Step 5, run early and read-only: what re-sourcing would actually change

diffreport.py fetches each resolved basis from BSE, converts it, and
compares value by value against the shipped file.  Nothing is replaced and
the live library is untouched; this is the report step 6 reviews.

Of the 130 single-file orbital sets that step 1 resolved:

    identical        78   every value agrees to the stored precision
    structural       35   shell layout differs
    coverage         10   same values, different element sets
    constant-ratio    4   all differing values share one ratio
    DIFFERS           3   values genuinely moved

So the great majority is exact, which is the reassuring part.

Two refinements the first pass needed.  Shell ORDER within an atom is not
chemistry -- a basis with its D shell before its SP shell is the same
basis -- so a mere reordering is now reported as such instead of as a
structural difference.  And fetching was filtering to ECCE's element list,
which made BSE answer 404 whenever ECCE covered an element it does not;
it now fetches everything and intersects afterwards.

The constant-ratio bucket is DESCRIPTIVE, not a verdict of harmless, and
that distinction was earned.  4-31G's Helium differs by exactly 1.688956,
the same factor as 6-31G's Helium, which was already shown to change no
energy (8e-9 Hartree in ORCA) because the code renormalises.  But the same
sweep found a constant ratio of exactly -1.000000 -- a sign flip -- and one
of exactly 0.100000, a factor of ten.  Neither is a normalisation constant,
and an earlier version of this tool labelled all three "normalisation",
which would have waved through the two most worth looking at.

Also worth knowing for step 6: the twelve structural results among the
CURATED mappings are evidence about those identity guesses, not about the
data.  "ADZP is jorge-adzp" was a proposal, and H having 5 shells against
6 says it needs checking rather than accepting.  Several Partridge
differences look like representation rather than data -- ECCE stores 2
shells where BSE lists 29 -- which is what an uncontracted set looks like
written two different ways.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EFMUNbw2UpRPmgsqviAowb

Author: andy