Phase angle difference constraints gated by sw, not the energized predicate (b*sw != 0)

[cameronkhanpour]

Raised by Copilot review on #55 (low-confidence/suppressed); confirmed to have merit.

Inconsistency. The energized-island logic (reference_buses, the topology
cache) treats a branch as energized iff b[e] * sw[e] != 0. The DC OPF phase
angle difference constraints, however, are gated by sw alone:

phase_diff_lb = @constraint(model, network.sw .* (network.A * va) .>= network.sw .* network.angmin)
phase_diff_ub = @constraint(model, network.sw .* (network.A * va) .<= network.sw .* network.angmax)

So a branch with b[e] == 0 and sw[e] != 0 (e.g. a near-zero-impedance branch,
which the DCNetwork constructor sets to b = 0 and treats as open) is counted
as de-energized by the island partition (it splits islands / adds a reference
bus), yet its angle-difference constraint still couples the endpoint angles as if
the branch were present. This contradicts the "Open lines should not constrain
angle differences" comment and the energized-topology definition used elsewhere.

Impact. Narrow. It only triggers for b[e] == 0 with sw[e] != 0, and the
coupling is usually slack because DC angles are small relative to the default
±60° window, so it rarely changes a solution. When it does bind it couples two
islands that are otherwise independent (each referenced to its own slack), which
distorts the primal and produces a spurious cross-island angle dual, so the
congestion/energy split and switching/susceptance sensitivities would be off in
that edge case.

Scope of a proper fix. Gate the angle-difference constraints by the same
energized predicate (an effective sw_eff[e] that is zero when b[e] == 0),
then mirror that gate consistently in:

• the KKT residual K_γ_lb / K_γ_ub (kkt_dc_opf.jl),
• the analytic KKT Jacobian gamma columns,
• the switching and susceptance Jacobian gamma contributions,
• the LMP congestion decomposition gate (calc_congestion_component).

A model-only change would desync the KKT system from the solved model and
corrupt sensitivities, so this needs to land as one coherent change with
finite-difference verification near b = 0. Tracking separately rather than
expanding the already-approved #55.