TIG-254: Embedding-based event extraction pipeline (staging → main)

[DIodide]

Promotes staging to main. Contents: PR #34 (embedding-based event extraction pipeline by @yubimamiya, with review fixes).

Summary

• Replaces LLM event extraction with clean → embed → logistic-regression classify → cosine-similarity tagging
• New event_tag_embeddings table (pgvector, 1536-dim) + seeded tag embeddings; committed classifier model
• Renamed/expanded event_tag (22 values) and org_category enums, with a verified single-migration USING-cast conversion (0002_update_enums.sql) — replay-tested against a scratch DB seeded with every old enum value
• Frontend fully aligned with the new enum values; users.ts types now derived from the schema
• Pipeline safety: needs_review events (fallback date, no location) insert as is_public = false
• scikit-learn pinned to the model's pickle version (1.6.1)

Verification

• tsc --noEmit passes in apps/web and apps/database; biome clean; py_compile clean
• Migration chain 0000 → 0001 → 0002 replayed in one transaction on pgvector:pg17 with old-format data, including the workshop+academic merge collision and the four previously-unmapped org categories

Deploy note

Local/staging containers created from a plain postgres:* image must be recreated from the compose file (now pgvector/pgvector:pg17) before migrating — the data volume carries over.

[linear]

TIG-254