fix(#304): implement ReferenceConstraint residuals and verify post-solve

[prjemian]

• closes ReferenceConstraint solvers still stubbed in v0.11.2: psi/incidence/emergence/specular modes raise NotImplementedError #304

Reframing the issue

Issue #304 reported that ReferenceConstraint.evaluate() / .is_satisfied()
were stubs raising NotImplementedError, concluding the reference-constraint
solvers were never finished. Investigation showed the premise is inverted:

• The forward solvers for every reference-constraint mode
  (incidence, emergence, specular, psi, omega) already exist in
  forward.py (_solve_surface, _solve_psi_mode, _solve_omega_mode,
  _solve_free_detectors, _solve_qaz_mode). The dispatcher routes those
  modes to the dedicated solvers and never calls the stubbed methods.
• ConstraintSet.is_implemented() / ReferenceConstraint.is_implemented()
  already return True once the reference vector (azimuth / surface_normal)
  is set.
• Every affected mode (psic, fourcv, sixc, zaxis, s2d2, kappa*)
  solves end-to-end through forward() once its reference vector is set —
  verified across all of them.

The real defect was therefore dead code + stale references, not a missing
solver. The two stub methods were reachable only from two tests that asserted
they raise, and the comments pointed at the long-closed "Issue J" (#157).

What this PR does

• mode.py: evaluate() returns real pseudo-angle residuals via reference.py
  (incidence/emergence/specular/psi/naz/omega); is_satisfied()
  delegates to it; the stale "Issue J" comments are removed.
• forward.py: reference constraints are wired into the post-solve verification
  loop (_validate_solutions), so reference modes get the same residual sanity
  check as bisect/sample constraints. psi is intentionally skipped there
  — it is enforced upstream as a validation filter in _solve_psi_mode, and its
  motor-frame residual is subject to ±360° azimuthal wraparound, so a naive
  per-solution check is unreliable. Unset reference vectors are guarded.
• Tests: the two NotImplementedError stub tests are replaced with real residual
  tests; the is_implemented parametrization is corrected (psi is True when
  azimuth is set; naz is False). A new cross-module regression file
  tests/test_regression_issue_304.py confirms every affected mode solves
  without spurious ConstraintViolation and that the loop rejects a bad
  reference solution.

Known remaining gap (out of scope, not a regression)

naz has no forward solver — forward.py has no _solve_naz_mode, and
ReferenceConstraint.is_implemented() deliberately returns False for it.
evaluate() will compute a naz residual (used for verification only), but no
demo geometry's YAML declares a naz mode, so nothing is broken. Per the
pre-1.0 no-volunteered-follow-up policy, this is noted here rather than filed as
a new issue.

QC

• Full suite: 2644 passed, 2 skipped, 100% coverage.
• pytest -m slow_benchmark: 3 passed (hot path: forward.py / mode.py).
• US-English grep on changed files: clean.
• pre-commit run: all hooks pass.

Contributed by: OpenCode (argo/claudeopus48)