[kvlonge]

Why

Reviews of SQL conversion and analysis turned up shapes where a Snowflake-to-DuckDB conversion gave a different answer, or where valid Snowflake SQL was reported as an error. DuckShop's edge-case area is where such shapes live as executable examples, with expected values captured on Snowflake, so this adds them there.

Changes

• New fixed-row edge__* sources and staging models: basket lines, date and time texts, amount checks, VARIANT attribute values and category links.
• Edge models and unit tests for:
  • ordering through arrays;
  • dates and zoned times read from text, and time arithmetic;
  • numbers, NULL arithmetic and literal types;
  • VARIANT comparisons;
  • text patterns;
  • lineage-sensitive clauses (CONNECT BY, recursive CTEs, joins on USING);
  • compile-only analysis patterns.
• Five edge Python functions, covering volatile values, NULL input and values returned as VARIANT.
• Expectations captured on Snowflake under the pinned edge session.
  • Values Snowflake does not define (ties, volatile calls, MEDIAN above scale 12) are checked by assertions.
  • Shapes Snowflake rejects are left out and listed in the README.
• README and RUNBOOK updates: the new groups and counts, known limitations, and capture guidance.

Verification

• sqb compile reports 0 errors and 0 warnings, both with the edge area and on the core copy. The core-copy checksum is unchanged.
• The edge build and tests pass on Snowflake under the edge target. Two capture runs in a row produced identical files.
• pytest benchmarks/duckshop/tests passes.
• ruff format and lint are clean.
• make check-ci passes.
• The public-repository hygiene scan passes.