2026-07-15: library 0.21.2 deployed (enriched /graph.md parsing)

Deploy PR 117 rolled the library to 0.21.2, restoring the floor's federation edges: the old column-count parser got zero edges (and phantom nodes) from the enriched six-column /graph.md the agent has published since 0.21.1. Verified in-cluster: fresh pod on image 0.21.2, ready, clean logs.

Two verification lessons from the incident that preceded this:

Check the artifact, not the traffic. The library's zero-downtime rolling update produces no external signal at all; a healthz dip watch (which worked for the broker) spun uselessly. Pod age plus image tag via kubectl is the reliable roll signal.
"Merged" needs its object confirmed. The code PR merged while the deploy bump sat committed but unpushed locally; Argo stayed correctly synced to old main and the 7-day-old pod was the first hard evidence. When a bump matters, verify origin/main actually contains it before watching for a roll.

Agent 0.21.3 deployed (H1 title fallback)

Deploy PR 118. Verified without cluster access via the self-proving output: /graph.md v766 (18:08Z) carries titles on all 147 soul rows (H1-extracted) while declared-metadata titles keep precedence on the cluster worlds; the old agent could only publish soul rows nameless. Also relearned the credential edge: the kubectl-based roll watcher silently reads expired gcloud auth as "not rolled yet" (2>/dev/null hid the failure), so prefer output-based verification when the artifact's output is version-distinguishing, and never discard stderr in a poll loop.

Library 0.22.0 deployed (rel-typed edges on world maps)

Deploy PR 119. Verified via kubectl: fresh pod on image 0.22.0, ready, clean logs (the roll watcher now fails loudly on auth/API errors instead of reading them as "not rolled yet"). World maps draw typed relations dashed with the predicate as tooltip; rendering is unit-pinned and the rel data was already confirmed live in the aggregate, so the remaining check is a human glance at the soul map's supersedes chains.

2026-07-15: library 0.21.2 deployed (enriched /graph.md parsing)

Two verification lessons from the incident that preceded this:

Check the artifact, not the traffic. The library's zero-downtime rolling update produces no external signal at all; a healthz dip watch (which worked for the broker) spun uselessly. Pod age plus image tag via kubectl is the reliable roll signal.
"Merged" needs its object confirmed. The code PR merged while the deploy bump sat committed but unpushed locally; Argo stayed correctly synced to old main and the 7-day-old pod was the first hard evidence. When a bump matters, verify origin/main actually contains it before watching for a roll.