Expose full-text conditions and $score through Table.search and REST

Outcome

Integrate full-text search into Harper's existing query mechanisms rather than adding a parallel search API.

matches, matches_all, matches_phrase, matches_prefix, matches_fuzzy, and fuzzy-prefix compile from the existing condition tree.
Bounded same-index Boolean groups and supported structured AND conditions are planned correctly. Cross-index, relationship, unanchored negative, and unsupported ordering shapes fail before native execution.
$score and opt-in $highlights are available for projection without becoming stored attributes.
REST query translation and Table.search() produce equivalent results and errors; raw Tantivy syntax is never exposed.
Native candidate IDs are resolved to current Harper records. Indexed/current versions, structured predicates, and bounded refill are checked before return.
Table and attribute authorization applies to searchable fields, predicates, returned records, and highlights. Restricted-field membership cannot become an oracle.
Filtering does not alter BM25 scores. Pagination and tie-breaking are deterministic within a local ready generation without promising cross-node snapshots.
Product autocomplete uses prefix or fuzzy-prefix against the product index. Curated suggestions remain a normal customer-managed table with its own @fullText declaration.
Query parity, permissions, stale candidates, refill, pagination, highlights, unsupported shapes, and REST status mapping are covered by integrated tests.

Comment generated by kAIle (GPT-6 Astra)