Before

• Every semanticize call reloaded tokenizer/session and semanticize_batch just looped, so batches still ran N single-document inferences.
• Missing ONNX/tokenizer assets caused hard errors despite the docs claiming a deterministic stub fallback.

After

• Introduced a per-thread cache that stores tokenizer + ONNX session handles keyed by model path.
• ONNX mode now performs a single batched inference that pads inputs and reuses the cached session.
• Errors resolving models/tokenizers fall back to the deterministic stub, matching the documented behavior, and docs/tests were updated accordingly.

Advantages

• Huge reduction in latency and filesystem thrash for steady-state semantic inference.
• Batching finally amortizes session setup costs, improving throughput for multi-doc workloads.
• Pipelines keep running even when assets are missing or temporarily unavailable.

Testing

• cargo fmt --all
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace --all-targets -- --nocapture