test(flash): isolate the summarizer's model factories from provider credentials

[basil-k-aji-dev]

Closes #41.

Problem

Four Flash test modules cannot run without Gemini credentials, contrary to the deterministic-suite contract in CONTRIBUTING.md. On this host (Linux, GOOGLE_API_KEY / GEMINI_API_KEY / OPENAI_API_KEY all unset) they account for 64 of the 89 failures in make test.

VisualStepSummarizer.__init__ resolves its utility model eagerly:

try:
    if model_name:
        self._llm = get_google_llm(model_name=target_model, temperature=0.0)
    else:
        self._llm = get_llm(ctx, name="summarizer", is_utils=True)
except Exception:
    self._llm = get_google_llm(model_name=target_model, temperature=0.0)

Both branches — including the except fallback, which catches any failure of the configured path — end in a real ChatGoogleGenerativeAI construction that raises API key required for Gemini Developer API. The tests assign their fake to summarizer._llm after construction, which is too late: the object never gets built. test_flash_runner* and test_flash_scrub_edge reach the same constructor indirectly through FlashRunner and ScrubEdgeCompressor.

Separately, test_flash_config_and_builder only asserts Flash configuration values but ran the builder's provider validation, failing with Planner requires GOOGLE_API_KEY in .env.

Change

tests/unit/agents/conftest.py (new) — a stub_summarizer_model_factories fixture replacing both factories at the summarizer module boundary, and a small StubChatModel that:

• exposes model, the one attribute the constructor reads back, so the model-name resolution branch stays exercised rather than silently skipped;
• raises from ainvoke, so a test that needs generation and forgets to install its own fake fails loudly instead of quietly reaching for the network.

Each of the four modules opts in with a one-line autouse fixture. The real summarizer, its bounded retry loop, the compressor, the ledger and the runner all stay under test — this replaces only the two credential-resolving factories.

test_flash_config_and_builder now calls .build(validate_profiles=False), the option the builder already exposes for configuration-only use.

No production code is touched and no credential validation is relaxed.

Why the module boundary rather than an autouse fixture for tests/unit/agents

A directory-wide autouse fixture would also cover test_explorer.py and test_video_analyzer.py, whose failures are a different bug (#33, MagicMock provider values). Keeping the opt-in explicit matches the issue's scope and keeps the two fixes independently reviewable.

Verification

Targeted, exactly the reproduction in the issue:

$ uv run pytest -q tests/unit/agents/test_flash_runner.py \
                  tests/unit/agents/test_flash_runner_ledger.py \
                  tests/unit/agents/test_flash_scrub_edge.py \
                  tests/unit/agents/test_flash_step_summarizer.py
77 passed in 1.30s

matching the 77 passed the issue predicts (from 64 failed, 13 passed).

Full deterministic suite on this host:

	Result
main	89 failed, 2096 passed, 10 skipped, 8 deselected
this branch	26 failed, 2159 passed, 10 skipped, 8 deselected

Exact failed-node comparison: 64 resolved, 0 new failures. The 26 that remain are the groups the issue already attributes elsewhere — #31 (turn-index snapshot, 4), #33 (explorer / video-analyzer provider values, 12), #18/#52 and adb-dependent tests.

Command	Result
uv run ruff check tests/unit/agents/	All checks passed
uv run ruff format --check tests/unit/agents/	41 files already formatted

make typecheck could not be run in my environment — the pyright wheel downloads a Node runtime on first use and my sandbox blocks it. No touched file is in pyright-core.json.

[basil-k-aji-dev]

Closing as a duplicate. #42 by @wellorbetter already covers issue #41 and was opened first — that one should get the review, not this.

My mistake: I didn't check the open PR queue for an existing claim before sending this. Sorry for the noise.