test(#285) regression test for horizontal-bisecting on trigonal-rhombohedral

[prjemian]

• closes Horizontal-bisecting modes cannot reach trigonal-rhombohedral (1, 1, 0) #285

Adds tests/test_regression_issue_285.py covering the three
horizontal-bisecting cases from the issue body:

fourch  / bisecting
psic    / bisecting_vertical
kappa6c / bisecting_horizontal

Each test asserts that forward(1, 1, 0) on a trigonal-rhombohedral
cell (a=b=c=7.0, alpha=beta=gamma=72) returns at least one solution
that round-trips through inverse() back to (1, 1, 0).

Investigation summary

Replayed the issue's verbatim self-contained reproducer in a clean
Python 3.14.4 / numpy 2.4.6 venv across released versions:

version	fourch / bisecting	psic / bisecting_vertical	kappa6c / bisecting_horizontal
v0.10.0	2 valid solutions	2 valid solutions	0 solutions (real failure)
v0.10.1	2 valid solutions	2 valid solutions	0 solutions (real failure)
v0.11.0 (PyPI)	2 valid solutions	2 valid solutions	2 valid solutions
main (post-#284)	2 valid solutions	2 valid solutions	2 valid solutions

Every returned solution round-trips correctly through inverse().

• The fourch bisecting and psic bisecting_vertical claims in the
  issue do not reproduce on any released version. Most likely
  explanation: the issue's "self-contained reproducer" was
  synthesized from a downstream cross-validation harness that uses a
  bootstrapped UB (from calc_UB of two seed reflections at
  chosen positions), and the simplification to ub_identity lost
  the specific UB orientation that produced the zero-solution branch
  selection.
• The kappa6c bisecting_horizontal zero-solution failure was real
  on v0.10.x and was eliminated by the issue Audit order of rotations used in forward() / inverse() #280 composition-order
  overhaul shipped in v0.11.0 (most likely the BL1967 B-matrix
  correction in phase 3, which changed the reciprocal-lattice
  geometry for non-orthogonal cells).

No code change in the package is required. The new regression test
pins down the working behavior so any future regression on the
rhombohedral path is caught immediately.

Notes

• No CHANGES.md update (test-only PR per AGENTS.md policy).
• No hot-path source files touched, so slow benchmark suite was not
  re-run; full pytest suite passes locally (2574 passed, 2 skipped, 3
  deselected, 100 % coverage).

Contributed by: OpenCode (argo/claudeopus47)