The map-coefficient to real-space transform was still gemmi's bundled pocketfft while the
map to structure-factor direction had moved to FFTW. MapFromFPhi does that inverse with
FFTW - the same half-l XYZ layout, the same 1/V scale, NaN coefficients read as zero, the
same conjugation convention - and the output maps are now built with it. Plans are
FFTW_ESTIMATE, cached per grid size and direction, and made under the shared planner lock.
No pocketfft code is instantiated in rugnux any more (75 symbols -> 0); gemmi's fourier.hpp
is still included for its non-FFT helpers (get_size_for_hkl, get_f_phi_on_grid), so the
vendored header and its notice stay.
Tested element-wise against gemmi on odd and even grids, with arbitrary phases, negative
indices and symmetry/Friedel expansion in three space groups, plus a round trip; both new
tests fail if the conjugation is dropped. On the eight --model audit sets the merged MTZ is
byte-identical, the map coefficients are identical, the CCP4 maps agree to 3e-7 relative
(map CC 1.00000000) and every reported key is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013nW6FNRP1bBJJ8pfHiByAT
leonarski_fandClaude Opus 5
1b414909cb