Fix DFTB+ Hessian parsing and harden coordinate/frequency readers

[galjos]

Follow-up to #46 (shared parsing via PQAnalysis). Verifying that refactor surfaced one critical regression plus a couple of lower-severity issues, fixed here with regression coverage.

Critical: Hessian reading was broken on real DFTB+ output

DFTB+ writes hessian.out as a flat stream of second derivatives wrapped at a fixed number of values per line (with a ragged final line per matrix row), not as a square grid. The refactor routed it through a reader that requires an already-square numpy.loadtxt result, so it raised on every real file:

• uniform-width output loaded as (n, 4) and failed the square check;
• ragged output made numpy.loadtxt itself fail.

This is invisible to CI because the integration tests need the dftb+/modes binaries and are skipped, while the unit test fed a synthetic square grid DFTB+ never emits. Restored flat-token parsing that reads every value and reshapes to (3N, 3N), keeping the explicit size validation. Verified bit-for-bit against real 72x72 and 78x78 Hessians.

xyz coordinate precision

read_xyz routed coordinates through XYZFrameReader, which stores positions as float32, perturbing every value by ~1e-7 relative to the source and to read_gen (float64). Now reads the coordinate columns in float64 while keeping XYZFrameReader for validation, atom count, symbols and cell detection.

Smaller fixes

• read_xyz no longer masks a missing file as "Invalid XYZ coordinate file."; it raises FileNotFoundError again.
• Modes.read stops at a trailing tag section (e.g. saved_modes :integer:..) instead of casting the section header to float and crashing.

Test coverage

• Regression test for the Hessian reader using the authentic ragged wrapped layout (asserts numpy.loadtxt cannot parse it, so the fixture can't be reduced to a square grid).
• Small fixtures (tests/data/calculator/modes/) plus an integration test that drives Hessian.read -> Modes -> wave numbers end to end with only the modes binary mocked, closing the no-binary gap.
• float64-precision and missing-file tests for read_xyz; a trailing-section test for Modes.read.

134 passed, 5 skipped (the 5 still need the DFTB+ binaries); pylint 7.94/10.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.63%. Comparing base (505074e) to head (65a9a26).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #47      +/-   ##
==========================================
+ Coverage   93.61%   93.63%   +0.01%     
==========================================
  Files          21       21              
  Lines        1191     1194       +3     
==========================================
+ Hits         1115     1118       +3     
  Misses         76       76              

Flag	Coverage Δ
unittests	93.63% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[galjos]

Validated end to end against the real DFTB+ binaries (DFTB+ 24.1 + MODES 0.03, 3ob-3-1 Slater-Koster set), not just mocks.

Full suite with the binaries enabled: 139 passed, 0 skipped (the five integration tests that normally skip without dftb+/modes now run: real geometry optimization, real second-derivatives Hessian, real modes frequencies, and the full thermo pipeline).

Concrete evidence on freshly generated output for water:

• The real hessian.out is the wrapped/ragged layout (line widths {1, 4}, 81 values -> 9x9). The reader in this PR parses it; the pre-fix square-grid reader raises ValueError: the number of columns changed from 4 to 1 at row 3 on that exact file — i.e. the regression reproduces on genuine DFTB+ output, and the fix resolves it.
• The real vibrations.tag is frequencies-only (single :real: section, no saved_modes), matching the committed fixture; the trailing-section handling in Modes.read is therefore defensive rather than load-bearing for the default flow.
• Resulting water frequencies are physically sensible (bend ~1462, stretches ~3604 / ~3876 cm^-1).