Define stable crawler test layers and fixtures

Problem

The crawler has useful parser, transport, lifecycle, database, queue, and live-model coverage, but there is no short test contract explaining which boundary owns which behavior. Some large tests cross several concerns, and future rule/runtime improvements could force broad fixture and assertion rewrites.

The goal is stable mocks plus representative integration tests, not more tests at every layer.

Relevant code:

• apps/server/src/scraper/**/*.test.ts
• apps/server/src/orpc/routes/external-sites/scrape-sources/**/*.test.ts
• docs/development/backend-testing.md
• #220

Desired outcome

Tests follow stable product/runtime seams so internal refactors do not require rewriting the suite.

Acceptance criteria

• Document the crawler test layers: pure parser fixtures; mocked HTTP/clock/model boundaries; database-backed lifecycle/service/sink integration; queue wiring smoke tests; live model evals in pnpm evals.
• Keep behavior assertions at public boundaries and avoid asserting internal call order unless order is the behavior.
• Add or extract only the small crawler builders/harnesses already justified by repetition, reusing the factory direction in #220.
• Keep historical rule fixtures owned by their version compatibility tests; current parser tests use only the current contract.
• Ensure review, price, and catalog each have one representative database-backed run from persisted configuration through the real sink, with remote HTTP mocked.
• Verify that changing a parser helper or run-plan implementation does not require updating route-level fixtures.

Non-goals

• Rewrite all existing crawler tests.
• Mock database transactions in tests whose purpose is persistence, ownership, or concurrency.
• Put live model behavior in deterministic pnpm test.

Start with the crawler tests under apps/server/src/scraper//*.test.ts and apps/server/src/orpc/routes/external-sites/scrape-sources//*.test.ts, then read docs/development/backend-testing.md and the factory direction in #220. Map existing tests to the requested layers, keeping current-contract parser fixtures separate from historical rule fixtures. Done means the layer contract is documented and review, price, and catalog each have a representative database-backed run with mocked HTTP, while pnpm test remains deterministic and live model evals stay in pnpm evals.