release: v0.3.3 add SPECTER2 + TEI support

[rkrug]

Summary

• Adds backend_specter2_tei() — thin wrapper around backend_config() for a local TEI server serving the merged SPECTER2 proximity model.
• Adds one-time setup material under inst/scripts/:
  • prepare_specter2_merged.py — Python merge of the proximity adapter into specter2_base into a HF-format model dir TEI can serve directly. Default output under R_user_dir() cache, overridable via OVC_SPECTER2_PATH.
  • start_tei_specter2.sh — launcher that follows the same path convention.
• Adds vignettes/specter2-setup.qmd — end-to-end walkthrough (Python env, merge, TEI start, R smoke test).
• Bumps version to 0.3.3.
• NOTES.md (Rbuildignored) records the follow-up option to publish the merged model to HuggingFace Hub so users can skip the merge step.
• CLAUDE.md is a Claude Code guide for the repo (init).

Rationale

For 4M+ academic-paper embedding, SPECTER2 is domain-tuned (citation-linked papers), produces 768-dim vectors, and runs locally on TEI at zero per-token cost. The proximity adapter is required for the topic-similarity objective, but TEI does not load adapter-transformers adapters — hence the one-time merge step.

Model preparation deliberately stays out of the R API surface to avoid a Python dependency on this R-first package (see DEVELOPMENT_CONTINUITY.md decision log entry).

Test plan

• Run inst/scripts/prepare_specter2_merged.py in a fresh venv with transformers + adapters + torch — produces a model dir with config.json and weights.
• Start TEI via inst/scripts/start_tei_specter2.sh and confirm /info reports the local model path and /embed returns 768-dim vectors.
• From R: backend_specter2_tei() returns the expected backend_config and backend_embed_texts() produces a 1 x 768 matrix.
• devtools::check(args = c("--no-manual","--no-multiarch","--no-examples","--ignore-vignettes")) passes locally.
• Render vignettes/specter2-setup.qmd locally.

🤖 Generated with Claude Code