Lower type-checked definitions to DuckDB SQL

[jat255]

Third of four PRs porting the data-dict definition compiler (kata f6hz, stage 1). Stacked on #270.

This is where conformance completes. translation.code is now asserted byte for byte and notes exactly, for all 42 corpus definitions, against the shared fixture and against an installed data-dict binary.

Numbers are formatted by widening the significant digits until the text reads back as the same double, which is the R algorithm rather than repr(). repr() switches to exponent notation and would emit 1e+20 where data-dict emits the digits. Perturbing the formatter to repr() fails only the dedicated boundary test and leaves the 42-definition corpus green, so that test earns its place: the corpus never reaches the case.

LIKE lowers to the narrowest form an authored pattern allows, so a wildcard-free pattern becomes an equality and a single anchored wildcard becomes a prefix or suffix test. A pattern that is itself an expression stays a LIKE, because none of that is knowable at compile time.

The notes are why the emitter is hand-written rather than generated through sqlglot: they are collected as the writer walks the IR, and sqlglot has no way to produce them.

_export.py imports the emitter inside _resolve_one rather than at module scope, because the emitter needs Ir from _export. Worth a look if you would rather the IR moved to its own module to break the cycle properly.

Verification: 545 tests pass, ruff and pyrefly clean.

[jat255]

Reviewed 2026-09-04 23:05 MDT — manual review complete. Findings: 3 MEDIUM (circular-import workaround avoidable via the codebase's TYPE_CHECKING idiom; stale translations docstring/comment; small-magnitude float notation unpinned by tests), 1 LOW (emitter errors bypass translation["error"]), no blockers. All three MEDIUMs fixed in 9daded3; full suite green (545 passed), roborev clean.