Restyle N-1 SCOPF on the recipe/args split, add DC, wire into AOT

[sshin23]

Summary

Builds on the N-1 SCOPF work by restyling it onto the package's recipe/args split, adding the DC formulation, wiring the security-constrained models into the ahead-of-time compilation workflow, and removing the duplicated per-model constraint-selector families that this work would otherwise have made a third copy of.

One file, built from the multi-period halves

src/scopf_simple.jl + src/scopf_twostage.jl become one src/scopf.jl, matching opf.jl / mpopf.jl / goc3.jl.

The monolithic SCOPF is the multi-period model with the time axis read as a scenario axis: naming the arrays barray/busarray/arcarray/genarray/refarray and the rep fields exactly as mpopf.jl names them lets its halves build the SCOPF body unchanged — the ramp becomes the corrective coupling and storage drops out. That removes the branching physics (if form == :polar … elseif :rect) in favour of the package's existing dispatch on OPFForm.

Both models are now *_recipe ∘ *_args, so the model solved here and the model ExaModelsC compiles come from one definition.

One selector family instead of three

Every formulation selector in opf.jl (c_ref, c_angle, c_bal_p, c_bal_q, ac_flow) now takes a trailing idx..., so one definition serves the static model (empty tail → V.va[i]), the multi-period model ((t,) → V.va[i, t]) and the SCOPF ((c,)). The 18-method _mp family in mpopf.jl is deleted; it differed from the static one only by the index tail. The splat resolves at compile time — the tail's length is in the type — so the traced expression is identical to the hand-written index.

DC formulation

scopf_model now takes DC() alongside Polar() and Rect().

A line outage is a mask, and the two families mask different fields. AC zeroes the admittance coefficients c1..c8; DC cannot, because its flow is built from br_r/br_x and zeroing br_x is 0/0 on the zero-resistance branches real cases contain (3 of case9's 9, 9 of case118's 186). So c_ohms_law_dcopf now takes the susceptance as an argument, dc_susceptance(br) computes it as data, and it rides on each branch row — zeroed for the tripped branch in its own scenario.

Because the outage becomes a pure data fact, add_flow_cons_mp!(::DC) is reused by the SCOPF unchanged: there is no SCOPF-specific DC code path.

scopf_twostage_model stays AC-only.

AOT

compile_all(ExaModelsPower) gains :scacp, :scacr and :scdcp (nine models). K is derived from SCOPF_DEFAULT_CONTINGENCIES rather than exposed as a compile_all keyword: the compiled library calls the argument function with a path and nothing else, so the recipe and the arguments cannot be allowed to disagree on K.

compile_all was red before this branch as well, on main and with no SCOPF sources present: four juliac --trim=safe verifier errors that all bottom out in one line of a dependency. ExaPowerIO's matpower parser built bus_map from an untyped [] and narrowed it by its Vector{Int} annotation, and the resulting convert(Vector{Int64}, ::Vector{Any}) copies element by element through a dynamic setindex! that the verifier will not resolve.

That fix is now released as ExaPowerIO v0.3.1, so this branch depends on the registered version rather than pinning a commit. [compat] moves from "0.3" to "0.3.1": "0.3" still admits v0.3.0, the exact version that cannot be compiled, so the floor is a build requirement rather than a preference and belongs in the resolver.

That break went unnoticed because test_aot() has never run in CI: it sat behind an EMP_TEST_AOT variable that no job in ci.yml sets. It is now the aot entry in the test matrix, beside goc3 — a slice cannot be forgotten the way an opt-in variable can, and it is the cheapest slice in the set. The variable is gone. test_aot() also no longer needs CNLPMODELS_PATH: compile_all defaults to path = "@emp", which resolves against that variable, so the test now passes a directory of its own and a plain Pkg.test() runs it with no setup.

Test wiring

• scopf_tests() sat outside both the EMP_TEST_SELECTION guard and the backend loop, so all four CI jobs ran it. It is now sliced like everything else.

• The SCOPF recipe/eager equivalence is checked for all three formulations. DC is not in static_forms (those entries carry a PowerModels type and a voltage test), so it gets its own loop.

• A DC outage regression checks the mechanism rather than the cost, and checks it on the formulation rather than on a solve. The DC model is exactly linear, so "the outaged line carries no flow" is a property of the constraint matrix: it holds exactly when pf[l, c] is pinned to a constant by the equalities alone and that constant is zero. That is a row-space membership test — exact rather than tolerance-bound, and true at every feasible point rather than at one. It is armed on both sides: nothing may pin the same line in the base case, and nothing may pin a different line in the same scenario, so neither a formulation that pinned every flow nor a dead scenario would pass. The arming doubles as a check on the index arithmetic, which is the one place this reaches into ExaModels' variable layout.

  It previously solved and read the flow out of the solution, which made a formulation check depend on the fragile part. Whether a particular SCOPF instance is solvable is a separate question from whether it is formulated correctly, and only the second is this package's responsibility — case9 DC at tol = 1e-8 is a case where the two answers differ. MadNLP reaches a primal-feasible point with the right objective (constraint violation 2.7e-15, objective agreeing to nine significant figures with a converged run) and then terminates INFEASIBLE_PROBLEM_DETECTED: the active constraint Jacobian is rank deficient by exactly K, because summing the bus balance rows over one scenario cancels every branch flow, so the difference of two scenarios closes against the corrective rows. The last mu level takes ten inertia corrections before entering restoration. The replacement runs in 35s where the solve took three minutes.

Test plan

Run locally on Julia 1.12.6, two NVIDIA GPUs (RTX 3080, RTX 2080 Ti).

• EMP_TEST_SELECTION=nothing — 746 passed, 0 failed, 0 errored, 13m26.9s.
  The ten that would previously have made this 756 are the aot slice
  below; the compile no longer runs in this one. Wall times here are off a
  shared desktop with unmeasured concurrent load, so they are not comparable
  run to run.
• EMP_TEST_SELECTION=aot — 10 passed, 0 failed, 2m18s, with
  CNLPMODELS_PATH unset, so the environment is the one CI will have. Ten is
  exactly test_aot()'s assertion count, so the slice selects that test and
  nothing else. Re-run after the pin came out, resolving ExaPowerIO v0.3.1
  from the registry rather than from a git checkout — 10 passed, 2m10s.
  Before the fix was available it errored inside compile_all before any of
  the ten ran. Both timings are against a warm local depot; a fresh runner
  pays precompilation on top.
• EMP_TEST_SELECTION=cpu — 450 passed, 0 failed
• EMP_TEST_SELECTION=cuda — 327 passed, 0 failed
• GPU SCOPF two-stage, in isolation — 5 passed, 0 failed, 2m54s. Run
  separately because a clean @testset prints no per-test breakdown, so the
  slice being green is not by itself evidence that this path executed.
• DC outage assertion, mutation tested on the CI machine under Julia 1.12.7
  (the combination that produced the original failure) — 11/11 green, and
  each half killed by the mutant aimed at it. Leaving the susceptance at its
  true value under outage fails exactly the two outage assertions (residual
  0.090 and 0.095, pinned value 0.382 and 0.581) and leaves both arming
  assertions green; outaging every branch in every scenario fails both
  arming assertions (1.2e-16 … 4.4e-16 against a > 1e-3 bar) and the rank
  precondition. Row-space separation holds on case9, case14_ieee and
  case5_pjm: 5.6e-17 … 2.8e-16 for the outaged pair against 0.058 … 0.25
  for both controls.
• compile_all end-to-end under juliac --trim=safe — verifier clean,
  libemp.so produced with all nine prefixes present, and the :scacp
  library model matches the eagerly built reference: same dimensions
  (81, 90), equal obj and cons at a point away from x0.

[codecov]

Codecov Report

❌ Patch coverage is 89.40678% with 25 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.53%. Comparing base (5f64c18) to head (ba19ee4).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/scopf.jl	88.29%	24 Missing ⚠️
src/constraint.jl	66.66%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #59      +/-   ##
==========================================
+ Coverage   80.63%   83.53%   +2.89%     
==========================================
  Files           8        9       +1     
  Lines        1477     1652     +175     
==========================================
+ Hits         1191     1380     +189     
+ Misses        286      272      -14     

Flag	Coverage Δ
aot	21.73% <45.92%> (?)
cpu	21.34% <12.34%> (-3.93%)	⬇️
cuda	32.52% <89.69%> (+7.25%)	⬆️
goc3	51.93% <0.00%> (-6.41%)	⬇️
nothing	33.89% <90.12%> (+8.28%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.