Iterate DeviceModel's device_cache, and narrow the device-collection signatures to Vector

[luke-kiernan]

Closes the POM half of IOM #79. IOM #157 was the prep; this is the "real meat" that PR's description promised.

1. Pass the cache instead of rebuilding the iterator

DeviceModel caches its components in a concretely-typed device_cache::Vector{D}, filled by make_device_cache! during template validation. The constructors still called get_available_components(model, sys) on every construct stage, rebuilding a FlattenIteratorWrapper — a wrapper over Iterators.Flatten of a generator of dict values, whose nested-tuple iterate state inference struggles with, and which offers no getindex, IndexStyle or HasShape. ~200 call sites now read get_device_cache(model).

Clones that never see template validation

Three sites build a DeviceModel outside validate_template!, so its cache is empty and construction silently proceeds against zero devices. Each now carries one over explicitly:

• initialization.jl — the IC template is assembled from get_initial_conditions_device_model clones and handed straight to build_problem!, bypassing validation entirely.
• load_constructor.jl — the controllable-load fallback rebuilds the model as StaticPowerLoad and recurses into construct_device!.
• test_utils/mock_operation_models.jl — the mock harness constructs devices with no template at all.

validate_available_devices deliberately keeps get_available_components: it runs once per DeviceModel rather than per device, so it gains nothing from the cache, and it is called directly on freshly built models that have none.

2. Narrow the signatures to Vector

With the cache being passed, Union{Vector{T}, IS.FlattenIteratorWrapper{T}} no longer earns its width. All of them narrow to Vector{T}, leaving one convention instead of two — including the 155 that already used the Union before this PR.

This is not a performance change. Julia specializes on the concrete argument type at each call site regardless of annotation width, so the Union compiled exactly as well. The speed came from passing the cache. What narrowing buys is a tighter contract and a single spelling.

_make_device_cache keeps IS.FlattenIteratorWrapper{T} — it is the boundary where the iterator becomes a Vector, so it is the one place the type belongs.

Five sources still produced an iterator and now materialize at the source: the make_system_expressions.jl areas; the AGC service constructors; get_available_reservoirs (a bespoke helper in utils/psy_utils.jl feeding 12 hydro call sites, invisible to a grep for get_available_components or PSY.get_components); and two sets of tests that call builders directly with PSY.get_components(...).

Two pre-existing devices_vec = collect(devices) calls in the hybrid models were written when devices was an iterator. They are redundant copies now that it arrives as the cache vector, and both uses are read-only, so they alias instead. No collect in src/ sits inside a loop.

One hazard worth flagging for review

Because IOM's methods still accept the Union, a stray iterator does not fail at the boundary. It falls through to a more generic method and surfaces far away — as get_fixed(::MarketBidCost) deep in the cost path, or as POM's own add_parameters! not implemented for ... Implement this method in PowerOperationsModels fallback, which reads as a missing feature rather than a dispatch miss. Narrowing IOM to match would restore a loud boundary. Reviewers may reasonably prefer to split section 2 into a coordinated POM+IOM change.

Testing

Full-suite baseline versus this branch, in the same environment: all 54 test files report identical Fail/Error counts, verified separately after section 1 and after section 2. test_aqua passes, so no new method ambiguities.

The 16 files that are red are red on main too, from the pre-existing ramp_limits requires a time on the units argument error that #304 fixes. CI will show them until that lands.

Follow-up, not done here

The 12 hydro reservoir/turbine call sites go through get_available_reservoirs(sys) rather than get_device_cache(model), so IOM #79's goal is not yet met for hydro — they still materialize per construct stage. Worth its own issue.

Also here

A separate commit repoints test/Project.toml's IOM pin from rh/cost_coefficient_ratio to main. That branch was IOM #166, which merged and was deleted, so the test environment did not resolve at all on main:

ERROR: Did not find rev rh/cost_coefficient_ratio in repository

🤖 Generated with Claude Code

[jd-lara]

Can we just make all the inputs vector{} I think we can remove iterator from the functions signatures

[luke-kiernan]

Can we just make all the inputs vector{} I think we can remove iterator from the functions signatures

Claude code says it's doable, requires adding collect in a dozen spots but that's it. I will change to that

[jd-lara]

Can we just make all the inputs vector{} I think we can remove iterator from the functions signatures

Claude code says it's doable, requires adding collect in a dozen spots but that's it. I will change to that

We should check the places where the collect is needed and we probably don't really do need it, it could be missing caches

[github-actions]

Performance Results

Version	Precompile Time
Main	5.245181516
This Branch	5.247914566

Version	Build Time
Main-Build Time Precompile	86.67194688
Main-Build Time Postcompile	1.054137022
This Branch-Build Time Precompile	86.364210345
This Branch-Build Time Postcompile	1.08109076

Version	Solve Time
Main-Solve Time Precompile	139.033501019
Main-Solve Time Postcompile	105.848773976
This Branch-Solve Time Precompile	310.774662814
This Branch-Solve Time Postcompile	264.874514182

[codecov]

Codecov Report

❌ Patch coverage is 94.17989% with 11 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/ac_transmission_models/branch_constructor.jl	87.32%	9 Missing ⚠️
...tic_injector_models/hydrogeneration_constructor.jl	92.30%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[jd-lara]

I think that test failures are unrelated but @luke-kiernan can you check them and fix the testing please?

[luke-kiernan]

can you check them and fix the testing please?

Should be fixed now