Member Junction
    Preparing search index...

    IT74 — the durable dispatcher actually executing graphs against a live database.

    Why this bundle exists. IT71 covers eighteen checks and not one of them runs a graph. Nine assert that metadata exists; nine drive Submit/Persist and verify the rows that come back. So everything from "the rows are correct" onward — claim, execute, propagate, roll up, settle — was unit-tested against fixtures and never against SQL Server. That gap is the same species as the one Phase 6 shipped and Phase 7 caught: a binding that looked right until something drove it.

    A stub agent runner, not a real one. The dispatcher takes TaskAgentRunner as an injected seam precisely so it can execute without the agent framework. Using a stub keeps this bundle in the deterministic tier — no model calls, no tokens, no network — while still exercising the real claim protocol, the real condition evaluator, the real rollup, and real rows. What is under test is the dispatcher, not the agent.

    Mutation-class throughout. Every check writes Tasks and Task Dependencies, so the whole bundle is gated behind RUN_MUTATION_TESTS=1 and tears down what it created, FK-ordered.

    Variables

    TaskGraphExecutionChecks