Tracker: migrate avoidable numerics to fortnum (drop GSL, audit vendored math)

[krystophny]

Scope

KAMEL is MIT-licensed but fetches and links GSL 2.4 (GPL-2.0+) through cmake/FetchGSL.cmake, and vendors several copies of special-function, quadrature, and ODE code with mixed provenance. fortnum replaces the generic numerics: special functions behind the direct GSL bindings, adaptive quadrature, generic ODE/RK steppers, and duplicate Lagrange-interpolation helpers. It also gives a clean-room home for the vendored Bessel, error-function, and hypergeometric code that currently has uncontrolled licensing. Domain physics stays in KAMEL: the KIM kinetic integrands, the KiLCA mode and conductivity solvers, and the equilibrium and Fourier preprocessing.

Planned leaves

leaf	replaces	needs	model
special functions to fortnum_special	KIM/src/math/gsl_mod.f90	fortnum M1 (special)	sonnet
drop GSL fetch glue	cmake/FetchGSL.cmake	after special/quadrature leaves merge	sonnet
quadrature and ODE replacement	KIM/src/math/quadpack_integration_m.f90, KIM/src/math/quadpack_compat_xerror.f90, KIM/src/kernels/integrals.f90, KIM/src/kernels/integrals_adaptive.f90, KIM/src/kernels/integrands_adaptive.f90, KIM/src/math/ddeabm/*, KiLCA/math/odeint_allroutines.f, KiLCA/solver/VER_5_STABLE/odeint_allroutines.f, KiLCA/flre/conductivity/odeint_allroutines.f, common/equil/rk4_integrator.f90, PreProc/fourier/src/rk4dc.f90	fortnum M2 (ode), M4 (integrate)	opus
dedupe Lagrange interpolation to fortnum_interp	KIM/src/util/plag_coeff.f90, QL-Balance/src/base/plag_coeff.f90, KiLCA/solver/VER_5_STABLE/plag_coeff.f90	fortnum M3 (interp)	sonnet

Provenance-audit pre-leaves

Each vendored special-function tree needs a provenance and license audit before any port. The audit is one pre-leaf per tree (opus); the replace-vs-vendor decision per #291 (clean-room fortnum_special, or deliberate vendoring with explicit license and tests) is a separate decision leaf (fable).

pre-leaf (provenance-audit)	paths	model
AMOS Bessel	KiLCA/math/bessel/amos/*	opus
TOMS 644 Bessel	KiLCA/math/bessel/644/*	opus
SLATEC support	common/math/slatec/*	opus
libcerf	KIM/src/math/libcerf-main/*, KIM/src/math/use_libcerf_mod.f90	opus
confluent hypergeometric	KiLCA/math/hyper/hyper1F1.cpp	opus

Leaf issues are created from this table only once the named fortnum milestone is closed. Until then this tracker is the only issue.

Execution conventions

Model ladder (implementation):

• haiku: label/milestone housekeeping, checklist updates, re-export boilerplate. No algorithms.
• sonnet: fully specified ports (source path + SHA given), wrappers, CI fixes with a clear failing log.
• opus: build-system and CI scaffolding, multi-file refactors, ports that change APIs.
• fable: API design issues, clean-room algorithm work, review of opus-implemented PRs.

Review ladder: sonnet PRs reviewed by opus; opus PRs reviewed by fable; haiku changes gated by CI only.

PR rules:

• One leaf issue = one PR. Branch from the base named in the issue.
• PR title = issue title. Body: Closes #N plus a ## Verification section with real failing-before and passing-after output.
• Squash merge. Stacked PRs only where the issue names a non-main base.

Executor rule: if any issue under "Blocked by" is open, stop and report instead of starting.

Umbrella

itpplasma/libneo#291

Original report

Problem

KAMEL is MIT-licensed. It links GSL (GPL-2.0+) at build time via cmake/FetchGSL.cmake, which downloads and compiles GSL 2.4. Any binary that links KAMEL is subject to GPL obligations, which contradicts the MIT grant.

Affected files

KiLCA (C++)

File	GSL usage
KiLCA/math/adapt_grid/adaptive_grid.cpp	gsl_heapsort_index, gsl_integration_workspace, gsl_integration_qag
KiLCA/math/fourier/four_transf.cpp	gsl_integration_qag
KiLCA/eigmode/calc_eigmode.cpp, calc_cond.cpp, sysmat_profs.cpp, others	GSL errno, integration
KiLCA/imhd/incompressible.cpp, compressible_flow.cpp, imhd_zone.cpp	integration, special functions
KiLCA/mode/calc_mode.cpp, QL-Balance/src/base/vel_integral.cpp	integration

KIM (Fortran)

KIM/src/math/gsl_mod.f90 defines direct C-bind interfaces to:

• gsl_sf_bessel_In (modified Bessel I of integer order)
• gsl_sf_erf, gsl_sf_erfc
• gsl_sf_dawson

These are used in integrands.f90, integrands_adaptive.f90, FLR2_asymptotics.f90, numerics_utils.f90.

Replacement directions

• heapsort: replace gsl_heapsort_index with std::sort + an index array using <algorithm> (no dependency).
• QAG adaptive integration: QUADPACK from Netlib (public domain) via a thin C or Fortran wrapper, or a self-contained Gauss-Kronrod 15/31-point rule.
• GL fixed quadrature: compute Gauss-Legendre nodes and weights in-tree with the Golub-Welsch algorithm (<50 lines).
• Bessel In: the C99 standard library does not cover this; Boost.Math (BSL-1.0) or the AMOS library (Netlib, public domain) are permissively licensed replacements.
• erf/erfc: these are C99 standard functions in <math.h>; drop the GSL wrappers and call erf/erfc directly.
• Dawson: implement the Rybicki (1989) algorithm in-tree (~30 lines) or use Cephes (BSD-compatible).

Build system

Remove cmake/FetchGSL.cmake and its invocation from cmake/Dependencies.cmake once all call sites are replaced. Remove KIM/src/math/gsl_mod.f90 when no callers remain.

[krystophny]

Derivative-ready update: when this tracker spawns leaves, each fortnum replacement should document whether it exposes primal-only or derivative-ready behavior. See lazy-fortran/fortnum#36-#42.

[krystophny]

Automated triage pass (Claude), grounded in the repository at 2722090f. Verify before acting.

The headline goal of this tracker — GSL is gone — landed in 41188c89 ("Drop GSL and finalize fortnum wiring (8/8)", #147), which deleted cmake/FetchGSL.cmake (42 lines removed) and its invocation in cmake/Dependencies.cmake. KIM/src/math/gsl_mod.f90 no longer exists. rg -i gsl over the tree at HEAD returns only comments and vestigial identifiers (KiLCA/imhd/imhd_zone.h:54 calculate_basis_incompressible_gsl, KiLCA/background/calc_back.cpp:26 #define GSL_SUCCESS 0) — no headers, no link targets, no fetch.

So this is not obsolete but it is badly out of date as written. Roughly three of the four planned leaves and three of the five provenance pre-leaves are done; the table still lists deleted files as targets. Keeping it open with the original body will send an executor after paths that no longer exist. Below is the re-scoped tracker.

Purpose

KAMEL is MIT-licensed and now ships without any GPL link dependency. The remaining value of this tracker is the second half of its title — audit vendored math — plus the duplicate-elimination leaf that was never executed. Two live, compiled numerical helpers (common/equil/rk4_integrator.f90, PreProc/fourier/src/rk4dc.f90) are verbatim-looking Numerical Recipes RK4 transcriptions, which is a stricter licensing problem than GSL was (NR code is proprietary, not merely copyleft, and cannot be relicensed under MIT at all). Three divergent copies of plag_coeff remain in-tree while LIBNEO::polylag is already linked, so the same 4-point Lagrange stencil is compiled three times with three different interfaces. Finishing this closes the license story for a public v1.0.0 release and removes the last hand-rolled, non-thread-safe steppers from the hot equilibrium path.

Current state

Done — remove from the table, do not re-open as leaves:

Original leaf / pre-leaf	Status at 2722090f
special functions → fortnum_special (KIM/src/math/gsl_mod.f90)	done, a2d641d5 (#140); file deleted
drop GSL fetch glue (cmake/FetchGSL.cmake)	done, 41188c89 (#147); file deleted
KIM Fokker–Planck quadrature	done, 95f02a62 (#141)
KiLCA/QL-Balance quadrature	done, 0e660b71 (#142)
KiLCA background/imhd ODEs	done, 8682b399 (#143), continuous rk8pd
KIM/src/math/ddeabm/*	done, eb2ed22d (#149); directory deleted
netlib QUADPACK + KIM/src/math/quadpack_compat_xerror.f90	done, 2341697e (#150); file deleted
AMOS Bessel (KiLCA/math/bessel/amos/*)	tree deleted; complex Bessel migrated 0f8cbbe1 (#145), shim at common/math/fortnum_amos_compat.f90
TOMS 644 Bessel (KiLCA/math/bessel/644/*)	tree deleted
SLATEC (common/math/slatec/*)	tree deleted, 2341697e
confluent 1F1	migrated eb511831 (#146); KiLCA/math/hyper/hyper1F1.cpp:11 now #include "fortnum.h", :50 calls fortnum_hyperg_1f1_a1

KIM/src/math/quadpack_integration_m.f90 survives under its old name but is now a fortnum wrapper — its own header says so (KIM/src/math/quadpack_integration_m.f90:1-6):

module quadpack_integration_m
    ! Adaptive Gauss-Kronrod integration wrapper for KIM Fokker-Planck kernels.
    ! Backed by the fortnum numerical core (fortnum_integrate); the
    ! rkf45_integrand_context_t flows through the integrand callback's class(*)
    ! ctx, so no module-scope thread context is needed.

Not done — the actual remaining scope:

(a) Numerical-Recipes RK4, live and compiled. common/equil/rk4_integrator.f90 is registered at common/equil/CMakeLists.txt:46 and PreProc/fourier/src/rk4dc.f90 at PreProc/fourier/CMakeLists.txt:47. Both use NR's exact local-variable vocabulary (HH, H6, XH, YT, DYT, DYM) and NR's argument order (Y, N, X, H, DERIVS):

      SUBROUTINE RK4C(Y,N,X,H,DERIVS)
      use rk4c_mod
      DOUBLE PRECISION X,H,HH,H6,XH
      EXTERNAL DERIVS
      double complex, dimension(n) :: Y

(PreProc/fourier/src/rk4dc.f90:14-19)

Both also carry module-level saved scratch arrays reallocated on dimension change (common/equil/rk4_integrator.f90:16-18, PreProc/fourier/src/rk4dc.f90:8-10), which is not thread-safe.

(b) plag_coeff triplication. Three divergent copies:

• KIM/src/util/plag_coeff.f90 (102 lines) — bare subroutine binsrc + plag_coeff, no module, registered at KIM/src/CMakeSources.in:36. ~9 call sites, e.g. KIM/src/electromagnetic/ampere_matrices.f90:27,35, KIM/src/background_equilibrium/calculate_equil.f90:171,179,237,245, KIM/src/grid/grid_mod.f90:230,259.
• QL-Balance/src/base/plag_coeff.f90 (115 lines) — module PolyLagrangeInterpolation with module state nlagr = 4, nder = 0, allocatable :: coef; registered at QL-Balance/CMakeLists.txt:69. Used by SingleStep.f90:92, get_dql.f90:27,517,541, time_evolution.f90:615, gengrid.f90:11, paramscan.f90:283, diff_coeffs.f90:10.
• KiLCA/solver/VER_5_STABLE/plag_coeff.f90 (59 lines) — dead: no solver/VER_5_STABLE/* source appears in the kilca_lib source list (KiLCA/CMakeLists.txt:184-240).

Meanwhile QL-Balance/CMakeLists.txt:147 already links LIBNEO::polylag, i.e. libneo/src/util/plag_coeff.f90, module plag_coeff_sub — a fourth copy of the same routine is already in the link closure.

(c) Dead vendored odeint_allroutines.f. Three copies exist — KiLCA/math/odeint_allroutines.f, KiLCA/solver/VER_5_STABLE/odeint_allroutines.f, KiLCA/flre/conductivity/odeint_allroutines.f — and none is compiled (rg 'odeint' */CMakeLists.txt matches only QL-Balance/CMakeLists.txt:100 cvodeint.cpp). The live callers QL-Balance/src/base/Wmn.f90:224,264 resolve the symbol from libneo, which is the subject of itpplasma/libneo#298.

(d) libcerf audit — resolvable now, no port needed. KIM/src/math/libcerf-main/LICENSE is the MIT text ("Copyright (c) 2012 Massachusetts Institute of Technology / Copyright (c) 2013 Forschungszentrum Jülich GmbH"), and KIM/src/math/use_libcerf_mod.f90:1-19 carries an MIT grant from Antonio Cervellino. Sole live consumer: KIM/src/math/plasma_disp_func.f90. This pre-leaf's answer is "MIT, compatible, keep vendored" — record it and drop the leaf.

(e) zersol include path. KiLCA/CMakeLists.txt:142 and :281 still add math/zersol-0.0.0/src to include directories although ZEAL was dropped in e2a2ab0f/37324177. Cosmetic, but it belongs to the same sweep.

Reproduction / evidence

Verification recipe (no bug to reproduce; this is a cleanup/licensing tracker). All commands from the repo root, using the real build described in README.md/Makefile:

# 1. GSL is truly gone (baseline: must print nothing)
grep -rn 'gsl/gsl_' --include='*.c' --include='*.cpp' --include='*.h' --include='*.f90' .
grep -rn 'FetchGSL\|GSL::\|-lgsl' CMakeLists.txt cmake/ */CMakeLists.txt
ldd build/install/bin/KIM.x | grep -i gsl        # expect no match

# 2. plag_coeff copies still in the build (currently 2 live + 1 libneo)
grep -rn 'plag_coeff' --include=CMakeLists.txt --include='*.in' . 

# 3. NR-derived RK4 still compiled
grep -n 'rk4_integrator.f90' common/equil/CMakeLists.txt   # -> line 46
grep -n 'rk4dc.f90' PreProc/fourier/CMakeLists.txt         # -> line 47

# 4. baseline before touching anything
make all && make test
make golden          # A/B against tag golden-baseline (eb511831)

Numerical guard for the RK4 replacement: common/equil/test/test_rk4_integrator.f90 is already registered (common/equil/CMakeLists.txt:88-96) — extend it rather than writing a new harness. For plag_coeff, KIM/tests/test_plag_coeff.f90 is registered at KIM/tests/CMakeLists.txt:93-99; it must keep passing bit-for-bit through the dedupe.

Behaviour (BDD)

Scenario: no Numerical-Recipes-derived stepper remains in the build
  Given a clean checkout at HEAD after this tracker's leaves land
  When I run "grep -rn 'DYT\|DYM\|H6' common/equil/rk4_integrator.f90 PreProc/fourier/src/rk4dc.f90"
  Then both files are absent from the tree
  And "grep -n rk4" common/equil/CMakeLists.txt and PreProc/fourier/CMakeLists.txt return no source entry
  And "make all && make test" succeeds with all CTest targets passing

Scenario: equilibrium field-line tracing is unchanged by the RK4 replacement
  Given "make golden" run against tag golden-baseline (eb511831)
  When the field-line tracer in common/equil uses the fortnum stepper instead of rk4_step
  Then test/golden/kim/cases/electromagnetic reports agreement within the tolerances
       configured in test/golden/kim/config.sh
  And test/golden/bin/gr_numcompare.py reports no field exceeding its threshold

Scenario: exactly one Lagrange-coefficient implementation is compiled
  Given QL-Balance already links LIBNEO::polylag (QL-Balance/CMakeLists.txt:147)
  When KIM/src/util/plag_coeff.f90 and QL-Balance/src/base/plag_coeff.f90 are removed
       and all callers switched to the single chosen provider
  Then "grep -rln 'SUBROUTINE plag_coeff' --include='*.f90' ." matches no file in this repo
  And ctest -R test_plag_coeff passes
  And "make golden" shows no change in qlbalance or kim cases

Scenario: libcerf provenance is recorded, not ported
  Given KIM/src/math/libcerf-main/LICENSE is the MIT licence text
  When the provenance audit is written to docs/
  Then docs/ contains a THIRD_PARTY entry naming libcerf, its upstream URL,
       its MIT licence, and its sole consumer KIM/src/math/plasma_disp_func.f90
  And no code change is made to KIM/src/math/libcerf-main/

Scenario: dead vendored ODE and interpolation copies are gone
  Given none of the three KiLCA odeint_allroutines.f files appears in any CMakeLists.txt
  When they and KiLCA/solver/VER_5_STABLE/plag_coeff.f90 are deleted
  Then "make all" succeeds with no unresolved symbol
  And QL-Balance/src/base/Wmn.f90 still links odeint_allroutines from libneo

Implementation plan

Order matters: step 1 first, because it is the only step with a licensing deadline, and step 5 last, because deletions are cheapest to review once the live replacements have settled.

1. Leaf R1 — provenance audit of the RK4 pair (opus). Create docs/THIRD_PARTY.md. Record, per vendored item: path, upstream, licence, consumer. Items to cover: KIM/src/math/libcerf-main/ (MIT), common/math/fortnum_amos_compat.f90, KiLCA/math/W/wofz.f (TOMS 680 — check the ACM licence), KiLCA/math/zgeev_tuned.f, KiLCA/flre/dispersion/dpsort.f (SLATEC-derived, verify it survived 2341697e cleanly), and both RK4 files. For common/equil/rk4_integrator.f90 and PreProc/fourier/src/rk4dc.f90 state explicitly whether the code is NR-derived; if it is, they cannot stay under MIT and step 2 becomes mandatory rather than optional.
2. Leaf R2 — replace the real RK4 (opus). Edit common/equil/rk4_integrator.f90: either delete it and switch callers to a fortnum fixed/adaptive stepper, or, if a fixed-step RK4 with no external dependency is genuinely wanted, rewrite rk4_step clean-room from the classical Butcher tableau with local (non-save) work arrays and a fresh header. Update common/equil/CMakeLists.txt:46 accordingly. Extend common/equil/test/test_rk4_integrator.f90 with a fourth-order convergence check: integrate y' = -y, y(0)=1 to x=1 at h = 0.1, 0.05, 0.025 and assert the observed order is 4.0 ± 0.15.
3. Leaf R3 — replace the complex RK4 (opus). Same treatment for PreProc/fourier/src/rk4c_mod/RK4C in PreProc/fourier/src/rk4dc.f90, registered at PreProc/fourier/CMakeLists.txt:47. A complex-valued fixed-step RK4 can be expressed as two real systems if fortnum has no complex stepper; note which choice was made in the PR body. Add PreProc/fourier/test/test_rk4c.f90 and register it in PreProc/fourier/CMakeLists.txt with add_test(NAME test_rk4c ...); assert y' = i·y, y(0)=1 reproduces exp(i) to 1e-10 after 1000 steps.
4. Leaf R4 — dedupe plag_coeff (sonnet). Pick one provider. Recommended: libneo's plag_coeff_sub (libneo/src/util/plag_coeff.f90), since LIBNEO::polylag is already in the QL-Balance link closure. Then:
   • delete QL-Balance/src/base/plag_coeff.f90 and its entry at QL-Balance/CMakeLists.txt:69. The module also exports nlagr, nder, coef and get_ind_Lagr_interp, which libneo does not — move those four into a small new QL-Balance/src/base/lagrange_state_m.f90 and update the seven use PolyLagrangeInterpolation sites (SingleStep.f90:92, get_dql.f90:27,517,541, time_evolution.f90:615, gengrid.f90:11, paramscan.f90:283, diff_coeffs.f90:10);
   • delete KIM/src/util/plag_coeff.f90 and its entry at KIM/src/CMakeSources.in:36; add use plag_coeff_sub, only: plag_coeff at the ~9 KIM call sites listed above. binsrc lives in the same KIM file — decide whether it also comes from libneo or moves to KIM/src/util/;
   • link LIBNEO::polylag into the KIM targets and into KIM/tests/CMakeLists.txt:97;
   • keep KIM/tests/test_plag_coeff.f90 passing unchanged. If libneo's variant differs numerically from KIM's, that is a finding, not a licence to change tolerances — report it and coordinate with itpplasma/libneo before merging.
     Ordering: this leaf depends on no libneo change if libneo's routine is already sufficient; if it is not, the libneo PR must merge first and LIBNEO_REF be bumped.
5. Leaf R5 — delete dead vendored copies (sonnet). Remove KiLCA/math/odeint_allroutines.f, KiLCA/solver/VER_5_STABLE/odeint_allroutines.f, KiLCA/flre/conductivity/odeint_allroutines.f, KiLCA/flre/conductivity/tests/odeint_allroutines.f, KiLCA/solver/VER_5_STABLE/plag_coeff.f90; drop the stale math/zersol-0.0.0/src include entries at KiLCA/CMakeLists.txt:142 and :281. Verify with a full make clean && make all that no symbol goes unresolved. Do not touch KiLCA/solver/VER_5_STABLE/ beyond those two files in this PR — issue refactor: KiLCA conductivity model selection by CMake filename-substitution breeds dead parallel variants #167 owns the wider dead-variant question.
6. Leaf R6 — derivative-readiness note (haiku/sonnet). Per the 2026-06-10 comment on this issue: add one line per migrated call site in docs/THIRD_PARTY.md (or a sibling docs/fortnum-migration.md) recording primal-only vs derivative-ready, covering the already-merged Migrate KIM special functions to fortnum (1/8) #140–Drop the netlib QUADPACK and SLATEC dependencies (Z3) #150 as well as R2–R4.
7. Housekeeping. Rewrite this issue body to the table above so the checklist matches the tree, and drop the AMOS / TOMS 644 / SLATEC / libcerf pre-leaf rows.

Acceptance criteria

• grep -rn 'gsl/gsl_\|FetchGSL\|-lgsl' . (excluding docs/ and comments) returns nothing — already true at 2722090f; guard against regression.
• docs/THIRD_PARTY.md exists and names, for every vendored numerical source still compiled, its upstream, licence, and consumer.
• No file in the repo matches grep -rn 'DYT.*DYM' --include='*.f90' --include='*.f' .
• grep -rc 'SUBROUTINE plag_coeff' --include='*.f90' . counts exactly zero occurrences inside this repository.
• make clean && make all && make test succeeds; every CTest target passes, including test_plag_coeff, test_rk4_integrator, test_integration_methods, test_equil_ode_ddeabm, test_kim_solver, test_kim_solver_em.
• New tests test_rk4c (PreProc) and the extended test_rk4_integrator order check pass; observed RK4 convergence order in 4.0 ± 0.15.
• make golden (A/B against tag golden-baseline = eb511831) shows no field regression for kim/cases/electromagnetic, the kilca cases, and the qlbalance cases at the thresholds in each config.sh.
• make KiLCA succeeds after the zersol include entries are removed.

References

• Press, Teukolsky, Vetterling & Flannery, Numerical Recipes, 2nd/3rd ed., CUP — the source whose RK4/odeint/rkqs/rkck naming the files in (a) and (c) reproduce. Its licence forbids redistribution of transcribed code, which is why step 1 gates step 2.
• Piessens, de Doncker-Kapenga, Überhuber & Kahaner, QUADPACK: A Subroutine Package for Automatic Integration, Springer 1983 — the algorithm behind the QAG/QAGS names retained in KIM/src/math/quadpack_integration_m.f90.
• Prince & Dormand, "High order embedded Runge-Kutta formulae", J. Comput. Appl. Math. 7 (1981) 67–75 — the RK8(9)/rk8pd family already adopted in 8682b399 for the KiLCA background and imhd ODEs.
• Johnson, S. G., libcerf / Faddeeva package (MIT) — KIM/src/math/libcerf-main/, used by KIM/src/math/plasma_disp_func.f90 for the plasma dispersion function.
• Comparable practice in the org and beyond: SIMPLE and NEO-RT are running the same fortnum consolidation (Tracker: adopt derivative-ready fortnum for RNG and generic ODE helpers SIMPLE#371, Tracker: replace vendored numeric helper copies with fortnum NEO-RT#43); outside the group, STELLOPT and BEAMS3D likewise route generic ODE/quadrature through a single shared numerics layer rather than per-module vendored copies.

Scope and risk

Six leaves, not one PR. R1 is documentation-only and can land immediately. R2/R3 are the risky ones: they change a stepper on the equilibrium and Fourier-preprocessing paths, so make golden against golden-baseline is mandatory — a changed step-size controller can shift near-resonant roots (see #164, which documents exactly that ULP sensitivity in ql-balance). R4 is mechanical but touches ~16 call sites across two subprojects and can surface a genuine numerical difference between the KIM and libneo variants. R5 is pure deletion and should be verified only by a clean rebuild.

Cross-links: umbrella itpplasma/libneo#291; itpplasma/libneo#298 (wrap odeint_allroutines on fortnum) owns the live symbol that KAMEL's QL-Balance/src/base/Wmn.f90:224 consumes; itpplasma/libneo#295 (verify no shipped target links GSL/FFTW) is the org-wide counterpart of the first acceptance box; #167 owns the dead parallel-variant sources in KiLCA/flre/conductivity/ that R5 deliberately leaves alone.