add vibe coded regression test suite

[jonmarks12]

Add golden-output regression tests and raise coverage

Why

The FSM is fully deterministic given the same inputs and the EMT calculator, but
the test suite never validated its output — the existing tests only ran
fsm_example.py in a subprocess and asserted the exit code. That left two gaps:

1. A change could silently alter energies or geometries and still pass.
2. Subprocess runs don't count toward --cov=src/mlfsm, so coverage was stuck.

This PR adds a golden-output regression test that makes it trivial to tell
whether a change is purely efficiency/flags or actually changes numerical output,
and adds targeted unit tests to raise coverage.

What

Golden regression suite

• For each interpolation style (cart, lst, ric), a small deterministic EMT
  run produces fsm.out, compared against a committed golden file.
• Comparison is token- and numeric-tolerance aware (abs_tol=1e-5), not raw
  byte equality — CI runs on both linux-64 and osx-arm64, where the last
  printed digit can drift. Structure and text changes remain hard failures, with
  line-numbered diffs on mismatch.
• The only non-deterministic line (Date/Time:) is masked before comparison.
• Goldens regenerate via pytest tests/test_golden.py --update-goldens
  (or UPDATE_GOLDENS=1); review the diff in git diff tests/data/golden.

Coverage

• Converted the existing subprocess tests (test_script, test_cartesian,
  test_stepsize) to in-process run_fsm calls so they count toward coverage.
• Added unit tests for utils, output, interp, and cos, plus tests for the
  comparison helpers themselves.
• Coverage rose to ~78% (interp.py 100%, output.py 93%, utils.py 98%).

Files

• New: tests/conftest.py, tests/golden_utils.py, tests/test_golden.py,
  tests/test_golden_utils.py, tests/test_utils.py, tests/test_output.py,
  tests/test_interp.py, tests/test_cos.py
• New data: tests/data/reactions/diels_alder/ (frozen inputs) and
  tests/data/golden/diels_alder_{cart,lst,ric}/fsm.out
• Modified: tests/test_script.py, tests/test_cartesian.py,
  tests/test_stepsize.py (subprocess → in-process)
• pyproject.toml: per-file ruff ignores for test code

No src/mlfsm behavior changes — the goldens exist precisely to detect such changes.

Notes

• utils.float_check's docstring claims 0-D ndarrays are supported, but the code
  raises TypeError on them. Tests assert the actual behavior; aligning the two
  is a possible follow-up.
• Remaining low-coverage areas: opt.py (the in-development InternalsOptimizer
  / optcoords="ric" path) and cos.py/coords.py frontier-growth edge cases.

Verification

• pytest tests/ — 44 passed
• pytest tests/ --cov=src/mlfsm — ~78%
• ruff check ., ruff format --check ., mypy . — all clean
• Regenerating goldens is byte-identical (confirms determinism end-to-end)

[joegomes]

Tests and example update LGTM!