Decompose architecture into 28 atomic implementation tasks

Break the @alkdev/taskgraph architecture specs into dependency-ordered
implementation tasks across 8 component directories: setup, schema,
error, graph, analysis, cost-benefit, frontmatter, api, and review.
Each task has clear acceptance criteria referencing specific architecture
docs. Three review tasks serve as quality gates at critical junction
points (schemas-and-errors, graph-complete, complete-library). The
dependency graph is validated acyclic with 9 topological levels enabling
significant parallelism across independent work streams.

tasks/implementation/cost-benefit/dag-propagation.md

@@ -0,0 +1,51 @@
+---
+id: cost-benefit/dag-propagation
+name: Implement DAG-propagation effective probability computation
+status: pending
+depends_on:
+  - cost-benefit/ev-calculation
+  - graph/queries
+  - schema/numeric-methods-and-defaults
+scope: moderate
+risk: high
+impact: phase
+level: implementation
+---
+
+## Description
+
+Implement the DAG-propagation cost model in `src/analysis/cost-benefit.ts`. This is the core algorithmic contribution beyond the Rust CLI — it captures the structural reality that upstream failures multiply downstream damage.
+
+Per [cost-benefit.md](../../../docs/architecture/cost-benefit.md), the algorithm:
+1. Processes tasks in topological order
+2. For each task with prerequisites, computes `pEffective` from intrinsic probability + upstream propagation
+3. Upstream propagation: `parentP + (1 - parentP) × qualityRetention` for each parent
+4. `pEffective` = intrinsic × product of all inherited quality factors
+
+## Acceptance Criteria
+
+- [ ] `computeEffectiveP(taskId, graph, upstreamSuccessProbs, defaultQualityRetention, propagationMode)` — internal helper
+- [ ] In `dag-propagate` mode: for each task in topological order:
+  - Get intrinsic probability from `resolveDefaults(risk).successProbability`
+  - For each prerequisite, compute inherited quality: `parentP + (1 - parentP) × qualityRetention`
+  - `pEffective` = intrinsic × product of all inherited quality factors
+  - Store task's **actual** success probability for downstream propagation (use `pEffective` if this is the task's real probability)
+- [ ] In `independent` mode: `pEffective = pIntrinsic` (no propagation)
+- [ ] Completed tasks (`status: "completed"`): propagate with `p = 1.0` when `includeCompleted: false`
+- [ ] `qualityRetention` per edge defaults to 0.9, can be overridden per-edge via `defaultQualityRetention` option or edge attributes
+- [ ] Throws `CircularDependencyError` if graph is cyclic (needs topo sort)
+- [ ] Unit tests: simple chain (verify compounding effect), diamond graph, independent vs dag-propagate comparison matches Python research model results, completed task exclusion/propagation semantics
+
+## References
+
+- docs/architecture/cost-benefit.md — DAG-propagation algorithm, qualityRetention semantics
+- docs/architecture/decisions/004-workflow-cost-dag-propagation.md — ADR-004
+- docs/architecture/decisions/005-no-depth-escalation-v1.md — no depth escalation in v1
+
+## Notes
+
+> To be filled by implementation agent
+
+## Summary
+
+> To be filled on completion