Fix GCC solution loss during recomputation

[zayenz]

Domain-level global cardinality propagation could lose solutions in two related matching-state paths.

First, assigned variables were counted in both the upper- and lower-bound narrowing passes when recomputation replayed several choices before propagation. Second, when a variable cardinality maximum shrank by more than the number of currently matched edges, the propagator retained an impossible conflict count. That stale conflict mark could remove supported value edges.

Count assigned variables only in the upper-bound pass, and cap conflict bookkeeping by the number of genuinely matched edges (excluding already-counted assigned occurrences). This retains the existing rematching strength without leaving stale conflicts.

The FlatZinc regression now covers both the original deterministic issue model (26 solutions, including a formerly missing assignment) and the minimized random-search model (11 solutions, including its formerly missing assignment).

Verified with:

• both focused FlatZinc regressions;
• the original model at recomputation distances 1, 2, 4, 8, and 16;
• the minimized model across those five distances and seeds 0 through 9;
• all 24 integer GCC test families at five randomized iterations;
• the exact variable-cardinality domain-propagation seed used to check fixpoint equivalence after cloning.

Closes #166.