Merge upstream main, and read the GOC3 case through PowerIO 0.9

[samtalki]

Targets agent/powerio-parser-backend, the head of upstream PR madsuite-org#50. Merging this updates that PR in place, the way fork PRs #1 and #2 did.

Do not merge yet. It needs PowerIO = "0.9", and PowerIO.jl 0.9.0 is not released. See "Before merging" at the end.

1. Catch up with upstream main

Upstream added four commits since this PR's merge base. The blocker was madsuite-org#55, which moved every model builder to the ExaModels macro API. Four files conflicted. Everything else merged on its own, and the merged tree has no remaining calls to the retired variable / constraint / objective / constraint! functions, because this PR changes parsers and not model builders.

File	Resolution
.github/workflows/ci.yml	Upstream's cuda runner label and per-backend matrix. Keeps the job-level eager registry preference. One job key covers all four legs.
src/ExaModelsPower.jl	Upstream's macro imports, and drop using DelimitedFiles.
src/scopf.jl	Keep destructuring lengths by name.
test/opf_tests.jl	Keep the new parser-boundary assertions and upstream's callback smoke check. Drop the solve upstream removed.

Two of those need explanation.

using DelimitedFiles is the one line a plain merge gets wrong. Upstream kept the import. This PR removed the dependency along with the last readdlm call. Keeping both fails precompilation.

lengths must be destructured by name. Upstream used positional destructuring. lengths is a NamedTuple that sc_parser merges more names onto, and PowerIO 0.9 adds K to it, so positional binding would silently bind the wrong values.

Two more things the merge left incomplete:

• test/powerio_parser_tests.jl ran its testset at include time, so after the CI split it would run in all four jobs. It is now a function that runtests calls from the serial slice only.
• The README user_callback example still used the old API and returned (vars, cons). Under ExaModels 0.11 ExaCore is immutable and the macros rebind it, so a callback must return (core, vars, cons) or it drops everything it added. Upstream fixed runtests.jl and missed the README.

2. Read the GOC3 case through PowerIO 0.9

PowerIO.jl PR eigenergy/PowerIO.jl#86 adds the fields this needs. sc_parser no longer reads parse_goc3_json's raw dictionaries for case data.

• parse_sc_data drops its data_json argument. K comes from lengths, which also removes five inline recomputations of the same expression.
• The four reactive capability arrays read p.beta_ub and friends instead of Float64(data.sdd_lookup[p.uid]["beta_ub"]). PowerIO types those fields, so a typed comprehension keeps a concrete eltype when the filter selects nothing. The four hand-written empty sentinels and the any(...) ? ... : empty ternaries are gone.
• producers_first comes from PowerIO. This derived it from the lexicographically smallest uid, which matches the uid number only when uids are zero padded.
• The violation costs come from scd.violation_cost.
• The shunt rows use PowerIO's j_sh instead of re-parsing the uid string.

What still reads data directly: the two goc3_add_status_flags! arguments, and the startup/shutdown capability builders. Those read fields SddRow already carries, so they need no PowerIO API and can be refactored after 1.0 freezes the surface.

3. Solve one DCOPF case

The upstream merge left DCOPF with no solve anywhere, so nothing checked that dcopf_model reaches the right objective, and test_dcopf_case sat in opf_tests.jl with no call sites at all. Static AC already carves out one solve for the smallest case and leaves every other case on callbacks. DCOPF now gets the same carve-out, calling test_dcopf_case on case3 against solve_opf with DCPPowerModel, with the reference cached per case in dcopf_ref_cache the way static_ref_cache already caches the AC references. The two model builds move out of the callbacks testset so the Float64 variables stay in scope for the solve, matching how the static AC block is arranged.

This adds 8 assertions on the slices that solve: status, objective, and case3's three generator and three branch flows.

Verification

goc3_model on C3E4N00073D1_scenario_303 gives the same summary before and after: lengths, nvar / ncon, objective and constraint values at x0, and the type and shape of all 60 sc_data_array fields. It was captured three times — before the merge, after the merge, and after the PowerIO work.

One field set changes on purpose. c_p, c_q, c_s and c_e were Vector{Int64}, because JSON3 reads an integer-valued number as Int64 and this wrapped the raw dictionary value untyped. They are Float64 now with the same values. On a GPU backend the old type mixed Int64 into Float32 arithmetic.

Local test slices, all green:

• EMP_TEST_SELECTION=goc3 — 750 tests
• EMP_TEST_SELECTION=nothing — 2591 tests, 2583 of them before the DCOPF solve was added
• EMP_TEST_SELECTION=cpu — 450 tests, unchanged, because CPU() is not a solving backend

CI also covers cuda.

Before merging

Project.toml sets PowerIO = "0.9". Merge this only after PowerIO.jl 0.9.0 registers in General. Until then the test environment cannot resolve.

The previous bound was PowerIO = "0.7.3", which resolves to >=0.7.3, <0.8.0 and so excluded every released 0.8.x. The branch could not install PowerIO at all.

🤖 Generated with Claude Code