feat: production-grade bring-your-own-model (BYO) support

[n24q02m]

Spec B Part A — make add_custom_model production-grade so any ONNX-able embedding model can be plugged in. Qwen3 stays the only built-in.

Bug fixes (verified file:line)

• dim/MRL swallowed on custom + pooled paths (CustomTextEmbedding, PooledEmbedding, PooledNormalizedEmbedding ignored **kwargs) — embed(..., dim=256) now truncates (pool → truncate → normalize, matching Qwen3).
• custom models broken under parallel>1 — registration lived in class vars (POSTPROCESSING_MAPPING/SUPPORTED_MODELS) that don't survive a spawned worker → KeyError. Folded pooling/normalization onto a new frozen CustomDenseModelDescription, keyed the registry case-insensitively, and propagated it to workers via a polymorphic _extra_worker_params() hook.
• case-sensitivity in the registry — fixed by the same refactor (lowercased keys).

Features

• CustomModelSpec — one-call BYO registration (CustomModelSpec(...).register()).
• qwen3_embed.export.export_to_onnx() — optional HF-id→ONNX helper (lazy imports torch/optimum; NOT a packaged extra because the export deps don't co-resolve with the lean tokenizers/huggingface-hub pins — documented manual install).
• NOTICE — retains fastembed (Qdrant) Apache-2.0 attribution (was missing).

Tests

• 496 unit passed (was 491), ruff/format/ty clean.
• Worker-registry propagation covered by test_custom_registry_survives_serialization. The end-to-end parallel=2 integration test is skipif(win32) (Windows multiprocessing+onnxruntime spawn deadlocks under pytest — runs on Linux).

Notes: custom INT8 models score per-row (batched INT8 isn't batch-invariant — Spec A Part 2 finding).

🤖 Generated with Claude Code

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None