fix(rig-postgres): filtered and thresholded searches generate valid SQL

[jackthepunished]

Fixes #2376

rig-postgres produced SQL that Postgres rejects for any VectorSearchRequest that carries a member filter, a single-condition filter, or a threshold; and the threshold compared in the wrong direction. Found by auditing the vector-store crates against rig-sqlite, which gets all of this right.

#	bug	severity
1	PgSearchFilter::member renders id is in ($3, $4) — not a Postgres operator	high (hard error)
2	threshold rendered as distance > $N in the inner WHERE, where distance is only a select-list alias → column "distance" does not exist	high (hard error)
3	> on a pgvector distance while threshold is documented as a minimum similarity → would keep the worst rows	high
4	"WHERE" + condition with no separator → WHEREprice >= $3 for any single-condition filter	high (hard error)

Generated SQL on origin/main, as printed by the new tests when run there:

WHEREprice >= $3
WHEREdistance > $3
WHERE(distance > $3) AND ((kind = $4) AND (id is in ($5, $6)))

and on this branch:

WHERE (price >= $3)
WHERE (1 - (embedding <=> $1) >= $3)
WHERE ((kind = $3) AND (id IN ($4, $5))) AND (-(embedding <-> $1) >= $6)

Fix

• member renders IN.
• The threshold is applied as a minimum similarity on a per-operator expression that repeats the distance operator rather than naming the alias: 1 - (embedding <=> $1) for cosine and jaccard, the negated distance for inner product / L2 / L1 / hamming (pgvector's <#> is already the negative inner product). This is the approach rig-sqlite takes. It is composed after the $ renumbering pass so the $1 inside the expression survives, and its bind value is appended after the filter's.
• WHERE is separated from its first condition.

Non-breaking: returned scores are still raw pgvector distances in ascending order; only requests that previously errored change behavior.

Tests

• 4 unit tests fail on origin/main and pass here (threshold_is_a_minimum_similarity_on_the_repeated_expression, threshold_placeholder_follows_the_filter_placeholders, single_condition_filter_is_separated_from_where, plus the existing every_parameterised_operator_uses_dollar_placeholders, whose assertion pinned is in). They call the private search_query on a store built with a lazy pool and rig_core::test_utils::MockEmbeddingModel, so no database is needed. similarity_expression_inverts_each_distance_operator covers all six operators.

Verified locally: cargo fmt --all -- --check, cargo clippy -p rig-postgres --all-targets --all-features, cargo test -p rig-postgres.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.