test: consolidate scripted MockClient doubles into a shared conftest fixture

[SuperMarioYL]

Summary

The runner, slot-worker, and eval-budget unit tests each ship their own
hand-rolled mock LLMClient. The three started as copies and have since
quietly diverged — they now disagree on observable behavior, so picking
"a" MockClient no longer gives a predictable contract.

Behavior	test_runner.py	test_slot_worker.py	test_eval_budget.py
exhausted responses	raises IndexError	raises IndexError	returns TextResponse("stuck")
send_stream	TEXT_DELTA + FINAL	raises NotImplementedError	FINAL only
call-spy lists	yes	no	no
api_format attr	absent	absent	"ollama"

The exhaustion split is the riskiest: a test moved between files (or a new test
copying the "wrong" double) silently changes from a hard error to a soft
fallback, which can mask a real defect.

Change

Replaces the three local classes with one configurable MockClient in
tests/conftest.py (previously empty). It is a behavioral superset, with
the variations promoted to explicit constructor knobs:

• on_exhausted — "raise" (default) or a fallback response object
• stream_mode — "deltas" (default) / "final" / "unsupported"
• api_format, context_length — defaulted to the common case
• call-spy lists (send_calls / send_stream_calls) are always populated

Each test file imports the shared class and passes only the kwargs needed to
reproduce its original behavior. No test assertion or test body was changed —
only the mock's source moved. The single-use inline NoFinalClient in
test_runner.py is intentionally left in place; it is a one-off for one
negative-path test, not a duplicate.

Why not just keep three mocks

A single shared double is only safe if it does not erase the per-test behavior
the three variants encoded — so the behavior is made configurable, not
collapsed. Each call site keeps its exact prior semantics, and the contract is
now visible at the construction site instead of hidden in three drifting class
bodies.

Test plan

• pytest tests/unit/test_runner.py tests/unit/test_slot_worker.py tests/unit/test_eval_budget.py — 105 passed, unchanged from before.
• pytest tests/unit/ — full suite green.

[antoinezambelli]

I appreciate this kind of cleanup! Doing some triage with in-flight stuff (including me, I'm literally traveling) but will circle back to this once some of the current work merges in.

[antoinezambelli]

Hey @SuperMarioYL — quick status, not a request for changes: I haven't forgotten this one. I took a closer look and the direction is right — consolidating the scripted MockClient doubles into a shared fixture, while keeping the runner / slot-worker / eval-budget behavior differences as explicit knobs rather than flattening them, is exactly how I'd want this done.

The reason I'm holding rather than merging: I've got a round of proxy/backend hardening still landing, and it's going to keep moving the test-suite surface for a bit. I'd rather not have you chase main through that churn and rebase repeatedly. Once the hardening settles, I'll come back to this and we can get it in against a stable base.

Appreciate the patience, and thanks for tackling the test-suite cleanup — it's genuinely useful work.

[SuperMarioYL]

Thanks for the detailed status — completely understood, and no rush at all. Consolidating the scripted MockClient doubles into a shared fixture while keeping the runner / slot-worker / eval-budget behavior differences as explicit knobs was exactly the intent, so I'm glad that lines up with how you'd want it. I'll keep an eye on main and am happy to rebase against a stable base whenever the proxy/backend hardening settles — just ping me when you're ready. Appreciate you taking the time to look.