fix(#222): wire SINGLE_NODE — chart default from hostPath.enabled + installer

[saadqbal]

Companion to tracebloc/client-runtime#92 (PR #93), which gates jobs-manager's GPU→CPU pending-job fallback behind a SINGLE_NODE env var. This PR wires the flag through the chart + installer.

Why

On multi-node/elastic clusters (EKS/AKS) a Pending GPU pod usually means a GPU node is still autoscaling in; the runtime's premature GPU→CPU downgrade silently moves the experiment to CPU and drives the client-runtime#80 token race. On single-node k3d the fallback is correct. SINGLE_NODE lets the runtime tell them apart.

Auto-update safe

The client self-upgrades via the auto-upgrade CronJob (helm upgrade --reset-then-reuse-values). Existing clusters never re-run the installer, so their stored values have no SINGLE_NODE. The chart default is therefore derived from a signal already in stored values — hostPath.enabled:

Deployment	hostPath.enabled	SINGLE_NODE default	Behavior
Installer k3d / bare-metal single-host	true	"true"	GPU→CPU fallback on (unchanged)
EKS / AKS / OpenShift (dynamic PVC)	false	"false"	Pending GPU pods left for the autoscaler

So existing EKS/AKS clusters auto-stop the premature downgrade on next upgrade with no value change, and existing single-node installs keep today's behavior. An explicit env.SINGLE_NODE always wins (e.g. a fixed multi-node bare-metal/NFS cluster that wants the fallback sets "true").

Changes

• jobs-manager-deployment.yaml: inject SINGLE_NODE into both containers = env.SINGLE_NODE if set, else hostPath.enabled (quoted; hostPath nil-guarded for --reset-then-reuse-values); added to the generic-range exclusion so it's never double-emitted.
• install-client-helm.sh: installer writes env.SINGLE_NODE: "true" for new k3d installs.
• values.yaml + values.schema.json: document the flag (quoted string; env.additionalProperties already allows it — schema correctly rejects a bare boolean).
• MIGRATION.md: 1.5.1 upgrade note.
• Chart.yaml: 1.5.0 → 1.5.1 so auto-upgrade actually ships the template change.

Tests

+3 helm-unittest (default "false"; "true" when hostPath.enabled; explicit override wins and is emitted exactly once) and +1 installer bats assertion. 178 chart tests green, helm lint clean. Rendered all scenarios manually (EKS→false, bare-metal→true, --set-string override).

Closes #222.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes GPU job scheduling behavior on upgrade for EKS/AKS (intended fix) and depends on a matching jobs-manager image; mis-set SINGLE_NODE could leave GPU jobs pending or trigger unwanted CPU downgrades.

Overview
Chart 1.5.1 wires env.SINGLE_NODE into jobs-manager so runtime can gate the GPU→CPU pending-pod fallback (client-runtime#92): on elastic clusters, long-pending GPU jobs stay on GPU instead of being downgraded after ~180s.

The jobs-manager Deployment now sets SINGLE_NODE on both containers: explicit env.SINGLE_NODE wins; otherwise it defaults to hostPath.enabled (quoted, nil-safe for --reset-then-reuse-values). The generic env range excludes SINGLE_NODE so it is not duplicated. The installer writes SINGLE_NODE: "true" for new k3d installs; values, schema, and MIGRATION.md document the flag and auto-upgrade behavior.

Tests: three helm-unittest cases for default/override behavior and one installer bats check for generated values.

^{Reviewed by Cursor Bugbot for commit eeaf810. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

👋 Heads-up — Code review queue is at 10 / 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-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
• data-ingestors#169 — feat: add token_classification (NER/POS) ingestion · author: @shujaatTracebloc · reviewer: @divyasinghds
• frontend-app#508 — Prod · author: @waqaskhanroghani · no reviewer assigned
• tracebloc-client#355 — fix(object_detection): scale RCNN targets to pixel space · author: @shujaatTracebloc · reviewer: @divyasinghds
• tracebloc-client#356 — Promote develop → staging: graceful exit on backend-stopped experiment (#353/#354) · author: @saadqbal · no reviewer assigned
• tracebloc-client#357 — chore(deps): bump pyarrow from 19.0.1 to 23.0.1 in /use_cases · author: @dependabot · no reviewer assigned
• tracebloc-client#363 — harden(tabular): fall back to random split when stratify is impossible · author: @LukasWodka · reviewer: @saadqbal

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