SOF-8005: parse NWChem initial and final geometry - #165

The cell was derived from the first geometry block alone, so a relaxation
that expands a molecule left atoms outside the box. Reproduced: a
diatomic going 0.2 -> 2.5 A landed at z = 4.0 and -1.0 in a 3.0 A box,
which reads as extra fragments and turns inchi 1S/H2O/h1H2 into
1S/H2O.2H2/h1H2;2*1H -- silently wrong chemistry, not a visible failure.

Still one cell for both, so the two structures stay comparable; it is now
sized to whichever geometry needs more room. Contracting relaxations,
the common case, are unaffected: the fixture still gives 5.72171236 for
both, and the H2O acceptance job still gives 3.163849.

Also drops four restatements of why the cell is shared, keeping the
explanation where the logic is. Found by review: the problem was
repetition, not volume -- express sits below this repo's docstring
density.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

tb-review blocker on #165. `(?:row\n)+` stops at the first non-row line,
so a log truncated mid-table parses as a shorter molecule and nothing
compared the two blocks. Reproduced: initial ['O','H','H'], final
['O','H'], serialized as H2O2. rupy publishes it, because
extract_structures() gates only on `if initial and final:` - and this is
a new failure mode, since on main these methods returned None.

_geometry_blocks() now parses every block once and returns nothing when
they disagree on which atoms are present, so both callers fail closed.
It also hosts the cell derivation, which previously reached three levels
into _geometry_block's return.

Also from the review:
- drop the `angstroms` verbatim branch. 1.889725989 * Constant.BOHR is
  1.0000000162985654, so it bought 5e-8 A of a 3 A cell for a units
  special case in every block.
- final_lattice_vectors aliases initial_lattice_vectors rather than
  repeating the body; a fresh dict per call, so no aliasing hazard.
- declare is_non_periodic = False on BaseParser instead of leaving it an
  implicit getattr protocol. Material keeps the getattr, being built
  with parser=None in tests.
- the shared cell is for comparability, not containment: a cell derived
  per structure always contains its own atoms. Centering a basis in the
  other block's cell is what put atoms outside. Docstrings said both.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The two flagged on the diff, plus the same kind elsewhere: the header
above the regex primitives, the gloss on ELEMENT_REGEX, and the first
line above is_non_periodic on NwchemParser. 14 comment lines down to 9.

What survives explains regex mechanics that are not readable off the
pattern - why the tag must start with a symbol, what the tempered gap
prevents, what bounds the table - and the one coupling between _basis
and _lattice_vectors.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

geometry_blocks() converts units and nothing else; basis(text, index)
picks a block by id, 0 initial and -1 final, which is the "extract all
structure entries and then choose by id" comment. The cell derivation and
the centering are gone, and with them the mat3ra-made import.

Both lattice delegates are deleted rather than reworded. They returned the
same invented cell, and IonicDataMixin declares them `pass` without
ABCMeta, so unimplemented is None - which is what VASP already relies on by
omitting initial_*.

references.py BASIS goes back to plain converted coordinates. It held
centered ones, which is what "Seems suspicious" was pointing at: a fixture
recording that the parser had moved the atoms.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The relocation is only worth anything if it changed no output, so assert
the numbers the live job produced: edge 3.163848897 and O at
[1.337365565, 1.554783897, 1.291226523], from web-app's
final-structure-nwchem-optimized.json fixture.

test-003 is that geometry translated to the origin, so the padding and the
centering are what have to put it back - fed its own published coordinates
the pipeline would be a fixed point and a missing centering step would pass.

Verified against a488316 rather than asserted: the full serialized material
for test-001, test-002 and test-003, initial and final, is byte-identical
before and after the move.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Reverses the fixture half of 9313866. test-003 was the relaxed cluster
geometry re-emitted in NWChem's column format - the only "NWChem output" in
the repo that no NWChem ever printed, and opaque fixture data is what
triggered this review in the first place.

The relocation invariant does not need it: the full serialized material for
test-001 and test-002, initial and final, is byte-identical to a488316, and
the cluster's own numbers stay pinned where they belong, in web-app's
final-structure-nwchem-optimized.json and the feature's HOMO/LUMO
assertions.

So the assertions move onto test-002, which is a real geometry
optimization. Mutation strength is unchanged or better: reverting the
centering fails one test, reverting the per-block sizing now fails two
(test-002's final block alone gives 3.05 A, not 5.72), reverting the
material.py hook fails three.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

TB-TEST-1. The restructure deleted test_shared_cell_fits_a_relaxation_that
_expands and replaced nothing, and the reviewer showed what that cost:
change material.py's initial branch to box_molecule(basis, [basis]) and the
whole suite still passes. test-002's initial block is the larger of its two,
so its padded edge already IS the shared edge - the expanding direction, the
one that leaves atoms outside the box and corrupts the InChI, is
unexercisable by any committed fixture.

tests/unit/properties/test_utils.py covers it with synthetic bases: the cell
is sized to the largest structure, every structure is centred inside the
shared cell, and Material hands over every parsed basis rather than only the
one it is centring. That last one is what fails under the reviewer's
mutation.

The three Material tests move out of the nwchem parser module and in beside
the other Material tests; the manifest is keyed on test name, so the move is
free. Three rather than the two asked for: leaving the third behind would
have kept a Material test in a parser module for no reason.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…cell

TB-NAME-1: `basis` and `bases` are one letter apart at a call site that
passes both. Now selected_basis and parsed_bases.

TB-DOC-1/3: NwchemParser's missing lattice methods are load-bearing control
flow - material.py reads None as "molecule, derive a box" - and nothing said
so once the old docstrings went with the methods. One sentence on the class,
where a reader wondering why the methods are absent will be. Not repeated at
the guard.

Also from the review: basis() returns None with no geometry block, so say
`dict | None` as the rest of the file does; box_molecule's docstring trimmed
and its Returns made truthful; the padded cell gets a named intermediate
instead of a bare [0][0]; and the fixture comment drops the SHA it cited,
which will not survive a squash-merge.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

a488316 swept black reflows and blank-line fixes into seven files no
nwchem change touches. Restore them to main byte-for-byte.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…xtures

Cuts the branch from 405 insertions over origin/main to 296. What went is
prose and code; the data and the docstrings stayed.

Code and prose:
- the three Material tests merge into one, since all three read the same
  fixture, and the manifest loses two entries with them
- test_converts_a_block_printed_in_atomic_units folds into the bond-length
  test, which asserts the same conversion more strictly
- test_nwchem_structures_of_optimization goes, with its two basis constants.
  Its numbers are still pinned one layer up: the cell edge comes from
  test-002's initial block and the crystal coordinates from its final one,
  so the serialized-material test covers both
- geometry_blocks() and box_molecule() keep the why-prose and lose the
  Args/Returns boilerplate; settings.py's regex comments keep their
  constraint and lose the restatement

Data was not golfed. The unit fixtures print a real NWChem geometry block
- header, column titles, dashed rule, column-aligned rows, atomic-mass tail
- and the bond-length test carries test-001's three rows verbatim with both
O-H distances asserted, because that is what the log contains. basis() keeps
its Example, as its siblings have one, and the two parser delegates keep the
full docstring the methods above them use.

Coverage went up, not down: the truncation guard had no test at all and now
has one. All four mutations still fail - centering, per-block sizing, the
material.py guard, and the initial call site. The serialized material for
test-001 and test-002 is byte-identical to 04f4d7d.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>