release-train: develop -> staging

[LukasWodka]

Automated promotion by the release train (RFC-0008 D14). Head is the train-managed release-train/to-staging branch (a mirror of develop), so it never collides with a human PR. Merged only when the fr-gate is green.

---

Note

Medium Risk
Changes when the image-refresh CronJob triggers control-plane rollouts (jobs-manager Recreate, resource-monitor DaemonSet flap path); logic is heavily gated and tested, but upgrades still force a one-time repair roll on unmarked legacy edges.

Overview
Bumps the client Helm chart to 1.9.113 and changes the image-refresh CronJob so fresh installs no longer pay an automatic digest re-pin rollout when the registry digest is unchanged but workloads still run the floating :tag.

The script now stamps tracebloc.io/first-observed-<image> on first observation and tracebloc.io/digest-applied-<image> after a successful re-image. When recorded == latest and the live workload is off-digest, it skips rolling only if both markers show a fresh install (first_observed set, applied absent). Established edges (digest was applied before) still re-pin after helm re-renders revert pins; legacy edges with neither marker still take a one-time repair roll on upgrade so the fleet is not left on a possibly stale tag. Digest annotation reads are fail-closed (no || true) so API errors cannot mis-stamp first_observed.

Helm unittest and bats coverage lock the marker keys, skip condition, and the three behaviors (fresh / established / legacy).

^{Reviewed by Cursor Bugbot for commit 186c494. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[LukasWodka]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 186c494. Configure here.}

[LukasWodka]

Functional review — staging batch of 2026-09-10/11 (hops 18:00, 19:07, 20:40 UTC + overnight mirrors) — PASS (FR reviewer: LukasWodka; self-signoff per RFC-BACKEND-1405 D6)

Decision: PASS. All 39 cards at FR on staging — e2e-test-agent 15, backend 9, tracebloc-engine 6, client 5, docs 2, client-runtime 2 — advance to Ready for prod. Evidence pulled 05:37–06:11 UTC against the current staging tips (backend 458de261, client 6acc2f9, client-runtime eb412941, docs ea8a47f5, e2e-test-agent aebcf75b, tracebloc-engine 4b503980).

Journey on the staging chart (governs). Two tier-a baselines with chart_ref=v1.9.113-rc.2 — the rc staging carries, which includes the ghcr.io control-plane registry default (#1038):

• tabular_classification https://github.com/tracebloc/e2e-test-agent/actions/runs/34566748806 — installed: chart=client-1.9.113-rc.2; node leaves 1150m / 3254Mi, envelope fits; experiment COMPLETED, leaderboard score 0.9.
• image_classification https://github.com/tracebloc/e2e-test-agent/actions/runs/34566754199 — same chart and arithmetic; experiment COMPLETED, leaderboard score 1.0.
• Two further baselines on the superseded rc.1 (34566665511, 34566670712) were also green.

Customer-path control. Fleet on the published chart 1.9.112, 16 tasks, https://github.com/tracebloc/e2e-test-agent/actions/runs/34566675826: 13 green through Train; seq2seq red at Train with the training pod exiting 137 at the 3Gi envelope — the known model-zoo template case (also red on the daily develop run), tracked upstream, not this batch; the tabular and image legs were superseded by the queued baselines above (same concurrency group), which cover them.

Channels. fr-assist run 34532009992 (21:24 UTC): all OK, 1 SKIPPED (credentials).

Per-repo staging deploys. Green on every tip: backend Tests / docker-build / Migrations check / Lint; client CodeQL / E2E mysql / Release Helm Chart (rc.2 published); client-runtime Publish images / Tests; tracebloc-engine Run Tests / Promote task images by digest; e2e-test-agent Unit tests / CLI surface; docs CodeQL. Staging API answers (401 unauthenticated on /). Frontend smoke on the staging branch https://github.com/tracebloc/frontend-app/actions/runs/34566891772 green.

Per-ticket read. All 39 PR bodies scanned for holds, interactive/TTY behaviour and migrations: none of the batch changes guided prompts; no body declares a hold on shipping. The one Medium Bugbot finding that shipped with the e2e mirror is recorded by the train on its own ticket.

What this evidence cannot tell you (README → Functional review): per-ticket behaviour beyond the journey's path, anything after a failing step, and steps a run skipped. The docs pages (docs#174) have no staging preview and were reviewed as content only.