The five offer tests this branch added all read the SHOP_OFFERS constant; none
ran the seed, so the one guarantee the stable-id design exists to provide --
re-seeding does not duplicate content -- was covered by nothing. This runs
`seedVisibleVerticalSlice` twice against an in-memory shop-offer repository and
asserts the five stable ids survive.
Confirmed catchable: reverting the conflict target to
['shopId', 'itemDefinitionId'] leaves 4 rows under the fake, because both bag
offers carry `itemDefinitionId: null` and collapse into one. On Postgres it
would fail outright.
The prices (12/30/40/35/60) and thresholds (reputation 25 and 40, renown 3) are
pinned in the same test: AGENTS.md §39 forbids silent rebalancing, and the
structural tests never looked at a number.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
1a7018d790