Refactor `gdp_col` to fix benchmark issues

[tristantc]

Description

This Issue aims to address and refactor the gdp_col model to fix benchmark issues as outlined in Issue #60.

Solution strategy causing issues:

• Big-M MINLP reformulation
• Hull MINLP reformulation
• GLOA
• Enumeration
• LOA
• LBB
• RIC

Branch

This PR targets the benchmark_zd branch.

Steps to Reproduce

1. Run the benchmark tests for the gdp_col model before applying the changes.
2. Apply the changes from this PR.
3. Re-run the benchmark tests to verify performance improvements.

Notes

• Ensure that all sub-solvers are tested thoroughly after refactoring.
• Additional optimizations may be required based on benchmark results.

---

Feel free to adjust any details as necessary before submitting the PR.

[bernalde]

Warning capture from PR #58 benchmark tooling found 29 Pyomo deprecation candidates for gdp_col: indicator variables are initialized with numeric 0/1, producing messages like implicitly casting 'tray[2].indicator_var' value 1 to bool and implicitly casting 'no_tray[2].indicator_var' value 0 to bool. This is part of the cross-cutting deprecation cleanup tracked in #55. Full generated table: #55 (comment)

[bernalde]

Benchmark evidence from PR #58 local/global Big-M smoke run (2026-05-10).

Configuration: gdplib-benchmark run --strategies gdp.bigm --timelimit 300; local profile = GAMS/DICOPT with IPOPTH/Gurobi role defaults; global profile = GAMS/BARON. Outputs are generated artifacts and are not committed.

Profile	Solver	Termination	Status	Objective	Lower bound	Upper bound	User time (s)
local	DICOPT	optimal	ok	20081.396823	20081.396823	20081.396823	5.012
global	BARON	optimal	ok	19740.340797	9705.12692	19740.340797	300.226

Warnings captured in build/Big-M transform pass:

• 31 unique warning event(s) in build/transform capture; 29 deprecation candidate(s); categories: pyomo.core=31; examples: implicitly casting 'no_tray[10].indicator_var' value 0 to bool | implicitly casting 'no_tray[11].indicator_var' value 0 to bool | implicitly casting 'no_tray[12].indicator_var' value 0 to bool

Follow-up:

• Resolve the warning/deprecation items or link them to Fix Pyomo.GDP and GDPOpt deprecations #55 if handled centrally.

Run artifacts: benchmark_runs/pr58_bigm_local_5min_20260510, benchmark_runs/pr58_bigm_global_5min_20260510, benchmark_runs/pr58_bigm_warnings_transform_20260510.

[bernalde]

During the issue #55 deprecation-cleanup benchmark rerun on branch fix/issue-55-pyomo-deprecations, gdp_col still emits non-deprecation Pyomo bound-initialization warnings during model construction/preflight:

WARNING (W1002): Setting Var 'H_L_spec_feed[benzene]' to a numeric value `0` outside the bounds (0.1, 16).
WARNING (W1002): Setting Var 'H_V_spec_feed[benzene]' to a numeric value `0` outside the bounds (30, 56).

I am not handling this in #111 because it is separate from the Pyomo.GDP/GDPOpt deprecation cleanup and changing initialization values could affect the benchmark comparison. Suggested follow-up for this issue: align the feed enthalpy initial values and bounds with the model intent, then rerun construction plus the PR #58 5-minute Big-M local/global smoke profiles for gdp_col.

[bernalde]

Coordination note from the PR #124 review (see the review and follow-up comment there):

• PR Fix gdp_col hull reformulation #124 partially resolves this tracker: hull now transforms cleanly (with a solver-free regression test) and local GDPopt/GLOA reports optimal at 20916.392216. Its Closes #66 link should be read as Refs #66; if the merge auto-closes this issue, reopen it for the items below.
• Remaining items this issue still owns:
  1. gdpopt.lbb — blocked upstream by GDPopt LBB time-limit path calls stale _get_final_results_object Pyomo/pyomo#3941.
  2. BARON declares the hull-transformed model infeasible in 0.42 s while Gurobi finds a 22355.24 incumbent on the same transformed model — investigate whether PR Fix gdp_col hull reformulation #124's exp-form Antoine rewrite trips a BARON presolve/domain issue.
  3. A validated global optimum for the hull-reformulated MINLP (the BARON GLOA run reached maxTimeLimit with an 8000 → 22355 gap).