen MOPAC
+  value is unchanged.
+
+### Performance
+
+- **The periodic SCF now uses the molecular path's accelerator.** Plain CDIIS
+  interpolates the Fock with unconstrained weights and has no notion of the
+  energy going down, so only damping held it and how much damping is enough
+  grows with the system: on a chain of identical waters it converged at 56
+  molecules, failed at 58, converged again at 66, then failed at every size
+  beyond. Restricted periodic runs now use A-DIIS until the commutator is small
+  and CDIIS after, sharing `crate::scf`'s history rather than a second copy.
+  Every size converges, in 25 iterations where damped CDIIS needed 110.
+- **The Ewald sum caches everything that depends on the geometry** — the
+  reciprocal enumeration, the neighbour list, and the `cos(G·r)`/`sin(G·r)` of
+  every site against every `G`. An SCF changes only the charges. Capped at
+  256 MiB, above which the trigonometry is recomputed.
+- **Divide and conquer no longer does `O(N²)` work on its own structural
+  zeros.** A partitioned density is zero wherever no subsystem holds both
+  orbitals — the approximation the method makes, not a rounding effect. DIIS,
+  damping, the RMS change and the energy trace now run over a recorded sparsity
+  pattern, and every matrix the loop touches is allocated once. At 960 atoms the
+  DIIS history alone fell from 7.17 s to 0.19 s and the whole run from 25.5 s to
+  6.5 s, with the converged energies unchanged to every printed digit.
+- **The ASE calculator computes forces unasked.** ASE requests one property at a
+  time and re-enters `calculate` for each, so naming only what was asked
+  converged the same SCF two or three times per MD step. With the Ewald cache, a
+  periodic single point went from 250 ms to 53 ms and an MD step from ~520 ms to
+  74 ms.
+- Periodic SCF: CDIIS on the `[F, P]` commutator, replacing plain damping
+  (about 38 iterations → a dozen, each costing a full lattice sum).
+- Periodic UHF starts from core-Hamiltonian orbitals occupied to the two aufbau
+  counts rather than a spin-scaled atomic-density guess. The scaled guess starts
+  inside the spin-symmetric subspace, which for the methyl radical contains a
+  stationary point 4.8 eV above the UHF minimum: damping needed ~25 wasted
+  cycles to escape it and CDIIS converged straight onto it. Methyl in a cell
+  now takes 9 iterations rather than 96, and reaches the correct solution.
+- The Ewald reciprocal sums enumerate one member of each `±G` pair and double.
+  Every quantity they produce is even under `G → −G`.
+
+## 0.1.2
+
+Performance and memory release. No change to any computed PM3 quantity: the
+MOPAC v23.2.5 oracle regressions (heats of formation, charges, gradients,
+optimized geometries, frequencies, special atoms, Sparkles) are unchanged.
+
+### Performance
+
+- The two-center Fock build — the `O(N²)` part of every SCF and CPHF iteration —
+  now runs batched-parallel over atom pairs instead of in a single serial loop.
+- The two-center Coulomb term is contracted as two packed mat-vecs over the
+  `(μν)`/`(λσ)` orbital-pair indices rather than a four-index loop: 100 instead
+  of 256 multiply-adds per sp/sp pair, with identical arithmetic.
+- The SCF accelerator keeps its `⟨E_i,E_j⟩` and `⟨D_i,F_j⟩` Gram matrices
+  incrementally, evaluating only the new row and column each iteration. The
+  A-DIIS difference matrices `D_i − D_n` / `F_j − F_n` are no longer
+  materialized, removing `2 × depth` full `nao × nao` temporaries per iteration.
+- `[F,P]` is formed with one matrix product instead of two (`F` and `P` are
+  symmetric, so `PF = (FP)ᵀ`), and is skipped entirely when no accelerator runs.
+- The two-electron pair table is a single flat buffer instead of a `Vec<Vec<_>>`,
+  removing one heap allocation per packed row and making each row contiguous.
+- The classical-correction (D3/H4/X) Hessian off-diagonal loop runs on rayon.
+- Elementwise reductions (`frobenius_dot`, RMS density change, history
+  combination) go parallel above 2^18 elements.
+- `symmetric_eigen` skips building a permutation when `faer` already returns
+  ascending eigenvalues.
+- 900-atom water cluster (nao = 1800), 16 cores: 35.7 s → 17.7 s wall.
+
+### Memory
+
+- The pair cache no longer retains the electron–core attraction blocks
+  `e1b`/`e2a` (2 × 81 `f64` per pair). They are consumed while `H_core` is
+  assembled and never referenced again; holding them cost 1.3 KiB per pair
+  (3.7 GiB for a 2400-atom system) for no benefit.
+- New `Pm3Options::integral_memory_mb` (env `PM3_MAX_PAIR_CACHE_MB`, default
+  4096 MiB): the `O(N²)` pair cache size is computed in closed form before the
+  allocation and reported as a `ResourceLimit` error if it exceeds the budget.
+- New `Pm3Options::scf_memory_mb` (default 512 MiB): bounds the SCF accelerator
+  history. The DIIS depth is reduced to fit, and A-DIIS degrades to CDIIS (which
+  needs no density history) rather than exceeding the budget.
+- 900-atom water cluster peak resident set: ≈ 1.4 GiB → 0.75 GiB.
+
+### Documented-API verification
+
+Every code block and stated guarantee in `README.md`, `docs/rust-api.md`, and
+`docs/python-api.md` is now executed as a test, and the CLI is exercised over
+every documented subcommand, flag, and error path.
+
+- `tests/api_surface.rs` (new, 15 tests) — the Rust surface: constructors,
+  `Pm3Options` fields and defaults, the `Pm3Result`/`GradientResult`/
+  `VibrationalModes`/`OptResult` field lists, unit conventions, `Variant::parse`,
+  reference selection, the memory budgets, and the documented error variants.
+- `tests/test_python_api.py` (new, 19 tests) — `pm3_rs.native` dict keys and
+  Hartree↔eV / Bohr↔Å conversions, the `pm3_rs.ase.PM3` calculator (ASE units,
+  lazy Hessian, every accessor), and both documented example blocks.
+
+Fixed along the way:
+
+- **`pm3_rs.ase.PM3`**: every accessor (`get_potential_energy`, `get_forces`,
+  `get_gradient`, `get_hessian`, `get_frequencies`) raised
+  `AttributeError: 'NoneType' object has no attribute 'get_atomic_numbers'` when
+  called with no argument before the calculator had been bound to a structure.
+  They now raise a `RuntimeError` that names the three ways to fix it.
+- `docs/python-api.md`: the cation example was labelled "forced UHF doublet" but
+  passed `multiplicity=1` (NH4+ is a closed-shell singlet), and its
+  `nh4_positions` was never defined, so the block could not be run as printed.
+  Corrected, given a runnable geometry, and extended with an open-shell example.
+- `docs/python-api.md`: documented what `atoms=None` means for the ASE
+  accessors — assigning `atoms.calc = PM3()` does not bind the structure.
+
+### Documentation
+
+- Removed the claims that this crate is derived from or structured after sibling
+  Rust projects; the PM3 Hamiltonian, parameters, and derivations are documented
+  against MOPAC v23.2.5 and the primary literature.
+- `third_party/pyseqm/NOTICE` added — `THIRD_PARTY_NOTICES.md` referenced a
+  `third_party/pyseqm/LICENSE` that was not present.
+- Corrected the MOPAC reference values quoted in the `scf.rs` unit-test comments
+  for water and the methyl radical; they did not match the asserted PM3 values.
+- `docs/rust-api.md`: documented the three memory budgets and corrected the
+  description of the Hessian `step` argument.
+
+## 0.1.1
+
+Initial `pm3-rs` release, backed by the PM3 Hamiltonian and the MOPAC v23.2.5
+PM3 parameter tables.
+
+- Rust implementation; linear algebra via `faer`, no BLAS/LAPACK dependency.
+- RHF/UHF energies, heats of formation, charges, dipoles, analytic gradients,
+  CPHF/UCPHF Hessians, optimization, and frequencies.
+- PM3-D3, PM3-D3H4, and PM3-D3H4X correction variants.
+- MOPAC special atoms `Cb`, `+`, `-`, and La-Lu trivalent Sparkles.
+- Rust library `pm3_rs`, CLI `pm3_rs_cli`, and Python distribution
+  `pm3-rs-python` with native and ASE APIs.
+- MOPAC v23.2.5 oracle regressions for closed-shell molecules, an open-shell
+  radical, a Sparkle complex, point charges, gradients, optimization, and
+  frequencies.
+- Configurable Hessian workspace budget and bounded integral-cache construction.
+- Release profile uses fat LTO with one code-generation unit.