Reduced cost interpretation for capacity investment variables with operational costs

I'm modeling an energy system for a fixed target year where I decide how much capacity to build for different generation technologies and how to operate them.

• Decision variables: continuous, non-negative capacity investment for each technology (MW), plus operational decisions (generation levels, etc.)
• Objective: minimize total system cost (investment costs + operational costs)
• Example: Nuclear has an investment cost of X (CHF/MW) and operational cost of 10 (CHF/MWh)

For a non-basic variable in a minimization problem, the reduced cost indicates how much its objective coefficient would need to decrease before it enters the basis at a positive level. This sensitivity is valid only locally around the current solution. In my case, for a capacity investment variable (e.g., nuclear capacity in MW), the objective coefficient is the investment cost (CHF/MW), so the reduced cost tells me how much the investment cost would need to drop.

I want to use sensitivity information from Gurobi to systematically identify which input parameters are most influential, so I can prioritize updating uncertain cost estimates starting with the most important parameters first.

The problem: In my base case, nuclear capacity stays at zero. The reduced cost suggests a ~5% reduction in investment cost would bring it into the solution. However, when I actually reduce the investment cost by 5% and re-solve, nuclear still doesn't enter (even for around 50%). This discrepancy makes sense—reduced costs are only valid locally around the solution. But here's the interesting part: when I run a modified scenario where I set nuclear's operational cost to 0 CHF/MWh (keeping investment cost unchanged), the reduced cost in that solution is ~55%. When I then apply a 57% reduction to the investment cost in the original base case, nuclear does enter the solution.

Additional observation: For gas-fired plants (lower investment cost, higher operational cost than nuclear), the base case reduced cost is very low, and the "set operational cost = 0" experiment doesn't provide more meaningful reduced cost values either.

Why does the reduced cost become more "accurate" (better predicts the actual cost reduction needed) when operational costs are removed? Is there a theoretical explanation for why this approach works for capital-intensive technologies (nuclear) but not fuel-intensive ones (gas)?

More broadly, what are better ways to use Gurobi's sensitivity information to systematically identify which parameters merit deeper investigation when decision variables have both investment and operational cost components in the objective function?