test(electric): verify compiled predicate meaning against backend results #1814
Send to my agentGoal#
Add semantic checks where compiler/request snapshots only prove the text stayed the same. Follow-up to #1808 and the existing integration portfolio in #772; reuse that infrastructure rather than introducing another generic adapter model.
First slice: Electric predicate compilation#
- Inventory which current snapshot tests already have an independent semantic counterpart in the adapter/conformance suites.
- Specify a small compatible predicate/value domain, including nulls, quoted strings, Boolean combinations, and supported comparisons. State exclusions and SQL three-valued logic explicitly.
- Generate finite row worlds and predicates; compare the production compiler's actual emitted query executed by the matching backend with a separately specified finite evaluator.
- Preserve snapshot tests for quoting, parameter binding, unsupported-expression errors and stable wire syntax. Semantic equality does not establish safe quoting.
- Calibrate with wrong bound values, lost predicates, incorrect null handling, and swapped comparison direction.
- Report backend/SDK versions and executed cells. A mock interpreter consuming original IR is not evidence that emitted SQL is correct.
Later adapter boundary work#
Use the same law/owner inventory to identify real SDK and native-storage gaps. Add provider-specific legal schedules to existing oracles, with captured or documented SDK contracts. Do not generate impossible SDK traces or replace durability/worker/IPC tests with callback mocks. Keep unexecuted service/device gates visibly separate from passing local models.
The desired result is a few independent semantic checks, not universal serialization, an all-adapter state machine, or new production abstractions. This issue proposes investigation and tests; it does not claim a newly reproduced adapter defect.
tannerlinsley closed this
Sign in to join this conversation.