feat(sqlserver): Part 4 — relation quoting and factory wiring

[axellpadilla]

Part 4 of the SQL Server Fusion adapter series. Part of dbt-labs#15714.

Stacked on #13 (Part 3). Base is part-3-sqlserver-auth, so this diff shows only Part 4's three files. Merge #11 → #12 → #13 first; this retargets automatically as each lands.

Two exhaustive matches stopped compiling when Part 1 added the variant, and relation_impl.rs has none — every arm it needs is a silent-correctness one that shows up as wrong SQL rather than a build failure, so this was worked from the list in the plan rather than from cargo build output.

backend_of — not claimed by any issue

adapter_factory.rs backend_of maps AdapterType::SqlServer to Backend::SQLServer, the same ADBC backend Fabric rides. It is one of the 45 E0004s and no issue in the series names it; #4 is the closest home, since without it no adapter can be constructed at all. Flagging it so the series checklist can be corrected.

relation_impl.rs

get_database joins the Databricks | Fabric | Postgres | … group that raises InvalidConfig when database is unset. The _ arm returns an empty string, so without the arm a relation missing a database renders as .schema.table rather than erroring.

get_canonical_fqn joins Fabric | Bigquery, which pass an unquoted path part through verbatim rather than case-folding it. SQL Server stores an identifier with the case it was given.

normalize_component is deliberately left alone. It models how the server resolves an unquoted name, which is a different question — folding to lower case is the correct canonicalization for a case-insensitive collation, and that is the bucket Fabric already sits in. Passing through here would make MyTable and mytable distinct semantic FQNs on a server that treats them as one object.

quoted now doubles an embedded delimiter for SQL Server. The shared default in dbt-schemas is:

fn quoted(&self, s: &str) -> String {
    format!("{}{}{}", self.quote_character(), s, self.quote_character())
}

so x"q rendered as "x"q" — unparseable, and v2 would reject names v1 accepts. dbt-sqlserver v1 fixed exactly this in dbt-msft/dbt-sqlserver#795, and the escaping is part of the quoting decision this port is built on.

Scoped to SqlServer rather than fixed in the shared default on purpose: quoted() is inherited by every adapter with a quote character, and widening it here would change rendering for Snowflake, Postgres, Redshift and Fabric in a SQL Server PR. Worth raising separately — the shared default looks wrong for all of them — but not worth burying in this diff.

new_sqlserver mirrors new_fabric. Its callers (metadata/get_relation.rs, the metadata module) arrive in Part 5, which is why it has none yet.

include_policy needs nothing. Its _ => Policy::trues() is already correct for 3-part naming; every explicit arm there is an adapter that drops a path part. #4's checklist lists it as work — it isn't.

Verification

Five tests in a new sqlserver module: three-part rendering, the missing-database error, delimiter doubling (dom\usr + x"q → "my_db"."dom\usr"."x""q"), case preservation through get_canonical_fqn, and construction through RelationStatic.

dbt-adapter still has 43 E0004s (Parts 5–7), so cargo test -p dbt-adapter cannot run on this branch as-is. To actually run these rather than just write them, I stubbed the remaining 43 arms locally with todo!(), ran the suite, and reverted the stubs before committing:

• 881 passed / 0 failed, the full dbt-adapter suite
• all 5 new tests pass
• the pre-existing generic::test_normalized_fqn_* tests still pass, which is what pins the quoted override to SQL Server only

The workspace count drops 46 → 44 here: relation/factory.rs and adapter/adapter_factory.rs. The rest are adapter_impl.rs (35), sql_types.rs (5), column_builder.rs (2), metadata/get_relation.rs (1) — and one outside this crate, dbt-df-providers seed_io.rs infer_seed_column_name_strategy, which no issue in the series claims. Earlier PRs in this stack said 45, all in dbt-adapter; that came from per-crate checks that never reached dbt-df-providers.

Collation

normalize_component is left on the lowercase arm above, which is correct under SQL Server's default _CI_ collation and wrong under a _CS_ one, where MODEL and model are two objects that fold to one semantic_fqn. Reachable only when a project turns quoting off — DEFAULT_RESOLVED_QUOTING is all-true, and the quoted path skips the fold entirely.

dbt-sqlserver v1 has the same defect and knows it: TestCachingUppercaseModel is @pytest.mark.skipped with "Fails because of case sensitivity. MODEL is coereced to model which fails the test as it sees conflicting naming." So this matches v1 rather than regressing against it, but nothing in dbt-core models a collation at all, and SQL Server is the only adapter where case behavior is configurable per database and per column. Marked TODO at the arm.

Closes #4