Centralize per-database logic behind a single SqlDialect abstraction

[lovasoa]

Introduce a SqlDialect abstraction to make per-database logic a single source of truth

Per-database knowledge is currently scattered across many files and keyed on two parallel enums that nothing forces to agree: SupportedDatabase (8 variants: Sqlite, Duckdb, Oracle, Postgres, MySql, Mssql, Snowflake, Generic) and sqlx's AnyKind (5 variants). Because the dialect-specific data is spread over ~6 independent match sites, several keyed on the wrong enum, the compiler cannot enforce completeness: each site silently falls through to a _ default. Adding or fixing a dialect today means hunting down every site and hoping none was missed, and a missed arm is a silent correctness/data bug, not a compile error.

Proof: the references are real and scattered

$ grep -rcn -E "SupportedDatabase::(Sqlite|Postgres|MySql|Mssql|Duckdb|Oracle|Snowflake|Generic)" src | grep -v ':0' ; echo "total:" ; grep -rn -E "SupportedDatabase::(Sqlite|Postgres|MySql|Mssql|Duckdb|Oracle|Snowflake|Generic)" src | wc -l
src/webserver/database/sql.rs:37
src/webserver/database/sql/parameter_extraction.rs:9
src/filesystem.rs:3
total:      49

$ grep -rn -E "AnyKind::" src | wc -l
      24

49 per-variant SupportedDatabase references (37 in sql.rs alone) plus 24 AnyKind references, all to be kept consistent by hand.

The scattered sites

• dialect_for_db returns the sqlparser dialect via an 8-arm match: sql.rs#L293-L304
• Placeholder style keyed on AnyKind (only 5 entries), with a silent fallback to a temp prefix when a kind is absent: parameter_extraction.rs#L25-L46
• Placeholder cast type keyed on SupportedDatabase (different enum, with a _ => default): parameter_extraction.rs#L96-L115
• Default pool size per kind: connect.rs#L83-L94
• Trace-context SQL inlined in the execution path: execute_queries.rs#L505-L514
• CSV import reaching into sqlx's private AnyConnectionKind to special-case the Postgres COPY path: csv_import.rs#L172-L177
• DB-filesystem CREATE TABLE per dialect: filesystem.rs#L265-L280
• Name detection in from_dbms_name, plus its inverse mapping duplicated inside the tests: sql.rs#L786-L792

The enum mismatch directly produces fragile fallthroughs that only exist because the key type is wrong:

• _ => Self::Generic in from_dbms_name
• _ => run_csv_import_insert (any non-Postgres connection)
• unreachable!(...) when a kind is missing from DB_PLACEHOLDERS, and the silent temp-prefix fallback for an unlisted kind (parameter_extraction.rs#L58-L60)

Proposed change

Introduce a single trait SqlDialect (or a method-rich impl on SupportedDatabase) that centralizes everything per-DB:

• placeholder style and placeholder cast type
• sqlparser Dialect
• otel db.system.name
• default pool size
• capability flags: native JSON, COPY FROM STDIN, trace propagation, rollback statement, :: cast support, MSSQL concat rewriting

Crucially, key the placeholder data on SupportedDatabase and make DB_PLACEHOLDERS (and the other tables) an exhaustive match, so the compiler forces every dialect to be handled and the unreachable!/_ => Odbc/_ => Generic fallthroughs disappear. Adding or fixing a dialect then becomes "implement one trait / add the arm the compiler demands" instead of editing ~6 scattered sites that nothing forces to agree.

Risks and mitigations

• Large but mechanical refactor. Mitigate by leaning on the existing dialect/placeholder tests in sql.rs (the ALL_DIALECTS table drives many of them) which already pin per-dialect behavior.
• Land it incrementally, one capability at a time, so each step stays reviewable and bisectable.

Expected win

• Single source of truth for per-DB behavior.
• Compiler-enforced dialect completeness instead of hand-synced matches across two enums.
• Removes the silent _ => Odbc / _ => Generic fallthroughs that can quietly mishandle (or corrupt data for) an unlisted dialect.

This refines the "split sql.rs by responsibility" suggestion in #1249: a SqlDialect abstraction is the natural seam to split along.