Fix solver correctness and input-validation issues from code review

- Refuse the uniform contact model when the fills form multiple
  disconnected copper groups that each touch both terminals: the
  prescribed injection split is ill-posed and the grounded system was
  singular, silently returning garbage (e.g. negative gigaohms).
  connected_restrict now reports the component count; a power-balance
  backstop (SolverError) catches any other inconsistent solve.
- Connect via/pad barrels to the nearest fill copper within the pad
  footprint (+1 cell) instead of only the exact center cell, so
  thermal-relief spokes still stitch layers; barrels that reach fill on
  fewer than two layers are warned about. ViaLink gains pad_nm
  (extracted from the padstack, JSON-roundtripped).
- Validate dialog input on OK (layers, current > 0, cell > 0, parseable
  frequency, extra Cu >= 0) with an inline error instead of silently
  substituting defaults; parse_frequency raises on garbage; pipeline
  rejects i_test <= 0; choose_cell_size rejects non-positive overrides.
- Warn when a contact part is dropped by the connectivity restriction;
  floor instead of truncate in cell_of; correct the uniform-model
  summary line; drop an unused variable; refresh plugin.json wording.

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

tests/test_contacts.py

@@ -3,9 +3,10 @@ import numpy as np
 import pytest
 
 from fill_resistance import raster, solver
-from fill_resistance.errors import ElectrodeError
+from fill_resistance.errors import ConnectivityError, ElectrodeError
 from fill_resistance.geometry import Electrode
-from tests.util import NM, make_problem, rect_mm, sigma_s, strip_problem
+from tests.util import (NM, make_multilayer, make_problem, rect_mm, sigma_s,
+                        strip_problem)
 
 
 def _solve(problem, h_mm, model, i_test=1.0):
@@ -161,6 +162,22 @@ def test_injection_area_partition_first_wins():
     assert total == pytest.approx(1.0, rel=1e-12)
 
 
+def test_uniform_multicomponent_raises():
+    """Two disconnected sheets that each touch both terminals: the
+    uniform model would build a singular system (one ground cell, pure-
+    Neumann second component) and previously returned garbage silently
+    (e.g. negative gigaohms). It must refuse; equipotential handles it."""
+    strip1 = [(0, 0), (10, 0), (10, 1), (0, 1)]
+    strip2 = [(0, 0), (10, 0), (10, 2), (0, 2)]   # asymmetric shares
+    p = make_multilayer([[(strip1, [])], [(strip2, [])]],
+                        (0, 0, 1, 2), (9, 0, 10, 1))  # contact 'all', no vias
+    with pytest.raises(ConnectivityError, match="disconnected"):
+        _solve(p, 1.0, "uniform")
+    res, _ = _solve(p, 1.0, "equipotential")
+    assert np.isfinite(res.R_ohm) and res.R_ohm > 0
+    assert res.power_balance_rel < 1e-9
+
+
 def test_touching_ok_uniform_error_equipotential():
     p = make_problem([([(0, 0), (10, 0), (10, 10), (0, 10)], [])],
                      rect1_mm=(0, 0, 5, 10), rect2_mm=(5, 0, 10, 10))