v0.1.0rc9

What's Changed

• test(pbc): p/d-shell and Li binding anchors for the wired derivative routes (Refs #498) by @mgoldey in #561
• feat(cc): integral-direct (RI) particle-particle ladder for closed-shell CCSD, dense when it fits ([cc] ladder = auto) by @mgoldey in #568
• fix(rpa): the Lanczos non-convergence warning names the real cause by @mgoldey in #570
• fix(scf): apply an unapplied cDFT constraint instead of returning the unconstrained state by @mgoldey in #569
• feat(scf): point-group-blocked closed-shell RKS, with Python and CLI (Closes #294) by @mgoldey in #567
• fix(rpa): Lanczos stall warning records which limit stopped it; no max_vecs advice for a full request by @mgoldey in #571
• fix(ecp): a ghost atom carries basis functions, not an ECP by @mgoldey in #572
• fix(rpa): solve a Lanczos request above the Krylov cap densely so it converges by @mgoldey in #573
• test(scf): port HeNe+ asymmetric cDFT-ET tests, re-measured on main by @mgoldey in #574
• ci: build the Python extension once in the coverage and validation jobs (fixes nightly 'Object is too small') by @mgoldey in #575
• fix(scf): cDFT-ET coupling uses the exact cofactor sum, not an S_TOL branch by @mgoldey in #576
• exp(pbc): shifted k mesh for the #230 parity test, result UNDETERMINED by @mgoldey in #577
• fix(rpa): solve every full Lanczos request densely; trunc_thresh 0 keeps exact-λ=1 modes by @mgoldey in #578
• feat(pbc): open-shell k-point SCAN / r2SCAN on RS-GDF J/K: energy, force and stress (library) by @mgoldey in #579
• fix(rpa): one v_inv_sqrt convention for the PDEP parts entries (gamma_urpa lindep panic) by @mgoldey in #580
• ci(test-shard): long-window zstd shard archives (~6x smaller), plus the #563 review nits by @mgoldey in #583
• feat(pbc): k-point ROKS SCAN / r2SCAN energy via the per-spin-tau evaluator (library) by @mgoldey in #582
• fix(cli-tests): give each ccsd_dispatch run its own TOML file by @mgoldey in #585
• feat(pbc): k-point ROKS SCAN / r2SCAN analytic force and stress (library) by @mgoldey in #586
• feat(dft,gpu): KS XC grid on the device, with a mixed-precision V_xc kernel (xc-vxc) by @mgoldey in #584
• ci: say when a shard's archive has expired (re-run ALL jobs), keep archives 2 days by @mgoldey in #587
• fix(rpa): cancellation-free trace-log summand near λ→1 by @mgoldey in #531
• fix(gate): flock -o so sccache cannot inherit the cargo lock by @mgoldey in #591
• feat(dlpno): opt-in add-back of screened pair energies; archive the LNO lane by @mgoldey in #530
• bench(pbc): unattended quiet-window driver for the remaining #228 plan (no timings) by @mgoldey in #589
• perf(pbc): measured, parallel, thread-count-invariant k-point XC build by @mgoldey in #590
• fix(gate): start the step timeout after the lock is held; report lock-wait expiry, timeout and failure separately by @mgoldey in #593
• feat(pbc): wire open-shell SCAN / r2SCAN energies (k-ROKS, Gamma UKS/ROKS, k-UKS RS-GDF) into Python and the CLI by @mgoldey in #588
• feat(mp2): scf_jk = exact | rijk | rijcosx for the MP2-family reference SCF by @mgoldey in #595
• exp(pbc): FIQ Stage 1 for the k-MP2 head residual (Refs #230) by @mgoldey in #598
• feat(python): AoIntegrals + einsum, building blocks for methods in numpy by @mgoldey in #594
• feat(basis): register every bundled set, add Pople/def2/pcseg/aug-JKFIT, pack them deduplicated by @mgoldey in #596
• feat(scf): Q-Chem-style BASIS2 secondary-basis SCF guess by @mgoldey in #597
• fix(scripts): all-or-nothing JKFIT import; --check flags an unapplied 6-311G override by @mgoldey in #599
• ci: 8 test shards; move >=120 s and thread-invariance tests to the weekly tier by @mgoldey in #600
• perf(gpu,dft): pinned staging + parallel operand scan for the device XC waves (bit-identical) by @mgoldey in #602
• perf(lmp2): stages A+B — stage timers, certificate, one-Cholesky VV-HV, metric bitmap, planned PCG by @mgoldey in #601
• perf(gpu,dft): strided-batched GEMMs for every XC wave, f64 and mixed, over a padded layout (stacked on #602) by @mgoldey in #606
• perf(lmp2): stage C — Cholesky + triangular-solve pair fits (stacked on #601) by @mgoldey in #604
• feat(scf,gpu): shared precision ladder with an f64 exit for the dfj-pack / dfk-occ J/K kernels by @mgoldey in #607
• ci(gate): scope push-gate clippy to the changed crates by @mgoldey in #608
• feat(precision): [precision] section and per-result precision provenance by @mgoldey in #603
• perf(lmp2): stage D — Hylleraas PCG energy + adaptive Fvv product (stacked on #604) by @mgoldey in #605
• docs(lmp2): sympy proofs of the WSHG23 Hylleraas energy and its error bound by @mgoldey in #609
• fix(lmp2): running out of iterations while tightening is a hard error by @mgoldey in #610
• test(lmp2): optional bitwise RHF cache for the direct bench by @mgoldey in #612
• fix(df): hold spilled RI-JK to the budget; a DF guess gives way on memory exhaustion by @mgoldey in #611
• feat(tools): box lease: a timing job holds the build locks and cargo waits by @mgoldey in #613
• feat(cli): log local MP2 stage timings in the JSON result record by @mgoldey in #616
• fix(tools): keep the cargo shim where a reboot does not wipe it by @mgoldey in #614
• feat(scf): SCF checkpoint format and identity checks, unused so far by @mgoldey in #618
• perf(lmp2): aux-major stage 3, one evaluation per needed triple by @mgoldey in #615
• fix(scf): decide RI-J/K before XC takes memory by @mgoldey in #617
• perf(dft,scf): faster Becke weights, batch screening and RI-J fold, same bits by @mgoldey in #619
• Memory opt for DFT by @mgoldey in #620
• feat(gpu): opt-in dfk-occ with device f64 ladder builds and a one-way latch; faster DF-K dressing and first build by @mgoldey in #621
• feat(dft): pruned default KS grid, DefGrid, weight cutoff, f32 RI-J, Split-RI-J by @mgoldey in #623
• perf(split-rij): batched kernel and far field; jk_storage = auto uses Split-RI-J when J does not fit by @mgoldey in #624
• fix: second CodeRabbit round on the grid and RI-J options (#623) by @mgoldey in #625
• truncate slow tests by @mgoldey in #626
• Correctness batch: XC density floor, meta-GGA response, PCM gradient, homogeneous response matvecs, certificates by @mgoldey in #627

Full Changelog: v0.1.0rc8...v0.1.0rc9

v0.1.0rc10

What's Changed

• fix(wheels): ship crates/ferric-xtb in the sdist by @mgoldey in #628

Full Changelog: v0.1.0rc9...v0.1.0rc10

ferric 0.1.0rc8

ferric 0.1.0rc8.

Files

• PyPI wheels and sdist (ferric-0.1.0rc8-cp3NN-...whl, ferric-0.1.0rc8.tar.gz):
  the same files pip install ferric==0.1.0rc8 downloads. Their libint2 is
  ferric's small 2.7.2 export, linked statically.
• Large-libint wheels (ferric-0.1.0rc8-1largeam-cp3NN-...whl): the same version
  and commit, built against conda-forge's libint2 2.13.1, which is bundled in
  the wheel as a shared library. It adds l = 7 shells, analytic 3- and
  2-centre second derivatives, multipoles above order 2 and G12 (F12)
  integrals. It is not a strict superset: its one-electron integrals go to
  l = 7 / 6 / 3 (energy / 1st / 2nd derivative) where the PyPI wheel's go to
  6 / 6 / 6. Analytic 4-centre Hessians stay limited to f functions in both
  wheels, and RI Hessians already work with the PyPI wheel. Install it with
  pip install --force-reinstall --no-deps <asset URL>; pip install -U ferric later replaces it with the PyPI wheel. Check which one is loaded
  with python -c "import ferric; print(ferric.build_info()['libint_max_am'])"
  (eri3 is [7, 7, 4] for this wheel, [6, 6] for the PyPI one).
• libint2-LICENSE.txt, libint2-COPYING.LESSER.txt: the licence texts of
  the libint2 the large-libint wheels bundle, from its conda-forge package.
• SHA256SUMS: sha256sum -c SHA256SUMS checks every file above.

libint2 source (LGPL-3.0-only)

The large-libint wheels contain a libint2 binary built by conda-forge. Its
source and build recipe:

• generated library source: https://github.com/loriab/libint/releases/tag/v0.1
  (libint-2.13.1-7-6-3-7-7-4_mm10f12ob2_0.tgz, sha256
  324b578f795d1f406f37a1d936f04ea51a85eae1b2ea37b5cf773afb4e25ab14)
• conda-forge recipe: https://github.com/conda-forge/libint-feedstock
• the exact package bundled: https://conda.anaconda.org/conda-forge/linux-64/libint-2.13.1-h83f5b4b_0.conda
  (sha256 d5219aae3bdd9629c6cff471f129f3aebbaddb058d0d19fbbab12c28924a0535)

The PyPI wheels' libint2 export is regenerated by
scripts/generate-libint-small.sh; its source tarball is the
libint-2.7.2-ferric-small-1 release of this repository.

v0.1.0rc7

What's Changed

• docs(scf): TRAH stall docs describe the current fix, not a refuted mechanism by @mgoldey in #221
• fix(hirshfeld): Python and CLI share the SCF proatom; Hirshfeld validation row by @mgoldey in #222
• fix(omega): tune both states with exact J; RSH omega-tuning validation row by @mgoldey in #223
• fix(scf): UKS TRAH defers to DIIS instead of cycling null steps by @mgoldey in #224
• fix(cc): wB97X-L-V puts λ inside the amplitude equations (paper eqn 22) by @mgoldey in #239
• test(rpa): PDEP-RPA dynamic α and C6 validation row (154) by @mgoldey in #240
• fix(scf): libint ERI precision 1e-20 for SCF J/K (was 1e-14) by @mgoldey in #226
• test(rpa): TS C6 and MBD@TS validation rows (153, 156) against pymbd by @mgoldey in #241
• test(scf): COSX validation row (175) — same-grid vs PySCF sgx, grid convergence, gradients by @mgoldey in #243
• test(cc): ωB97X-L-V component validation row (106) by @mgoldey in #242
• feat(scf): RHF internal stability descent; SCF ladder validation row (179) by @mgoldey in #244
• test(scf): free-atom Becke volumes for Z = 1–18 (row 152) by @mgoldey in #245
• test(scf): cDFT state-selection row (171); the 0.563 eV gap was a Becke-partition offset by @mgoldey in #246
• fix(scf): cDFT-ET coupling uses F = E + λN (was 3x high); validation row (170) by @mgoldey in #247
• fix: lint passes on Rust 1.99 (fetch_update deprecation, redundant rustdoc links) by @mgoldey in #250
• fix(scf): safeguard the single-constraint cDFT λ loop; correct the #225 measurement by @mgoldey in #248
• fix(scf): TRAH handles the trust-region hard case and keeps control of pending steps; unbreak complexity gate by @mgoldey in #249
• feat(rpa): Becke per-atom α is the Krishtal intrinsic polarizability; export the CT remainder by @mgoldey in #251
• fix(rpa): free-atom Pd uses the closed-shell 4d¹⁰ polarizability and the TS C6 by @mgoldey in #253
• feat(cli): [scf] stability_descent for kind = "rhf" and closed-shell ksdft by @mgoldey in #252
• chore(scf): FERRIC_TRAH_RHO_TRACE goes through ConfigVar by @mgoldey in #254
• feat(python): run_rhf(stability_descent=True) by @mgoldey in #255
• fix(rpa): sternheimer guard tests pass in release builds by @mgoldey in #257
• feat(dispersion): MBD@rsSCS energy, KS-DFT correction, exact gradient via KS Z-vector by @mgoldey in #256
• fix(dispersion): free-atom volumes for the isolated atom; validate HSE06 by @mgoldey in #258
• feat(properties): smooth (C2) Hirshfeld proatom interpolant by @mgoldey in #259
• feat(frequencies): harmonic frequencies with D3(BJ) and MBD@rsSCS dispersion by @mgoldey in #260
• feat(scf): unrestricted KS Z-vector; exact MBD@rsSCS gradient on UKS by @mgoldey in #262
• feat(scf): RIJCOSX, pruned sgx COSX grid, and a COSX final-grid pass by @mgoldey in #261
• feat(scf): ROKS Z-vector; exact MBD@rsSCS gradient on ROKS references by @mgoldey in #263
• feat(scf): COSX defaults to the pruned sgx (35,194) grid plus a final pass by @mgoldey in #267
• feat(correlation): separate the method from its local approximation ([local]) by @mgoldey in #297
• fix(gw): apply the Σx − v_xc shift in closed-shell COHSEX on a KS reference by @mgoldey in #300
• test(mp2): validate local RI-MP2 (in-core and integral-direct) against PySCF DF-MP2 by @mgoldey in #301
• test(validation): exact and local dRPA against an independent numpy plasmon dRPA by @mgoldey in #303
• feat(python): expose run_optimize_uhf; fix a libint init race and the gate's stale-import blind spot by @mgoldey in #298
• docs(terf): warn that base_terfc_closed.py is not the shipped kernel by @mgoldey in #299
• docs(validation): grade local RI-MP2 Proven (narrow) by its exact limit by @mgoldey in #306
• fix(scf): check RHF external (RHF→UHF triplet) stability, not just internal by @mgoldey in #304
• fix(dispersion): integrate Hirshfeld volumes where the density is, not on a lattice through its cusps by @mgoldey in #305
• test(cc): validate amplitude-threshold LinLCCD, all three variants, against numpy by @mgoldey in #302
• test(mp2): give AO-Laplace MP2 an external reference that separates quadrature error from implementation error by @mgoldey in #307
• test(ci): run every MPI cross-rank test at np 1, 2 and 3, not one at np 1-2 by @mgoldey in #308
• fix(cdft): the He2+ 3.50 A row needed 106 inner iterations, not a new solver by @mgoldey in #309
• fix(bse): report the ill-conditioned QP energies instead of replacing them by @mgoldey in #310
• fix(cdft): the He2+ lambda=0 start needed outer iterations, not a new solver by @mgoldey in #311
• fix(cdft): an iteration-capped inner solve is not a derivative by @mgoldey in #312
• test(trah): pin the closed-shell null-step guard with an iteration bound by @mgoldey in #316
• test(mp2): validate MP2-V's VV10 half and erfc control against PySCF/numpy (terfc half pending) by @mgoldey in #317
• test(rpax): validate the static RPAx alpha against an independent numpy kernel by @mgoldey in #318
• fix(omega-tuning): seed the cation across omega; the onset is invisible without the Hessian by @mgoldey in #313
• test(qmmm): FD/stationarity harness for optimize_qmmm across boundary schemes and free sets by @mgoldey in #321
• test(terfc): validate terf/terfc integrals and SCS-MP2(2terfc) against a Fourier-space reference by @mgoldey in #320
• fix(pipeline): salvage the pose-handoff guards; tier 2 no longer displaces the docked pose by @mgoldey in #325
• test(terfc): independent 1-D radial oracle for the s-type k-space references (completes #275) by @mgoldey in #326
• fix(scf): range-separated orbital Hessian with the Fock's own SR/LR exchange kernel by @mgoldey in #322
• feat(frequencies): dispersion-corrected frequencies on UKS and ROKS references by @mgoldey in #328
• fix(pipeline): quantum tiers score in the pocket field; refuse total-energy cuts across formulas by @mgoldey in #332
• Build identity: version + git commit at runtime; verified release stamping by @mgoldey in #333
• UKS and ROHF/ROKS analytic gradients with COSX exchange (fit off) by @mgoldey in #327
• Open-shell Z-vector: project the axis-rotation null mode of ²Π radicals (MBD gradients on OH, NO, CH) by @mgoldey in #334
• fix(qmmm): QM-MM Lennard-Jones follows the force field's exclusions across the cut by @mgoldey in #336
• Follow-ups to #333 by @mgoldey in #335
• fix(trah): a declined closed-shell null step falls through to DIIS (#315) by @mgoldey in #338
• fix: test_reproducibility ranks one formula; rebaseline two Z-vector functions by @mgoldey in #339
• fix(units): one Å↔Bohr constant across ferric (#329) by @mgoldey in #337

Full Changelog: v0.1.0rc6...v0.1.0rc7

v0.1.0rc6

What's Changed

• feat(scf): sample ESP on the vdW surface, not just at the nuclei by @mgoldey in #67
• perf(scf): 3.6x on benzene/aug-cc-pVTZ — DF-guess, ShellPair cache, DF-corrected increments by @mgoldey in #68
• Apply rustfmt and add git-blame-ignore-revs for formatting commits by @mgoldey in #71
• perf(integrals): terf kernel 3.1x — tail form, Abel summation, large-S asymptotic by @mgoldey in #69
• Add frozen core orbital selection with "auto" convention support by @mgoldey in #70
• perf(mp2): drive DLPNO-MP2 from the RI B tensor, never from dense (ia|jb) by @mgoldey in #72
• test(export): round-trip every NpzBundle field, not just mo_coeffs by @mgoldey in #73
• test(scf): certify optimize_qmmm stops at a stationary point of its own energy by @mgoldey in #74
• fix(gw): hard-error on an unphysical static polarizability instead of returning it by @mgoldey in #75
• fix(dft): 7 of 12 wB97X-L-V coefficients had flipped signs by @mgoldey in #78
• test(scf): make the cDFT coupling's (lambda, W) pairing non-vacuous by @mgoldey in #80
• fix(scf): Allreduce the rank-local J/K partial, not the caller's accumulate buffer by @mgoldey in #81
• fix(gw): real (energy, oscillator-strength) assignment for the BSE-TDA benchmark by @mgoldey in #76
• test(gw): per-molecule BSE-TDA breakdown and a diagnosis of the 4 unassigned states by @mgoldey in #77
• feat(export): make the C6 partition travel with the numbers; validate the pair sum by @mgoldey in #79
• ci(mpi): OpenMPI CI job + a system-MPI-linked wheel (not vendored) by @mgoldey in #82
• fix(scf): serialize the MINAO per-element guess build under multi-rank MPI by @mgoldey in #85
• fix(scf): open-shell SCF discarded its initial guess and converged to the wrong state by @mgoldey in #83
• feat(memory): one shared pool that reservations debit, across five crates by @mgoldey in #86
• fix(scf): ROHF ignored the configured guess and returned whichever state hcore fell into by @mgoldey in #87
• docs: QM/MM, golden paths, and xtb geometry initialization by @mgoldey in #89
• fix(ci): install uv in the coverage job, which has been red since 2026-09-16 by @mgoldey in #90
• feat(tools): read PDB/PQR/SDF/mol2/SMILES by converting to the one Rust entry point by @mgoldey in #91
• ci: run check/clippy/rustdoc as their own job, not behind a 63-min test step by @mgoldey in #94
• fix(pipeline): carry the docked pose into the quantum tiers by @mgoldey in #93
• fix(test): drop energy_conv=1e-10 from the MPI RHF reference; it never converged by @mgoldey in #95
• perf(gradient): batch the KS-DFT XC gradient grid; 7.4x lower peak by @mgoldey in #92
• feat(scf): TRAH, AURORA-SCF, and redundant internal coordinates (all opt-in) by @mgoldey in #88
• feat(python): expose FrequencyResult.normal_modes by @mgoldey in #97
• ci(mpi): retry the mpirun launch once; it flakes ~7% of runs by @mgoldey in #98
• feat(pipeline): propose active-site substitutions, scored relative to the parent by @mgoldey in #96
• fix(gate): the complexity baseline was keyed on absolute paths by @mgoldey in #104
• docs(tox): the catalog-construction claim was 5x, not "dominates" by @mgoldey in #108
• feat(scf): P-RFO transition-state search — ferric can now FIND a saddle by @mgoldey in #106
• docs: commit the measured pipeline notes, which only existed locally by @mgoldey in #103
• feat(pipeline): TierResult carries the resolution it can actually deliver by @mgoldey in #109
• feat(danuglipron): the substitution golden path, end to end on 7LCJ by @mgoldey in #102
• ci: shard the test tier 4 ways behind one nextest archive build by @mgoldey in #101
• fix(scf): the saddle cost model undercounted by 3 gradient evaluations by @mgoldey in #110
• exp(danuglipron): M12+M13 — all four pose protocols closed, including 'select a pose' by @mgoldey in #107
• docs(golden-path): C3 merged and demonstrated under QM/MM embedding; TS cost +3 by @mgoldey in #111
• feat(viz): figures for pipeline outputs — energy plots and molecule depictions by @mgoldey in #105
• feat(d3): native D3(BJ) dispersion, Z=1-103, zero new dependencies by @mgoldey in #99
• feat(viz): reaction_path — the two minima an IRC connects, with both barriers by @mgoldey in #116
• feat(scf): IRC, plus four run_saddle defects CodeRabbit caught on #106/#110 by @mgoldey in #114
• fix: two silent-wrong-answer findings left open on merged PRs (#107, #109) by @mgoldey in #115
• fix(tiers): a comment declared code missing that had existed for three weeks by @mgoldey in #117
• docs: tier costs were wrong in four more places, and one claimed MEASURED by @mgoldey in #113
• feat(viz): accept a per-(substituent,site) noise floor, and render an unmeasured cell as unbounded not zero by @mgoldey in #119
• feat(morph): pair poses by scaffold -- a paired ddE estimator the five closed routes never tried by @mgoldey in #118
• docs(golden-path): correct the TS crossover to N=(n_steps-1)/12 -- QM region size, not step count, sizes the job by @mgoldey in #120
• feat(python): thread point_charges into run_saddle/run_irc/run_frequencies so C3-C5 run in an MM field [MERGE BEFORE #122] by @mgoldey in #121
• feat(runlog): emit each method's own total in a result record, so the log carries the answer not the SCF reference by @mgoldey in #100
• docs(golden-path): scope the embedded-catalyst claim to Rust, and record the Python chain #121 enables by @mgoldey in #122
• docs(golden-path): the TS cost model counts gradients as if they were energies (~1.7x low) by @mgoldey in #125
• fix(viz): two-line grid captions printed on top of the molecules by @mgoldey in #124
• docs(golden-path): a retraction and the claim it retracts, live in one document by @mgoldey in #126
• docs+viz: one answer table for the six pharma use cases, plus optimization_trace to catch a climbing optimization by @mgoldey in #123
• feat(cli,solvate): a [qmmm] TOML section, and an explicit TIP3P droplet to fill it by @mgoldey in #127
• test(mm): cover ring exclusions, which both OpenMM cases miss by @mgoldey in #129
• fix(structure): two-letter elements read as carbon because the membership test could never match; add a .gro reader by @mgoldey in #128
• fix(solvate,ci): a short radii list dropped solute atoms and shifted every qm_indices entry by @mgoldey in #130
• docs(golden-path): the catalyst branch runs; fix a self-contradiction, a C4 API error, and two plot layout defects by @mgoldey in #131
• fix(golden-path): the answer table's FF cost was 3.5x high and carried no provenance by @mgoldey in #132
• fix(golden-path): three more answer-table costs, including a hidden network call in the default liability path by @mgoldey in #133
• docs: current state, not changelog — QM/MM page plus a site-wide sweep by @mgoldey in #134
• fix(tiers): refuse a cached geometry that is not the candidate molecule by @mgoldey in #135
• fix(gate): clear a squatting sccache before every cargo stage, not just at startup by @mgoldey in #137
• fix(tools): audit the geometry writers — two wrote a header that lied about the body by @mgoldey in #136
• fix(docking): a docked pose was coming back as the wrong enantiomer by @mgoldey in #138
• feat(tox): a command-line entry point for the toxicity screen by @mgoldey in #140
• feat(campaign): stream long...

libint2 2.7.2 export for the PyPI wheel (ferric-small-1)

libint2 v2.7.2 source export generated by scripts/generate-libint-small.sh: 4-centre ERI derivatives to order 2 (l <= 6,6,3), 3- and 2-centre to order 1 (pure SH), one-body to order 2 (l <= 6,4,3), generic code, no unrolling. Built by .github/workflows/wheels.yml, which pins its sha256. Compiled: libint2.so 191 MB raw / 53.8 MB zipped. Not a ferric release.

v0.1.0rc5

What's Changed

• COSX, LinK added for HF/KS
• Integral-direct assembly added for local methods with amplitude thresholds
• Memory management and disk quota management

Full Changelog: v0.1.0rc4...v0.1.0rc5

v0.1.0rc4

feat(wheels): ship `ferric` as a console script in the Python wheel (…

v0.1.0rc3

fix(wheels): Cargo.toml needs SemVer, not the raw PEP 440 tag string …

v0.1.0rc2

fix(wheels): prepare-container.sh uses curl, not wget (not installed)…