decompose architecture into 38 atomic tasks across 12 parallel generations

Decompose the reviewed architecture specs into taskgraph-managed tasks:
- 2 setup tasks (project init, test infrastructure)
- 4 schema tasks (enums, node attrs, edge attrs, graph schemas)
- 1 error hierarchy task
- 6 graph tasks (FlowGraph class, 3 construction paths, queries, validation)
- 5 analysis tasks (type-compat, build-type-edges, ordering, template-validation, defaults)
- 5 component tasks (Operation, Sequential, Parallel, Conditional, Map)
- 2 host config tasks (GraphologyHostConfig, ReactiveHostConfig)
- 4 reactive tasks (WorkflowRoot, node-status, max-concurrency, retry-semantics)
- 3 review tasks (foundation, reactive-and-hosts, complete-library)
- 5 meta cluster tasks (schema, graph, component, reactive, analysis layers)
- 1 API exports task

Validated with taskgraph: zero cycles, 38 tasks, 12 parallel generations.
Critical path: 12 tasks through reactive execution layer.

tasks/graph-construction-call.md

@@ -0,0 +1,44 @@
+---
+id: graph/construction-call
+name: Implement call graph construction (fromCallEvents, updateFromEvent, addCall, addDependency, updateStatus)
+status: pending
+depends_on:
+  - graph/flowgraph-class
+  - schema/enums
+scope: broad
+risk: high
+impact: phase
+level: implementation
+---
+
+## Description
+
+Implement the construction and mutation methods for call graphs: `FlowGraph.fromCallEvents()`, `updateFromEvent()`, `addCall()`, `addDependency()`, `updateStatus()`, `updateCall()`, and `removeCall()`. These handle the dynamic call graph populated from call protocol events.
+
+## Acceptance Criteria
+
+- [ ] `FlowGraph.fromCallEvents(events: CallEventMapValue[]): CallGraph` — processes events in order, idempotent (duplicate events have no effect)
+- [ ] Event processing: `call.requested` → add node (status: "pending", add triggered edge if parentRequestId), `call.responded` → status: "completed" + output + completedAt, `call.error` → status: "failed" + error + completedAt, `call.aborted` → status: "aborted" + completedAt, `call.completed` → status: "completed" + completedAt
+- [ ] `updateFromEvent(event: CallEventMapValue): void` — single event processing for real-time subscription
+- [ ] `addCall(attrs: CallNodeAttrs): void` — adds call node; if `parentRequestId` present, creates triggered edge from parent to child
+- [ ] `addDependency(source, target): void` — creates depends_on edge with composite key `${source}->${target}:depends_on`
+- [ ] `updateStatus(requestId, status, extra?): void` — validates transition per state machine, throws `InvalidTransitionError` on invalid transition
+- [ ] `updateCall(requestId, attrs): void` — partial merge of call attributes
+- [ ] `removeCall(requestId): void` — removes node and all attached edges
+- [ ] Status transition validation: pending→running→completed/failed, pending→aborted, running→aborted (terminal states: completed, failed, aborted)
+- [ ] Call graph errors are resilient: unknown requestId in responded/error/aborted → log warning, ignore; unknown operationId → create node anyway with status "pending"
+- [ ] Unit tests: fromCallEvents with full event sequences, updateFromEvent real-time pattern, addCall with/without parent, status transitions valid/invalid, addDependency, idempotency
+
+## References
+
+- docs/architecture/call-graph.md — fromCallEvents, updateFromEvent, status lifecycle, mutations, abort cascading
+- docs/architecture/schema.md — CallEventMapValue mapping to CallNodeAttrs, edge key convention
+- docs/architecture/error-handling.md — call graph error boundary, InvalidTransitionError
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion