fix: placeholder spacing and unsatisfiable constraints during introspection

[uwemaurer]

Fixes #10 — two gotchas reported by @mathijs81.

LIMIT ${param} generated LIMIT$1

Postgres lexes LIMIT$1 as the identifier limit$1 and rejects the query.

Root cause: when the parser reached a ${var} node it sliced the preceding SQL up to the end of the last token instead of up to the start of the variable, dropping the whitespace in front of the placeholder. Whitespace after a variable was kept — that asymmetry is why only the leading side broke.

Placeholders now go through pushPlaceholder(), which separates a placeholder from a preceding word character in all three rendering paths. This replaces the old fix-up that patched only the joined sql string, leaving it inconsistent with sqlParts (what the TS/Java template-literal renderers emit).

One existing snapshot already contained select 1 as n where$1 ::int is null, so the bug was baked into the test corpus.

Unsatisfiable foreign keys during introspection

Each QUERY/EXEC is introspected in its own transaction, so an INSERT with a NOT NULL foreign key can never be satisfied — the reporter worked around it with DEFERRABLE INITIALLY DEFERRED.

After initialization, referential integrity is no longer enforced:

Engine	Mechanism
SQLite	PRAGMA foreign_keys = OFF (better-sqlite3 turns it on by default)
PostgreSQL	session_replication_role = replica — best effort, needs superuser; SET LOCAL behind a savepoint in external-url: mode
DuckDB	no such switch, so a constraint violation in an EXEC warns instead of failing

Tolerating an EXEC failure is safe because nothing in the generated code depends on it running — parameter types come from the prepared statement. A QUERY (including INSERT ... RETURNING) still executes and still fails, since its result shape is what gets introspected.

Fixtures stay checked. The relaxation starts after BASELINE/MIGRATE/TESTDATA, so a foreign key violated by a fixture is still an error — that data is real and a broken reference in it is worth reporting. The alternative, requiring TESTDATA to satisfy every EXEC, would couple fixture primary keys to each query's @set literals (@set author_id = 42 demands an author with id = 42) and gets unworkable with self-referencing, composite, or multi-level FK chains.

Tests

• placeholder-spacing.test.ts — spacing across all three renderings, sqlParts/sql consistency, no spurious space after operators
• foreign-keys.test.ts — SQLite and DuckDB generate despite an unsatisfiable FK, and a FK violation in TESTDATA still throws
• sqltool-pg.test.ts — same for Postgres against the testcontainer

22 snapshots updated; git diff -w over them is empty, i.e. whitespace restoration only (priority =$1 → priority = $1, VALUES (?,?,?) → VALUES (?, ?, ?)).

Verified: vitest 210 passed, just test-java BUILD SUCCESSFUL, just test-python 54 passed, pnpm build clean, biome unchanged from baseline (13 warnings / 2 infos before and after).

🤖 Generated with Claude Code