v0.13.6

Fixes

• Fix nested DAG/foreach source resolution, incomplete required artifact reads, literal control-marker preservation, optional output sections, schema validation and public runtime-task types.
• Settle dynamic controller cancellation and deadlines before returning; preserve cross-round findings, immutable decision evidence, canonical task IDs and ancestor-relative nested workflow depth.
• Align raw artifact writers and readers with bounded hard-link ownership, host-bound integrity and actual runtime source aliases. Fix storage, retention, lease handoff and artifact-sidecar settlement races.
• Isolate parent usage by session and recover failed usage writes without reviving pruned runs. Bound previews and workflow listings, reject invalid command options, and correct CLI failure exits and RPC presentation.
• Require bounded local byte evidence and complete requirement reconciliation in bundled review helpers and scaffolds.
• Avoid repeated scans when aggregating stage metrics, and clarify evidence, retention, environment and runtime contracts.
• Bundle @agwab/pi-subagent@0.6.1, including its temporary hard-link cleanup race correction.
• Include the release-verification fix that accepts multiple valid npm registry signatures while retaining cryptographic verification.

Compatibility and operational notes

• Node.js requirement remains >=22.19.0; CI now covers Node 22.19.0 and 24 on Ubuntu and macOS. Runtime validation uses locked Pi 0.84.4; this release does not claim Pi 0.85.0 compatibility.
• Stricter checks intentionally reject invalid specifications, incomplete reads, inconsistent artifact ownership and unsupported review dispositions that could previously pass silently.
• New runtime-owned raw artifacts require host-bound integrity checks. Legacy raw artifacts remain scoped current-content reads marked original-bytes-unattested; no automatic migration or historical byte-attestation claim is made.
• Raw publication performs additional integrity work and can take longer. Bounded hard-link storage savings are retained without weakening ownership checks; no general end-to-end speedup is claimed.
• Pruning remains explicit and dry-run by default. Default model/thinking, verification requirements and internal-only experimental workflow boundaries are unchanged.