fix(ingestor): spawn by floating tag + retire chart-pinned digest machinery (companion to client-runtime#94)

[LukasWodka]

Summary

Companion to tracebloc/client-runtime#94. That PR makes jobs-manager spawn the ingestor Job by its floating tag (ghcr.io/tracebloc/ingestor:0.3) with imagePullPolicy: Always by default, with digest-pinning kept as an opt-in. This PR wires the chart to that model and removes the now-dead machinery the old pinned-digest design required.

Why

The ingestor was the only spawned image pinned by a chart digest. A helm upgrade that reset values reverted images.ingestor.digest to the chart baseline, silently downgrading the ingestor — the failure mode behind the recurring "empty columns rejected as non-numeric" customer reports (a stale ingestor lacked the NaN-tolerance fixes). Training pods never had this problem because they spawn by floating tag + Always. This makes the ingestor match — the cluster always runs the current published ingestor, and a helm upgrade can no longer revert it because there is no pinned digest to reset.

What changed

jobs-manager-deployment.yaml — wire INGESTOR_IMAGE_REPOSITORY + INGESTOR_IMAGE_TAG (new) alongside the now-optional INGESTOR_IMAGE_DIGEST. Empty digest (the default) → floating tag.

values.yaml / values.schema.json — images.ingestor: add repository, default digest: "" (opt-in pin), keep tag: "0.3"; drop the dead autoRefresh flag. The v0.3.6 digest is retained as a comment for easy pinning. Repository + tag are overridable for air-gapped mirrors. imageRefresh: drop ingestorResolveFailureThreshold (class-2 only).

image-refresh-cronjob.yaml + _helpers.tpl — retire the class-2 (ingestor) pass entirely. The kubelet now resolves the ingestor digest at each spawn, so there is nothing for the CronJob to reconcile. It keeps only the class-1 pass (jobs-manager + pods-monitor rollout restart). imageRefreshEnabled no longer factors in the ingestor (the CronJob is skipped when both class-1 images are pinned).

helm-ci.yaml — ingestor-multiarch now validates the floating tag (the default spawn target) and only checks a pinned digest when one is set, instead of failing on the now-empty digest.

ingestor/ subchart (README + values) — update docs that described the old auto-upgrade / INGESTOR_IMAGE_DIGEST currency mechanism.

Backward compatibility

• Pinning still works: set images.ingestor.digest (or pass image_digest per-request via the subchart) → repo@digest + IfNotPresent. The INGESTOR_IMAGE_DIGEST env is retained as the opt-in pin.
• --reuse-values from older releases: all new env reads are nil-guarded.

Validation (local)

• helm lint --strict on all four platform values (aks / eks / bm / oc) — clean.
• helm template — renders the new env wiring; CronJob is class-1-only.
• helm unittest ./client — 167/167 pass.
• sh -n on the rendered image-refresh script — valid; if/fi + for/done balanced.

Note

Supersedes the stopgap baseline-digest bump (v0.3.6) as the default currency mechanism — the pinned digest is now opt-in. The v0.3.6 value is preserved in values.yaml as the documented pin baseline.

⚠️ High-blast-radius prod-infra change across the chart + CronJob that can't be fully exercised without a live cluster. The CronJob class-2 excision and the deployment env wiring deserve a careful look before promotion.

🤖 Generated with Claude Code

[LukasWodka]

👋 Heads-up — Code review queue is at 12 / 8

Above the WIP limit. The team convention is to review existing PRs before opening new work.

Open PRs currently in Code review (oldest first):

• client#215 — Post-install self-verification of CLI usability (#738) · author: @LukasWodka · no reviewer assigned
• client#223 — test(charts): add helm-unittest suite for requests-proxy-service template · author: @saadqbal · no reviewer assigned
• client#224 — fix(Wire SINGLE_NODE flag: chart default from hostPath.enabled + installer sets it #222): wire SINGLE_NODE — chart default from hostPath.enabled + installer · author: @saadqbal · no reviewer assigned
• client#225 — fix(mysql): probe over TCP so a moved socket can't kill a healthy DB · author: @LukasWodka · reviewer: @saadqbal
• client#229 — fix(proxy): make all egress workloads proxy-aware (wire the dead HTTP_PROXY_* values) · author: @LukasWodka · reviewer: @saadqbal
• client-runtime#93 — fix(chore: revert default CODEOWNERS (back to author-picks-reviewer) #92): gate GPU→CPU pending fallback to single-node clusters · author: @saadqbal · no reviewer assigned
• client-runtime#94 — feat: spawn ingestor by floating tag (Always) when no digest pinned — revert-proof across helm upgrade · author: @LukasWodka · no reviewer assigned
• data-ingestors#169 — feat: add token_classification (NER/POS) ingestion · author: @shujaatTracebloc · reviewer: @divyasinghds
• data-ingestors#184 — Fix: backtick-quote column names in upsert VALUES() clause (unblocks proteomics ingestion) · author: @LukasWodka · no reviewer assigned
• design-system#38 — feat(molecules): add ButtonGroup + InputButtonCombo · author: @waqaskhanroghani · reviewer: @saadqbal

Pull from review before opening new work. (This is a nudge from the kanban WIP check, not a block.)