feat: generic ONNX embedding backend + multilingual-e5-small preset

[isolovyev77]

What does this PR do?

Local embedding models are currently fixed choices — minilm and embeddinggemma are hardcoded in embedding.py, and anything else requires standing up a separate openai-compat inference server. This PR adds the missing piece requested in #1563 / #1261:

• embedding_model: "generic-onnx" — run any HuggingFace-hosted ONNX encoder, selected entirely from configuration: repo, ONNX file, subfolder, pooling (mean/cls), tokenizer truncation length, instruction prefixes, and the persisted EF identity name all come from embedding_onnx_* keys in config.json, each overridable via the matching MEMPALACE_EMBEDDING_ONNX_* env var. No source edits, no separate server.
• embedding_model: "e5-small" — a bundled preset for intfloat/multilingual-e5-small (384-dim, fp32, ~120 MB): a lighter multilingual alternative to embeddinggemma-300m (~1/4 the download), also relevant to Feature request: support configurable embedding model (e.g. embeddinggemma-300m for multilingual) #1663.

Why e5-small specifically

This exact class + preset has been running in production since July on a 200k-drawer, RU-heavy palace. Measured on that corpus:

model	MRR	R@1
minilm	0.536	—
e5-small, symmetric passage:	0.684	—
e5-small, symmetric query:	0.754	0.65

One measured detail worth flagging: canonical e5 usage is asymmetric (passage: for documents, query: for queries), but the embedding wrapper routes both sides through __call__, applying one prefix symmetrically. On a real corpus symmetric query: beats symmetric passage: (0.754 vs 0.684), so the preset pins query: on both paths. GenericONNXEmbedding keeps separate doc_prefix / query_prefix (and embed_documents / embed_query), so if call sites later split the two paths, the canonical asymmetric pair is one config change away.

Design notes

• Same lazy-loaded deps as EmbeddinggemmaONNX (huggingface_hub, tokenizers, numpy, onnxruntime) — no new requirements.
• Follows the EmbeddinggemmaONNX conventions: double-checked lazy load, session assigned last, _intra_op_session_options thread cap, ImportError with reinstall hint.
• EF identity: the default collection name for generic-onnx is derived from repo + ONNX file, so pointing the config at a different model trips the embedder-identity check instead of silently mixing vector spaces; embedding_onnx_ef_name overrides it for vector-affecting knobs the derived name cannot see (pooling, max_len, prefixes). The e5 preset persists as e5_small_384.
• Switching models on an existing palace still requires mempalace repair rebuild-index — docstrings repeat the caveat.

Closes #1563, closes #1261; addresses #1663.

How to test

uv run pytest tests/test_generic_onnx.py -v   # 21 offline tests, no network
uv run pytest tests/ --ignore=tests/benchmarks

tests/test_generic_onnx.py mirrors test_embeddinggemma.py: fully offline via mocked hf_hub_download / Tokenizer / InferenceSession (module opted into _REAL_EMBEDDING_TEST_MODULES in conftest). Covers pooling (masked mean vs cls), L2 normalization, padding invariance, per-path prefixes, token_type_ids feed detection, batching, identity-name stability, config resolution (env + file, incl. trailing-space prefix preservation), and factory dispatch/caching.

To try it live: set "embedding_model": "e5-small" in ~/.mempalace/config.json on a fresh palace (or run mempalace repair rebuild-index on an existing one).

Full suite locally (macOS, py3.12): 4320 passed; the one failure is test_mcp_server::TestWriteTools::test_checkpoint_added_by_accepted_via_dispatch, which fails identically on clean develop.

Checklist

• Tests pass (python -m pytest tests/ -v)
• No hardcoded paths
• Linter passes (ruff check .)