chore: release 0.9.20

[github-actions]

🚀 Release ${version}

0.9.20 (2026-08-24)

Features

• foreman: opt-in archival of task audit records and transcripts (#1655) (c577133)
• foreman: report the Kubernetes node on FleetNode.status, not as identity (#1649) (7563e8a)

Bug Fixes

• controller: carry Model tolerations onto the prefetch Job (#1622) (d0db838)
• controller: clear the controller's own schedulingStatus once a service is Ready (#1633) (315f34d)
• Empty payload.repo silently defeats the upstream-base fetch: task branches cut from a stale fork HEAD (#1625) (#1626) (27a015a)
• foreman: carry the in-pod result extras on every Job-mode branch (#1657) (d1ce51c)
• foreman: free the agent's in-process slot while a Job-mode task runs (#1635) (0d64fbb)
• foreman: gate the make-invoked CI checks, and pin them against the workflows (#1642) (4e7ab7b)
• foreman: make the CustomResourceState config produce usable metrics (#1650) (1620443)

Documentation

• add ROCm host-retune runbook for >64GB on Strix Halo (#1387) (fff5e59)
• fix guide instructions that fail against the shipped API (#1629) (fba0bf4)
• proposals: foreman run, an unattended orchestration loop (#1652) (df57f49)
• reframe the multi-GPU guide from an Issue #2 validation plan into a deployment guide (#1631) (907ac8d)
• runbook: correct the metal-agent memory-pressure runbook against pkg/agent (#1630) (f665674)
• serving one model across two DGX Sparks via llama.cpp RPC (#1620) (7eaf04b)

---

This PR was generated with Release Please. See documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

🤖 Created releases:

• v0.9.20

🌻