Search: capability-gated cross-lingual retrieval #79

Context

Locale-scoped lexical search cannot find equivalent meaning expressed with different words in another language. A Thai query for วัฒนธรรม does not imply a match for an English document containing only “culture.”

That expectation requires query translation, multilingual embeddings, semantic retrieval, or another cross-lingual mechanism. It is separate from multilingual indexes, mixed-script literal recall, and source-locale fallback.

Goal

Reserve and design a capability-gated cross-lingual retrieval surface without weakening the guarantees of lexical search.

Questions

• Does translated-query lexical retrieval belong in the provider contract, an upstream query-expansion seam, or a host plug-in?
• How are translation provenance, privacy, latency, cost, and failure reported?
• How do semantic results compose with lexical matching and highlights?
• What relevance and language-pair evaluation is required before a provider declares support?
• How does the existing capabilities.semantic flag evolve to distinguish monolingual semantic retrieval from cross-lingual retrieval?

Constraints

• Never infer this capability from a translated interface or from nativeAnalysis.
• Never silently translate a query without a documented host policy.
• Lexical results must remain available when translation or semantic services fail.
• Result locale remains explicit on every hit.

Trigger

A real installation requires concept-level retrieval across languages and can supply representative multilingual relevance judgements.