Accommodate ORM schema reorg for search/browse queries

[skearnes]

Summary

ord-schema's upcoming 0.8 release reorganizes the ORM into role-based schemas (open-reaction-database/ord-schema#859) and splits dataset loading into separate ingest and derivation stages (open-reaction-database/ord-schema#860, open-reaction-database/ord-schema#861):

• public.reactions — the served Reaction proto payload, keyed by reaction_id (moved off ord.reaction.proto).
• public.datasets — per-dataset metadata num_reactions / submitted_at (moved off ord.dataset).
• derived.* — generated SMILES + rdkit_*_id links: reaction_smiles, compound_smiles, product_compound_smiles (moved off ord.reaction / ord.compound / ord.product_compound).
• ord.* — now a pure search index (decomposed message tables only).
• Loading is now two stages — add_dataset / add_parquet_dataset ingest ord.* + public.*, and update_derived_data populates the derived.* SMILES / RDKit links / reaction classes.

This PR updates the search/browse queries to read from the new locations and updates the test-fixture loader to the two-stage API.

Changes

• ReactionSmartsQuery / ReactionComponentQuery (queries.py) — join through derived.reaction_smiles / derived.compound_smiles / derived.product_compound_smiles to reach the rdkit_*_id links, replacing the dropped ord.reaction.rdkit_reaction_id / compound.rdkit_mol_id columns.
• fetch_reactions — reads proto from public.reactions.
• _fetch_dataset_most_used_smiles — joins the derived SMILES tables for the smiles column (new smiles_table / smiles_foreign_key params).
• get_datasets / get_dataset (search.py) — join public.datasets for num_reactions / submitted_at.
• setup_test_postgres (testing.py) — the removed add_dataset(..., rdkit_cartridge=...) keyword is gone; ingest with add_dataset(dataset, session) then derive with update_derived_data(dataset.dataset_id, session, rdkit_cartridge=...), mirroring the ord-schema ORM conftest and preserving the rdkit-cartridge gating.
• queries_test.py — the _max_similarities helper mirrors _mols_join, updated to the same derived-table joins.
• pyproject.toml — bump the pin to ord-schema[orm]>=0.8,<0.9.

Reaction classes (ord-schema#854)

No changes needed: ord-interface does not read the reaction class/name fields anywhere (no query, search, or frontend reference), so #854 — and its move to derived.reaction_classes in #859 — is transparent here.

Correctness note

All joins to the derived.* tables are INNER, which preserves the prior NULL-column filtering: update_derived_tables inserts a derived row only when a SMILES can be generated, so un-derivable entities are dropped exactly as the old rdkit_*_id IS NULL / smiles IS NOT NULL paths did. The API response shapes (DatasetInfo, QueryResult) are unchanged, so no frontend changes are needed.

Blocked on release ⚠️

Draft until ord-schema 0.8.0 is published:

• uv.lock still pins 0.7.1 and must be regenerated (uv lock) once 0.8.0 is on PyPI.
• CI (and the local ty check pre-commit hook) cannot resolve >=0.8 until then; these commits were pushed skipping only the ty-check hook for that reason — ruff check, ruff format, and addlicense all pass.

🤖 Generated with Claude Code