docs: ADR for embedding profile system and model decisions

[aaronsb]

Context

Over PRs #320 (profile abstraction), #322 (provider reporting + SigLIP removal), and #323 (task prefix support), we made several architectural decisions about embedding that should be captured in a formal ADR:

Decisions to document

1. Profile-based embedding config (migration 055) — text + image model slots stored in DB, hot-reloadable, with protection flags and vector space tags

2. SigLIP removal — evaluated as a dead end due to lack of shared text/image embedding space with our text model. Removed seed profile.

3. Jina CLIP v2 rejection — CC BY-NC 4.0 license makes it unsuitable for unrestricted commercial use

4. Nomic v1.5 as primary — best available option balancing quality, license (Apache 2.0), shared text/image space, 8K token context, and sentence-transformers compatibility

5. Purpose-aware task prefixes (migration 056) — search_query: / search_document: prefixes stored per-profile, applied via prompt_name for sentence-transformers or raw prepend for transformers loader

6. Qwen3-VL-Embedding-2B noted as future candidate — Apache 2.0, 2048-dim, 32K context, but requires vLLM/custom loader infrastructure we don't have yet

References

• ADR-039 (original embedding config)
• ADR-057 (Nomic vision research)
• Issue bug: Image ingestion fails with NomicVisionModel AttributeError #313 (Nomic vision attributeerror)
• PR feat: unified embedding profile with text + image model slots #320, Fix embedding provider reporting, remove SigLIP seed #322, feat: purpose-aware embedding with task prefixes #323