PipelineTestModel cannot be constructed: pipeline_integration_infrastructure.py written against an API that does not exist

[jeremymanning]

Symptom

Every test in tests/integration/test_pipeline_integration_infrastructure.py fails — 27 at HEAD, all with a TypeError from constructing PipelineTestModel.

Root cause

src/orchestrator/testing/pipeline_integration_infrastructure.py passes keyword arguments that the core dataclasses do not define. Fixing one simply reveals the next; I peeled four before stopping:

Line	Passes	Actual field
89	ModelCapabilities(max_output_tokens=...)	max_tokens
96–97	ModelRequirements(gpu_memory_gb=0, network_access=True)	network_access does not exist; requires_gpu expresses the intent
102	ModelMetrics(tokens_per_second=...)	not a field
112–114	ModelCost(input_cost_per_token=..., fixed_cost_per_request=...)	input_cost_per_1k_tokens, base_cost_per_request

max_output_tokens is Google's generation-config key (used correctly in google_model.py and langchain_adapter.py) — it was evidently copied into a ModelCapabilities call where it means nothing.

This module has therefore never run successfully. It is not a regression; it appears to have been written against an imagined API and never executed.

How it stayed invisible

Two methods in TestPipelineTestProvider shared the name test_pipeline_test_provider_initialization (lines 238 and 335), so the second silently replaced the first and one test never ran at all. Ruff's F811 catches this, but CI lints only src/orchestrator, not tests/.

Fixed in #434 by renaming the second method — which is what surfaced the TypeError chain above.

What this needs

A decision before any code: is this module wanted?

• It is "testing infrastructure" living in src/, and it is built around mock_responses — which the contributing guide forbids outright ("No mock objects. Ever.").
• It overlaps tests/test_infrastructure.py, which does work.
• It may belong to the frozen implementations in [EPIC] Phase 5: retire the frozen competing implementations (321 modules) #430.

If it is wanted, the four constructor calls need correcting and the mock-response design reconsidering. If not, delete it and its test file rather than leaving 27 permanently red tests that suggest coverage that does not exist.

I deliberately did not half-fix it: correcting two kwargs left it still 100% broken while making the diff look like progress.

Reproduce

ORCHESTRATOR_RUN_INTEGRATION=1 pytest tests/integration/test_pipeline_integration_infrastructure.py -q
# 27 failed

Suggestion beyond this issue

Extend the CI correctness lint to tests/ — F811 there means a test that silently does not run, which is worse than a failing one. There are currently ~73 correctness findings under tests/.