support multi-dot partial search via index-side edge-grams

[mj-palanker]

Context

#139 made progressive typing of dotted identifiers work (ad.ph02t1.admin.se → ad.ph02t1.admin.securitygroup) by appending a :* prefix operator to the last search term. On a prefix-dense index like ours (edge-grams mean every word contributes all its prefixes as lexemes), :* forces GIN to range-scan and union the posting lists of every lexeme starting with the term — a trailing common token like prod turned a sub-second search into a 43s query on a 2.1M-row tenant (OPS-2310). #139 is reverted in the PR below this one; this PR restores the multi-dot matching it provided, keeping all queries as exact lexeme lookups.

Root cause of the original miss

jargon's tokenizer merges dotted segments (admin.securitygroup stays one token), so edge-grams were computed on the merged string and a bare sub-token gram like se (prefix of securitygroup) was never indexed. #139 papered over that at query time; this fixes it at write time.

Changes

• Index side (normalizeVectorDocs): additionally emit edge-grams for each symbol-split sub-token (securitygroup → s, se, sec, …) at lowered weight. Purely additive — new vectors are a strict superset of old ones.
• Query side (FullTextSearchQuery): build symbol-aware variants of the input — original, symbols→spaces (ad ph02t1 admin se), and symbols-deleted (adph02t1adminse) — each with its stemmed twin, deduped and OR'd as exact websearch_to_tsquery('simple', …) terms. No :* anywhere. Inputs without symbols produce byte-identical queries to pre-fix: prefix match last search term for multi-dot queries #139 behavior.

Deployment properties

• The joined variant matches joined-token prefix lexemes already present in existing indexes, so from-the-start dotted typing works immediately on old rows — no backfill required.
• The split variant matches the new sub-token edge-grams on rows written after this change.
• Known gap: mid-path partials (admin.se) won't match rows indexed before this change until they're rewritten or backfilled. (fix: prefix match last search term for multi-dot queries #139 covered this case at query time; accepted trade-off.)

Testing

• TestSearchMultiDotPrefix restored verbatim from fix: prefix match last search term for multi-dot queries #139 (DB-backed, runs in CI).
• New unit tests: sub-token edge-gram emission, dot-variant query construction, and a table-driven guard asserting generated queries never contain :*.

[mj-palanker]

We are not moving forward with this, while we believe it could be correct, it's not cleat that is it. There are significant costs needed to re-index the world for this to properly work, and we are likely going to attempt to re-architect this system in some way instead of continue to put out small fires.
if you are a future person looking at this, and you feel so inspired as to pick up this approach and merge it, feel free to do so given that you have audited it and strongly believe it to be correct.