Embedding engine hardcodes last-token pooling, producing wrong vectors for CLS/mean models (e.g. bge-m3)

[unverbraucht]

Problem

src/engine/optimum/optimum_emb.py always applies last-token pooling in generate_embeddings:

embeddings = self.last_token_pool(outputs.last_hidden_state, batch_dict["attention_mask"])

This is correct for decoder-style embedding models like Qwen3-Embedding-*, but wrong for the much larger family of encoder-style models that ship via sentence-transformers — most notably:

• BAAI/bge-m3, BAAI/bge-large-en-v1.5, etc. → CLS pooling
• sentence-transformers/all-MiniLM-L6-v2, intfloat/multilingual-e5-* → mean pooling

Loading any of these through OpenArc's optimum engine today returns numerically valid but semantically wrong vectors (last hidden state of the final non-pad token instead of [CLS] / masked mean). There's no error and no warning — retrieval quality just silently collapses.

Reproducer

1. Convert BAAI/bge-m3 to OpenVINO IR via optimum-intel (preserves the 1_Pooling/config.json shipped with the model).
2. Register it in openarc_config.json with engine: "optimum", model_type: "emb".
3. Hit /v1/embeddings and compare against the PyTorch reference — cosine similarity is ~0.3–0.6 instead of ~1.0.

Proposal

Make pooling mode metadata-driven, with the same precedence sentence-transformers itself uses:

1. runtime_config.pool_mode (operator override; one of "cls" | "mean" | "last"). Unknown values raise at load time so typos don't silently fall back.
2. <model_path>/1_Pooling/config.json (auto-detect from the file the model ships with — pooling_mode_cls_token → cls, pooling_mode_mean_tokens → mean).
3. Default: "last" (preserves current Qwen3-Embedding behavior; no change for existing users).

This keeps the registry config minimal for the common case (correct pooling auto-detected from model files) while giving a clear escape hatch for models that ship without sentence-transformers metadata or with the wrong metadata.

Scope

• src/engine/optimum/optimum_emb.py: add cls_pool / mean_pool, dispatch in pool(), resolve mode in load_model.
• Unit tests for each pool, _detect_pool_mode, the override path, and unknown-value rejection.
• Integration test loading bge-m3 and asserting cls auto-detect + 1024-dim unit-normed vector.

Happy to send a PR — branch is ready (feat/embedding-pool-dispatch on KIntegrated/OpenArc), verified end-to-end against PyTorch (cos > 0.999) and live via /v1/embeddings on GPU.