test(backends): cover the _embed_texts fallback and empty-batch guard

[mbeacom]

Refs #2190. Follow-up to #2187 — tests only, no production code changes.

Note on scope. I independently diagnosed and fixed #2190, then found #2187 had already landed the same one-line .tolist() fix while I was verifying. Rather than open a duplicate, I rebased onto develop and kept only the part that isn't already covered. mempalace/backends/embedding_wrapper.py is byte-identical to develop on this branch.

What does this PR do?

#2187 fixed the default-backend ingest failure and covered the numpy path well — an ndarray-returning EF plus a real EmbeddingCollection.upsert round-trip. Two branches of the fixed function are still unexercised:

1. The float(x) fallback arm. _embed_texts branches on hasattr(v, "tolist"):

return [
    v.tolist() if hasattr(v, "tolist") else [float(x) for x in v]  # numpy | plain sequence
    for v in vectors
]

Only the numpy side runs under test. The fallback is the arm serving embedders that return plain sequences — custom/BYO EFs, and rows that arrive as tuples. I confirmed the gap by mutation: replacing that arm with list(v) leaves the suite fully green. A regression there would surface only in the field, on a non-default embedder, as the same production ValueError #2190 describes.

2. The if not texts: return [] guard. Also untested. Callers pass empty batches (a drawer set fully filtered by dedup), and constructing the EF is the expensive part — on the ONNX default it spins up a native session. Deleting the early return is likewise green today.

Both tests are written so they assert the behaviour, not just the shape:

• the plain-sequence EF yields Decimal, so == [[0.5, 0.25], ...] plus type(x) is float proves a real conversion happened rather than values passing through unchanged;
• the empty-batch test asserts by making get_embedding_function raise, so it verifies the EF is never constructed, not merely that [] came back.

Neither test depends on the NumPy version. They live in tests/test_embedding.py for exactly the reason #2187 documented: conftest's autouse _stable_embedding_function_for_tests replaces _embed_texts outright for every module outside _REAL_EMBEDDING_TEST_MODULES, so tests placed elsewhere exercise the stub and pass against broken code.

How to test

Mutation-verified — each new test was proven to fail against a deliberately broken tree, then the tree was restored:

Mutation	Result
fallback arm → list(v)	❌ test_embed_texts_handles_plain_sequence_embedders fails (type(x) is float → False)
delete if not texts: return []	❌ test_embed_texts_short_circuits_on_empty_input fails (get_embedding_function must not be called for an empty batch)
unmodified develop	✅ both pass

The load-bearing point: deleting either branch of the shipped function leaves the suite entirely green today. That is the argument for these tests existing.

uv run pytest tests/ -q --ignore=tests/benchmarks   #  3854 passed, 31 skipped
uv run ruff check .                                 #  All checks passed!
uv run ruff format --check .                        #  212 files already formatted

git diff develop -- mempalace/ is empty — the change is confined to tests/test_embedding.py (+49).

Correcting the root-cause framing (this was mine, and it was wrong)

An earlier revision of this PR described #2190 as a "NumPy 2.x" failure, and I repeated that on the issue. That attribution is incorrect and I want it on the record rather than quietly edited away.

np.float32 has never been a float subclass in any NumPy major. Verified directly across the full supported range (pyproject.toml pins numpy>=1.24):

NumPy	isinstance(np.float32(0.5), float)	isinstance(np.float64(0.5), float)
1.24.4	False	True
1.26.4	False	True
2.4.4	False	True

float32 mro on 1.26.4: ['float32', 'floating', 'inexact', 'number', 'generic', 'object'] — no float.

So normalize_embeddings would have rejected list(arr) output on NumPy 1.x just the same; there was no version boundary. The accurate framing is the one #2187's own docstring used, which pointedly never claimed version-specificity:

A latent defect exposed by a capability change — 0e79797 added requires_explicit_embeddings, routing the default Chroma backend through _embed_texts for the first time — and missed because conftest stubs _embed_texts out of every test module outside _REAL_EMBEDDING_TEST_MODULES.

A corollary worth stating explicitly: a numpy<2 upper bound would not have prevented this, and shouldn't be added on this basis.

None of this affects the change here — both tests target version-independent branches and remain mutation-proven.

Independent confirmation that #2190 is resolved on develop

• Pre-fix: _embed_texts(['hello world']) → <class 'numpy.float32'>, and normalize_embeddings raises.
• On develop: → <class 'float'>, normalize_embeddings OK, dim 384.
• End-to-end, real mine (not --dry-run) of 3 Claude transcripts into a throwaway palace: pre-fix 0 drawers filed, aborting in _validate_and_prepare_upsert_request; on develop 86 drawers filed, and mempalace search returns the drawer verbatim.

Checklist

• Tests pass (python -m pytest tests/ -v) — 3854 passed, 31 skipped
• No hardcoded paths
• Linter passes (ruff check .) — plus ruff format --check . clean

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[igorls]

Wave 3 minimal for 3.7.0: LGTM. Docs/tests/tiny fix only.