perf(tests): four guard modules cost 64s at import time, and every changeset touching repo_tests/_reach.py pays all of it #16188

Four guard modules spend 64 seconds at IMPORT time, before a single test runs

Measured on origin/Dev_new_gui by importing each repo_tests/*_test.py and timing
the import alone — no test execution:

Importing the full guard set takes 67s. So 95% of it is these four, and the
work is at module scope rather than inside a test.

One of them starts a live client during import:

llm_shared.mock_providers - INFO - LocalLLM initialized with Ollama at http://127.0.0.1:11434

Why this is a real cost and not a curiosity

repo_tests/reach_declarations_test.py imports every guard module to enumerate
declare() calls. Any change to repo_tests/_reach.py therefore pays all 67
seconds before running one assertion — against a 128-second pre-push budget.

That is the mechanism behind the timeouts on #15928: not that reach discovery is
expensive (measured at 0.02–0.04s per declaration), but that importing the
guards to find the declarations costs 67s.

The shape

Import time is a cost every consumer pays, whether or not it uses the expensive
thing. A test module that builds its fixture at module scope charges that to:

• the pre-push hook, for any changeset that selects it
• every other module that imports it
• collection itself, even for a run that deselects all its tests

no_duplicate_dict_keys_test at 25s is paid by anyone who touches anything it
transitively reaches.

Acceptance criteria

• Each of the four is measured to find what it does at import — a fixture build, a corpus parse, a client construction — with the specific line named
• Expensive module-scope work moves into a fixture or a lazily-evaluated helper, so it is paid by the tests that need it
• No guard module constructs a network client at import; the Ollama line above is the instance to start from
• Re-measured after: full guard-set import time quoted before and after
• A guard caps import time per test module, so the next one is caught when it lands rather than when it blocks a push

Not the fix

Raising the pre-push budget. #16182 is a real and separate defect — the budget
scales with changed-file count rather than with the selected test set — but a
64-second import is a cost worth removing regardless of what the budget is.

Related: #16182 (the budget) · #15928 (blocked by this) · #16183 (blocked by #15928)