Fix gdp_col hull reformulation

[bernalde]

Summary

• Initialize gdp_col feed liquid and vapor enthalpy variables from the existing feed enthalpy expressions so construction no longer emits out-of-bounds Pyomo W1002 warnings.
• Rewrite the active-tray Antoine vapor-pressure relation in algebraically equivalent exponential form so Pyomo's default gdp.hull transformation no longer evaluates log(0) in the hull perspective.
• Add focused regression coverage for feed enthalpy initialization and solver-free gdp.hull reformulation.
• Add an optional Pixi gurobi environment for direct gurobipy/Pyomo Gurobi checks without making licensed solver bindings part of the default environment.

Tests run

• /home/bernalde/.pixi/bin/pixi run pytest tests/test_pyomo_deprecations.py -v --tb=short -> 6 passed.
• /home/bernalde/.pixi/bin/pixi run pytest tests/test_module_imports.py -v --tb=short -> 68 passed.
• /home/bernalde/.pixi/bin/pixi run test -> 279 passed, 1 skipped.
• /home/bernalde/.pixi/bin/pixi run lint -> passed with exit code 0. The configured non-fatal flake8 inventory still reports existing repository style findings under --exit-zero.
• /home/bernalde/.pixi/bin/pixi run gdplib-benchmark warnings --instances gdp_col --strategies gdp.hull --mode transform --no-check-solvers --run-id issue66_hull_warning --metadata-dir /tmp/gdplib-issue66-warning-metadata -> 0 warning events, 0 unique warning rows, 0 deprecation candidate events, 0 capture errors.
• /home/bernalde/.pixi/bin/pixi install -e gurobi -> installed the optional Gurobi Pixi environment.
• /home/bernalde/.pixi/bin/pixi run -e gurobi python -c "import gurobipy as gp; from pyomo.environ import SolverFactory; print(gp.gurobi.version()); print(SolverFactory('gurobi_direct').available(False))" -> (12, 0, 3), True.
• /home/bernalde/.pixi/bin/pixi run -e gurobi pytest tests/test_release_workflow.py -v --tb=short -> 5 passed.
• Direct Pyomo/Gurobi smoke solve through gurobi_direct -> ok, optimal, objective 4.0.
• git diff --check -> passed.

Solver-backed benchmark evidence

The benchmark comparisons now include GAMS/Gurobi cases for all direct GDP reformulations in this PR scope (gdp.bigm and gdp.hull), in addition to the local DICOPT and global BARON profiles. The GAMS/Gurobi rows below were run through the benchmark CLI with --solver-profile gams-gurobi, so they use the same single-threaded GAMS options as the other campaign rows.

Commands:

/home/bernalde/.pixi/bin/pixi run gdplib-benchmark run --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-local --run-id issue66_local_5min_20260511
/home/bernalde/.pixi/bin/pixi run gdplib-benchmark run --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-baron --run-id issue66_global_5min_20260511
/home/bernalde/.pixi/bin/pixi run -e gurobi gdplib-benchmark run --instances gdp_col --strategies gdp.bigm gdp.hull --timelimit 300 --solver-profile gams-gurobi --run-id issue66_gurobi_5min_20260511 --no-summary

Profile	Strategy	Solver	Termination	Status	Objective	Lower bound	Upper bound	User time (s)
gams-local	gdp.hull	DICOPT	intermediateNonInteger	warning	10273.524906	13110.109796	10273.524906	3.516
gams-gurobi	gdp.bigm	Gurobi	maxTimeLimit	ok		11014.509755		300.289
gams-gurobi	gdp.hull	Gurobi	feasible	ok	22355.240680	19223.858427	22355.240680	300.047
gams-baron	gdp.hull	BARON	infeasible	ok		1000.000000		0.423
gams-local	gdpopt.gloa	DICOPT/IPOPTH/Gurobi roles	optimal	ok	20916.392216	20916.392216	20916.392216	29.078
gams-baron	gdpopt.gloa	BARON roles	maxTimeLimit	ok	22355.240679	8000.000000	22355.240679	302.049
gams-local	gdpopt.lbb	DICOPT/IPOPTH/Gurobi roles	failure	error
gams-baron	gdpopt.lbb	BARON roles	failure	error

Additional Gurobi checks:

• GAMS/Gurobi accepts both direct reformulations as nonconvex MINLPs.
• For gdp.hull, GAMS/Gurobi finds a feasible incumbent at objective 22355.24067954157 within 300 seconds.
• For gdp.bigm, Gurobi reaches the 300-second time limit with no incumbent (Solution count 0, Best objective -) and a lower bound of 11014.509754975745. The GAMS wrapper reports OBJVAL/upper bound as 17000.0, but the Gurobi log does not support treating that as a valid incumbent, so the table leaves objective and upper bound blank.
• Pyomo's direct Gurobi interfaces are available in the optional Pixi environment, but they cannot currently write this nonlinear transformed hull instance directly: gurobi rejects nonlinear LP-format output, while gurobi_direct and appsi_gurobi reject expression degree None.
• Future solver-backed benchmark comparison tables should include the gams-gurobi profile for each direct reformulation alongside gams-local and gams-baron when Gurobi is licensed locally.

Notes

• gdp.hull now transforms cleanly, but direct DICOPT and BARON behavior remains solver/formulation sensitive: DICOPT stops with intermediateNonInteger, while BARON reports the transformed hull model infeasible despite GAMS/Gurobi finding a feasible incumbent.
• gdpopt.lbb remains blocked by Pyomo GDPopt runtime behavior: GDP_LBB_Solver calls stale _get_final_results_object() in the time-limit path. I opened upstream Pyomo issue GDPopt LBB time-limit path calls stale _get_final_results_object Pyomo/pyomo#3941 with a minimal reproducer.
• Generated benchmark artifacts were not committed. Per-case JSON/log files from the local checks remain under ignored benchmark output directories.

Closes #66

[bernalde]

Local solver campaign update

I ran bounded 5-minute solver campaigns for gdp_col on this branch to check the remaining issue #66 strategy surface after the hull transformation fix.

Commands:

/home/bernalde/.pixi/bin/pixi run gdplib-benchmark preflight --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-local --build-models --check-solvers
/home/bernalde/.pixi/bin/pixi run gdplib-benchmark run --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-local --run-id issue66_local_5min_20260511
/home/bernalde/.pixi/bin/pixi run gdplib-benchmark preflight --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-baron --build-models --check-solvers
/home/bernalde/.pixi/bin/pixi run gdplib-benchmark run --instances gdp_col --strategies gdp.hull gdpopt.gloa gdpopt.lbb --timelimit 300 --solver-profile gams-baron --run-id issue66_global_5min_20260511

Environment metadata:

• Python 3.12.13
• Pyomo 6.10.0
• GAMS executable: /home/bernalde/packages/gams51.2_linux_x64_64_sfx/gams
• Git head: 7c47e62039baa354cbcfa0f7846500dcc84f8faa

Results:

Profile	Strategy	Solver	Termination	Status	Objective	Lower bound	Upper bound	User time (s)
gams-local	gdp.hull	DICOPT	intermediateNonInteger	warning	10273.524906	13110.109796	10273.524906	3.516
gams-local	gdpopt.gloa	DICOPT/IPOPTH/Gurobi roles	optimal	ok	20916.392216	20916.392216	20916.392216	29.078
gams-local	gdpopt.lbb	DICOPT/IPOPTH/Gurobi roles	failure	error
gams-baron	gdp.hull	BARON	infeasible	ok		1000.000000		0.423
gams-baron	gdpopt.gloa	BARON roles	maxTimeLimit	ok	22355.240679	8000.000000	22355.240679	302.049
gams-baron	gdpopt.lbb	BARON roles	failure	error

Failure details:

• Both LBB runs fail with the same Pyomo GDPopt exception: AttributeError: 'GDP_LBB_Solver' object has no attribute '_get_final_results_object'. Did you mean: '_get_final_pyomo_results_object'?
• Failure logs:
  • gdplib/gdp_col/benchmark_result/issue66_local_5min_20260511/gdpopt.lbb_gams_dicopt_failure.log
  • gdplib/gdp_col/benchmark_result/issue66_global_5min_20260511/gdpopt.lbb_gams_baron_failure.log

Runner note:

• Solver calls completed, but gdplib-benchmark run exits during summary generation because _generate_summary() imports missing module generate_benchmark_summary_all. Per-case JSON/CSV artifacts were still written under benchmark_runs/issue66_* and gdplib/gdp_col/benchmark_result/issue66_*.

Interpretation:

• The PR fix moves gdp.hull past the previous transformation-time math domain error.
• gdpopt.gloa with the local profile now solves to an optimal local result in this bounded run.
• Direct hull solve quality is still not clean: local DICOPT stops intermediateNonInteger, and BARON reports the transformed hull model infeasible.
• LBB remains blocked by a GDPopt runtime error that appears independent of the gdp_col formulation.

[bernalde]

Follow-up: GDPopt LBB failure root cause

I dug further into the gdpopt.lbb failure reported in the local solver campaign.

Root cause:

• This is a Pyomo GDPopt LBB time-limit exit bug, not a gdp_col formulation failure.
• In Pyomo 6.10.0, pyomo/contrib/gdpopt/branch_and_bound.py::GDP_LBB_Solver._solve_gdp() calls self._get_final_results_object() only in the reached_time_limit(config) branch.
• The base GDPopt algorithm class no longer provides _get_final_results_object; it provides _get_final_pyomo_results_object().
• The same stale call is present on upstream Pyomo main as of this check.

Evidence:

• The installed class has hasattr(GDP_LBB_Solver, "_get_final_results_object") == False and hasattr(GDP_LBB_Solver, "_get_final_pyomo_results_object") == True.
• A minimal GDP model reproduces the same AttributeError by forcing the LBB time-limit branch, without using gdp_col or any GAMS solve.
• A local monkeypatch aliasing _get_final_results_object = _get_final_pyomo_results_object lets the time-limit path return a normal Pyomo results object.

What the monkeypatch shows for gdp_col:

• With a 30-second local-profile LBB run, the patched solver returns maxTimeLimit, no feasible incumbent, and infinite bounds.
• So the method-name fix would make benchmark reporting cleaner, but it does not make LBB solve gdp_col.

Related upstream context:

• Pyomo issue GDPopt LBB doesn't solve trivial GDP Pyomo/pyomo#2483 tracks broader known LBB problems.
• I did not find an existing upstream issue/PR for this exact _get_final_results_object stale call.

Practical implication for this PR:

• gdpopt.lbb should remain tracked as blocked by Pyomo LBB runtime behavior, not by this gdp_col hull transformation fix.
• A separate follow-up could either report/fix the Pyomo method-name bug upstream or add a temporary compatibility shim in gdplib-benchmark if maintainers want local campaign reports to classify LBB as maxTimeLimit instead of a Python exception.

[bernalde]

Maintainer review of head e1018a6.

Both changes are correct, verified line by line. The feed-enthalpy initializers now come from the existing feed_liq_enthalpy_expr/feed_vap_enthalpy_expr Expressions (well-defined at construction; hand-computed values ≈10.1 and ≈34 land inside bounds, eliminating the W1002 warnings). The Antoine/Wagner rewrite from log form to exp form is algebraically equivalent given Pvap in [1e-3, 5] > 0 and Pvap_X in [0.25, 0.5], and under hull disaggregation the exp form stays evaluable at the origin where the base log(Pvap) failed; the constraint is valid in both its disjunct and global-block contexts, and the involved disjunct variables are globally bounded as hull requires.

Findings:

Blocking: closes an unresolved tracker — the body says "Closes #66", but issue #66's checklist has three open items (Hull, GLOA, LBB) and this PR's own benchmark table shows gdpopt.lbb still failing on both profiles (upstream Pyomo/pyomo#3941) and BARON declaring the hull-transformed model infeasible. Merging as-is auto-closes a tracker that is not resolved. Fix: treat the link as Refs #66 (recorded in a PR comment since the body stands), or close #66 only alongside a follow-up owning LBB and the BARON discrepancy.

Nonblocking: stale body bullet — the body's "Add an optional Pixi gurobi environment" is stale: [feature.gurobi] already exists at the merge-base (via PR #144); this diff touches only two files. The corrected scope should be recorded so reviewers are not misled.

Nonblocking: test placement — test_gdp_col_reformulates_with_hull and its sibling are not deprecation tests; placing them in tests/test_pyomo_deprecations.py is mild scope creep. Move both to a new tests/test_gdp_col.py.

Question: BARON infeasibility — BARON declares the hull-transformed model infeasible in 0.42 s while Gurobi finds a 22355.24 incumbent on the same model. Is the exp rewrite triggering a BARON domain/presolve issue worth a tracked follow-up on #66?

Evidence note: the gams-local gdpopt.gloa optimal 20916.392216 row (LB = UB) is the strongest solver evidence and supports the GLOA fix; the hull rows prove hull transforms and is solvable in principle, not that it is robustly fixed.

Summary — Blocking: 1. Nonblocking: 2. Questions: 1. Tests: static line-by-line verification of both fixes, including bound/domain arithmetic; CI is green at this head (6/6 checks). I would not merge this until the blocking issue above is addressed (an issue-linkage record; the code itself is sound). Posting as COMMENT (author account); the main ruleset requires one approving review from another maintainer.

[bernalde]

Addressed the review of head e1018a6 in commit a1c16ec:

• test-placement: the two gdp_col regression tests moved from tests/test_pyomo_deprecations.py to a new tests/test_gdp_col.py; the deprecation module is back to deprecation-only scope.
• closes-vs-refs-66 (Blocking, metadata): recorded here since the body stands — this PR partially resolves issue Refactor gdp_col to fix benchmark issues #66 (hull transformation fixed with a test; local GLOA optimal at 20916.392216), while gdpopt.lbb remains blocked upstream (GDPopt LBB time-limit path calls stale _get_final_results_object Pyomo/pyomo#3941) and the BARON-infeasibility discrepancy is unresolved, so the link should be read as Refs #66. If merged as-is, the auto-closure of Refactor gdp_col to fix benchmark issues #66 should be reverted (reopen) or explicitly accepted with the remaining items tracked; a coordination note is posted on Refactor gdp_col to fix benchmark issues #66.
• stale-gurobi-bullet: corrected scope for the record — the optional Pixi [feature.gurobi] environment already existed at the merge-base (PR [codex] Consolidate benchmark support changes #144); this PR's diff is exactly two files (gdplib/gdp_col/column.py, tests), and body bullets referring to the gurobi environment describe pre-existing state, not changes made here.
• baron-infeasible-followup (question): tracked on Refactor gdp_col to fix benchmark issues #66 via the coordination note — BARON declares the hull-transformed model infeasible in 0.42 s while Gurobi finds a 22355.24 incumbent on the same model; worth investigating whether the exp-form rewrite trips a BARON presolve/domain issue.

Tests: pixi run pytest tests/test_gdp_col.py tests/test_pyomo_deprecations.py — 6 passed; black --check clean. CI for the new head was starting when this was posted; the previous head was green (6/6).

[bernalde]

Follow-up on the BARON-infeasibility question from the review: it does not reproduce at this head. Hull + GAMS/BARON (120 s, optcr=1e-6, threads=1) returns the 22355.24 incumbent with max constraint violation 4.1e-9, and Big-M + BARON returns a feasible 19449.85 — which, for this minimize model, is strictly better than the body's "GLOA optimal 20916.39", so that gams-local GLOA label is falsified as a global claim. Details and re-baselining follow-ups recorded on #66.