Fix spectralog benchmark cases

[bernalde]

Summary

• Add finite source-scale bounds for spectralog objective-value variables so GDPopt/GLOA gets a bounded relaxation instead of fallback val[j] bounds.
• Keep the existing GDP model structure and public build_model() API intact.
• Remove spectralog/gdpopt.enumerate from the reusable PR58 local benchmark campaign and document why: 30 independent disjunctions imply 2^30 realizations, and Pyomo currently materializes the enumeration list before enforcing time_limit.
• Add focused spectralog tests for the derived bounds and Big-M/Hull transformation paths.

Tests run

• pixi run pytest tests/test_spectralog.py tests/test_benchmark.py::test_committed_pr58_local_cases_cover_default_matrix -v --tb=short -> 4 passed.
• pixi run gdplib-benchmark run --instances spectralog --strategies gdpopt.gloa --timelimit 60 --solver-profile gams-local --run-id issue75_gloa_fixed --no-summary --no-skip-existing --fail-fast -> 1 row, 0 failures; GDPopt/GLOA optimal in 2 iterations, LB = UB = 15.540610755468409.
• pixi run pytest tests/test_module_imports.py -v --tb=short -> 74 passed.
• pixi run pytest tests/test_benchmark.py -v --tb=short -> 21 passed.
• pixi run lint-black -> passed.
• pixi run lint-flake8-critical -> passed, 0 critical errors.
• pixi run lint-flake8 -> exit 0; existing style inventory remains.
• pixi run typos --config ./.github/workflows/typos.toml -> passed.
• pixi run test -> 338 passed, 1 skipped.
• git diff --check -> passed.

Notes

• I did not run spectralog with gdpopt.enumerate to completion. A focused reproduction exceeded the requested 60-second solve limit before reaching a solve result, matching the known Pyomo behavior that enumeration materializes all discrete realizations before honoring time_limit.
• Issue Plan PyPI release workflow and Pixi platform support #104 is closed release/Pixi policy work and this PR leaves packaging, release workflow, and Pixi platform support unchanged.

Closes #75

[bernalde]

Benchmark follow-up for spectralog after adding finite val[j] bounds.

I ran the requested 5-minute matrix for this instance: reformulations plus GDPOpt except gdpopt.enumerate, across local, Gurobi, and global solver profiles.

Run id:

issue75_spectralog_5min_matrix_20260514

Result summary:

Strategy	Local	Gurobi	Global
gdp.bigm	optimal, obj 12.089262, ~0.4s	optimal, obj 12.089262, ~6.7s	optimal, obj 12.089262, ~2.3s
gdp.hull	optimal, obj 12.089262, ~0.2s	optimal, obj 12.089262, ~6.8s	optimal, obj 12.089262, ~2.7s
gdpopt.loa	maxTimeLimit, UB 59.326029, LB 4.0	maxTimeLimit, UB 59.326029, LB 4.0	maxTimeLimit, UB 12.089262, LB 4.0
gdpopt.gloa	optimal, obj 15.540611, ~3.6s	optimal, obj 15.540611, ~3.4s	optimal, obj 15.540611, ~3.5s
gdpopt.lbb	failed	optimal, obj 12.089262, ~148.2s	optimal, obj 12.089262, ~87.7s
gdpopt.ric	maxTimeLimit, UB 13.275276, LB 4.0	maxTimeLimit, UB 13.275276, LB 4.0	maxTimeLimit, UB 80.634986, LB 4.0

The local gdpopt.lbb failure is solver routing rather than a Spectralog construction failure:

ValueError: GAMS writer passed solver (dicopt) unsuitable for model type (nlp)

The suspicious result is GLOA. It certifies 15.540611, while direct reformulations and LBB with Gurobi/BARON find 12.089262. This appears consistent with Pyomo/pyomo#3939, where gdpopt.gloa currently uses the concave MC++ slope for both concave and convex affine cuts.

I checked the installed Pyomo source in this environment and confirmed it has that exact ccSlope reuse pattern. I then ran a temporary control monkeypatch that uses ccSlope for the concave lower cut and cvSlope for the convex upper cut. With that patch, Spectralog GLOA no longer false-converged:

termination: maxTimeLimit
iterations: 127
lower_bound: 6.0
upper_bound: 12.08926220096831
relative_gap: 50.36918%
time: 300.16660s

So the Pyomo fix would not make this PR solve Spectralog globally with GLOA in 5 minutes, but it would make the result conservative: GLOA times out with the correct incumbent and an open gap instead of incorrectly reporting optimal at a worse solution. For this PR, I would treat the reformulation and LBB results as the reliable benchmark evidence, and flag unpatched GLOA certificates for this nonconvex instance as unreliable until Pyomo/pyomo#3939 is fixed.

[bernalde]

Maintainer review of head b36cb3f.

The bound derivation is valid: with R identity in the committed data, val[j] = sum_k r_k^2 >= 0, and with P in [0, 1000], A >= 0, the per-k corner maximum gives an upper bound over the whole variable box — a superset of the feasible region — so the bounds cannot cut off any feasible point. The removal of the spectralog,gdpopt.enumerate row from pr58_local.csv is campaign configuration, not committed results, benchmark_cases/ already exists on main, and the cited upstream Pyomo/pyomo#3953 exists and matches the stated cause (2^30 enumeration; 3 compounds x 10 wave numbers = 30 two-disjunct disjunctions checks out). The strengthened test_committed_pr58_local_cases_cover_default_matrix adds exact set equality on top of the count, strictly stronger than base.

Findings:

Nonblocking: bound helper assumptions — _val_bounds silently assumes R is identity and A, C >= 0; the sum-of-squares decomposition breaks if m.R ever gains off-diagonals. Also max_absorbance is loop-invariant in k (hoist it) and misnamed — it bounds the modeled signal sum P*A, not absorbance. Hoist, rename (e.g. max_modeled_signal), and state the identity-R assumption in a comment or assertion.

Nonblocking: vacuous assert — assert not any(... LogicalConstraint ...) in tests/test_spectralog.py is vacuously true: build_model() declares no LogicalConstraint (only BooleanVar + associate_binary_var). Drop the line or comment it as a forward guard; the Disjunction/Disjunct deactivation asserts are the real checks.

Question: optimum discrepancy — the module docstring (untouched) states "Optimal value: 12.0893" while this PR's GLOA run proves LB=UB=15.5406 on this implementation. Since the added bounds provably contain all feasible points, they cannot explain the gap. Is 12.0893 from the source papers' differing formulation/scaling, or is a subsolver bound wrong? Reconcile — either update the docstring with the verified optimum for this implementation (with provenance) or document the formulation difference.

Nonblocking: closes-vs-refs — issue #75's checklist leaves GLOA and Enumeration unchecked; this PR fixes GLOA and excludes Enumeration pending Pyomo/pyomo#3953. "Closes #75" is defensible only if the exclusion is accepted as the resolution; otherwise treat as Refs #75 and track re-adding the enumerate row once the upstream fix lands.

Sibling note: PR #145 also touches tests/test_benchmark.py; the hunks are disjoint and merge-tree reports no conflict.

Summary — Blocking: 0. Nonblocking: 3. Questions: 1. Tests: independent numeric re-derivation of all 8 bound constants (all match to 4 decimals); CI is green at this head (6/6 checks). Merge-ready once the optimum discrepancy and issue-linkage are resolved. Posting as COMMENT (author account); the main ruleset requires one approving review from another maintainer.

[bernalde]

Addressed the review of head b36cb3f in commit a959a54:

• val-bounds-assumptions: _val_bounds now documents the identity-R and nonnegative-data assumptions; the loop-invariant bound is hoisted out of the compound loop and renamed max_modeled_signal.
• vacuous-logical-assert-147: kept as an explicitly commented forward guard.
• docstring-optimum-mismatch (question) — resolved with a decisive local run: BARON on the Big-M reformulation of this head proves the global optimum is 12.089262 (LB = UB, optcr=1e-6, 120 s limit), so the module docstring's 12.0893 is correct and stays. The PR body's "GLOA optimal LB=UB=15.5406" is therefore not a global optimum: that run used --solver-profile gams-local, whose local NLP subsolvers void GLOA's global-bound guarantee, so GLOA's "optimal" label there certifies only its own converged bounds. The new variable bounds are unaffected — the 12.0893 optimum was found with them in place, confirming they cut nothing. Follow-up worth tracking on Refactor spectralog to fix benchmark issues #75: re-run the spectralog GLOA row with a global-subsolver profile (gams-baron) before checking the GLOA box.
• closes-vs-refs-75: recorded here since the body stands — with GLOA's evidence now known to be non-global and Enumeration excluded pending GDPopt enumerate materializes all discrete solutions before checking time_limit Pyomo/pyomo#3953, the link should be read as Refs #75; if merged as-is, reopen Refactor spectralog to fix benchmark issues #75 for the GLOA re-run and the enumerate-row restoration.

Tests: pixi run pytest tests/test_spectralog.py tests/test_benchmark.py — 24 passed; black applied and clean. CI for the new head was starting when this was posted; the previous head was green (6/6).