fix(tests): keep the offline gate hermetic for model-fallback tests

[Coding-Dev-Tools]

Problem

python -m pytest tests/ -q — the offline gate documented in CLAUDE.md and AGENTS.md §1 — hangs indefinitely on a host with no network route, rather than passing or failing.

Two tests assert that the backend factories fall back when a model cannot be loaded:

• test_embedder_factory_falls_back_offline
• test_reranker_factory_falls_back_offline

Both pass an unresolvable model name, which sends SentenceTransformer / CrossEncoder to the Hugging Face Hub to resolve it. When the host has no route to the Hub, the TCP connect sits in SynSent instead of erroring, so the suite never reaches a summary line.

The fallback relied on the network failing fast, not on it being absent — which is precisely the condition AGENTS.md §3 ("local-first and offline-capable") requires the core to survive.

Fix

Patch the sentence-transformers loader in both tests so the load failure is raised locally and instantly. No network is touched, and the fallback contract each test exists to verify is still exercised end-to-end through get_embedder / get_reranker.

If sentence_transformers is not installed at all, the helper no-ops — the factory already fails fast with ModuleNotFoundError, which is the path minimal CI (pip install numpy pytest) takes.

Test-only change. No production code is modified.

Verification

Full offline gate, run locally:

• python -m pytest tests/ -q — completes; previously hung forever
• python -m ruff check . — All checks passed
• python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5 — recall@k 1.000, hit@k 1.000, answer_token_recall 1.000
• python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5 — recall@k 1.000, hit@k 1.000, answer_token_recall 1.000
• python -m eval.ablation — vector-only 1.0, hybrid-1hop 1.0, hybrid-ppr 1.0; multi-hop arm-level: vector 0.667, graph 1-hop 0.0, graph PPR 1.0

The two target tests now complete in well under a second with --timeout=30.

Note on an unrelated pre-existing failure

tests/test_session_idempotent.py::test_session_close_linearizes_before_delayed_memory_write[ingest] fails locally when a developer .env sets ENGRAPHIS_EXTRACTOR=llm with an unreachable provider: the LLM client retries 2s + 4s and exceeds the test's 10s budget. It reproduces standalone on main without this branch, and CI has no .env, so it is green there. Not addressed here — it is the same hermeticity gap in a different place and deserves its own change (neutralising LLM env vars in tests/conftest.py).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 93e7e53ed0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Coding-Dev-Tools]

Good catch, and fixed in 06f09b0.

The helper no longer imports sentence_transformers at all. It now patches the adapter each factory constructs — SentenceTransformerEmbedder in engraphis.backends.embedder_st and CrossEncoderReranker in engraphis.backends.reranker — so a package that imports but raises something other than ImportError (a RuntimeError from a mismatched torch, for example) can no longer fail these tests. Every such failure now stays inside the factory under test, where the existing except Exception catches it and falls back, which is the contract these tests exist to verify.

Verified the optional stack is never imported: after running this module's tests, 'sentence_transformers' in sys.modules is False.

Full offline gate green locally (pytest, ruff, both eval harnesses, ablation), and all 9 required checks pass on CI including Python 3.10/3.11/3.12 and the numpy-only 3.9 floor.