[kvlonge]

Why

DuckShop should exercise every construct that SQL translation tools refuse, guard or fail on, so the benchmark alone can measure translation coverage without any other project.

Changes

• 86 new edge-case models with unit tests (160 edge models and 185 edge unit tests in total). Most place a construct refused only under strict checking before one refused by default, so one model shows both.
• Two fan-in marts (rpt_edge_case_model_keys, rpt_edge_case_array_model_keys) read the new models, so a refusal also reaches a downstream model.
• An open (uncontracted) source edge__supplier_notes with its loader and staging model, a SQL table function edge_basket_lines_for, and a JavaScript function created by a pre-hook.
• relations.py also checks and drops functions that SQL hooks create.
• Fixes sqb test on main: three core unit-test mocks now list columns in declared order (planning failed with S000 since the declared-columns change); core checksum updated (compiled tests only). A new e2e guard runs sqb test --inspect on DuckShop so fixture errors fail CI.
• README, RUNBOOK step 7 and sqlbuild_local.toml.example describe the new group. They also note that the edge connection sets database and schema, so unqualified calls resolve to the hook-created function.

Verification

• sqb compile: 0 errors, 0 warnings; sqb format --check clean.
• New test --inspect guard fails before the mock fix and passes after; DuckShop compile guards and benchmarks/duckshop/tests pass.
• Expectations captured on Snowflake under the edge target; a second capture left every file unchanged.
• sqb test --target edge --select tag:edge_case: 185 of 185 pass.
• make check-ci (verify hook), public-repository hygiene scan.