Remove ontology; read participant aliases via cross-core join

[samseaver]

Two related fixes to reaction quick-search, plus a guard so the same class of bug cannot return silently.

The outage this came from

Reaction search was failing with what the browser reported as a CORS error. It was not CORS. Solr answered HTTP 400 undefined field participant_aliases, and a 400 carries no Access-Control-Allow-Origin header, so the browser discarded the response and blamed the wrong layer. The real error was in the body all along.

participant_aliases had never existed — not in any schema, not in the compiler, not on any core. This UI has queried it since v3.6.5.

1. ontology removed entirely

Gone from the biochemistry JSON and both Solr schemas, but the UI still carried nine references: two type declarations, two special cases that existed only to filter it out of compound queries, a detail row on each of the reaction and compound pages, an entry in the reaction fl list, an explanatory comment, and two tests asserting the workaround.

2. Participant aliases via cross-core join

Searching reactions by a participating compound's alias did not work. The clause beside the undefined field — aliases — looked like a working fallback and was not: stoichiometry children carry no aliases, so of 262,517 children exactly 0 matched it. Participant search worked by compound id and participant name only.

It now reads the aliases from the compounds core at query time:

{!join from=id to=compound fromIndex=${SOLR_COMPOUNDS_COLLECTION}}aliases:X

inside the existing block join. fromIndex comes from config so it follows the environment — compounds_staging when querying reactions_staging. A mismatch resolves against the other environment's data rather than erroring, which is why it is threaded from config rather than written literally.

Verified against live cores, not mocks:

term	by participant name	via join
ChEBI 30616	0	4,702
L-Glutamine	237	266
cpd00053	0 (200 via compound:)	—

30616 is not ATP's name, not its ModelSEED id, and not in any reaction's own fields; 4,702 matches cpd00002's participant count exactly. L-Glutamine shows the join adding 29 reactions name matching misses. The join is genuinely applied rather than parsed away: drop the {!join} prefix and the same query returns 0.

3. The guard, generalised

The existing test asserted "the query must not contain ontology" — protection against exactly one field. It now extracts every field: token from the generated query and asserts each is one the core defines. Three fields have reached production as undefined-field 400s — ontology, stoichiometry, participant_aliases — each fixed individually, none of them prevented.

423 unit tests pass.

Deployment ordering

This must ship before the backend field is withdrawn. Production currently serves participant_aliases as a denormalised field on the stoichiometry children, because the deployed UI needs it. Once this is deployed, ModelSEEDDatabase 42d9e70f is reverted and the cores reposted without it. Doing that in the other order takes reaction search down.

🤖 Generated with Claude Code