ADR 0005 — Distributed Execution Graph (fan-out / join)

• Status: Accepted (Step 5 of 5 — Cloud Run compute provider)
• Date: 2026-07-07
• Builds on: ADR 0004

Context

An incident fans out into independent sub-analyses (metrics, logs, cost, queries). The original plan (Epic 8) modeled these as one execution running on four Cloud Run instances that “synchronize through the Event Store.” That violates the ExecutionJournal single-allocator invariant — four instances allocating sequence numbers for one executionId corrupts the fold. There is no distributed sequencer.

Decision

New package @loom/graph-executor with executeGraph(dispatcher, nodes, opts). Each graph node becomes its own execution (distinct executionId) dispatched concurrently through the ExecutionDispatcher (bounded by limit via parallelLimit). Placement is free — in-process, or one Cloud Run instance per node — because each execution has exactly one writer. The graph relationship lives in the BranchRegistry as lineage edges (recordLineage parent→child), not in a shared, concurrently-written log. The join is the barrier: results return only once all branches settle, and branchRegistry.children(parent) enumerates the branches.

This reuses the exact machinery the branching examples already rely on (per- branch executionId + BranchRegistry), just fanned out instead of speculated.

Consequences

• The single-allocator invariant holds unchanged — no distributed sequencer, no multi-writer.
• Fan-out is placement-agnostic: the same executeGraph runs on LocalWorkerRuntime or CloudRunWorkerRuntime with no code change.
• Lineage (not a merge) is the join model — consistent with Loom’s branch-only, no-merge stance.
• Sub-executions here are independent flows; when branches are speculative same-flow variants, WorkflowRuntime.fork() (prefix copy) remains the right tool and composes with the same registry.

Verification

• In-process: a 3-node incident graph (metrics/logs/queries) runs to completion, each node a distinct executionId, each effect once, and lineage children returns exactly the three branch ids.
• Cloud Run: the same graph dispatched via CloudRunWorkerRuntime over a platform that boots one fresh instance per branch — three concurrent instances, all branches complete, three distinct executions.
• Full repo: build + typecheck clean, 66/66 test tasks green, biome clean.