chore(stack): update secure LLM component pins - #1321

No actionable comments were generated in the recent review. 🎉

_{Comment @coderabbitai help to get the list of available commands.}

🌿 Preview your docs: https://nvidia-preview-mcamp-chore-complete-secure-llm-component-pins.docs.buildwithfern.com/nvcf

Fresh source-driven multi-region validation exposed one propagation gap: the BDD environment configured Pylon 0.14.1 under api.env, but the stack-generated nvcf-api values omitted configured API environment entries. The API chart therefore retained its Pylon 0.10.0 fallback in the worker launch configuration.

Fixed by 6233968:

• merge every configured api.env entry over the built-in stack environment
• preserve all unrelated built-in entries and emit each key exactly once in deterministic order
• serialize configured values as data without template evaluation
• prove Pylon 0.14.1 reaches both generated Helm values and the rendered nvcf-api ConfigMap

Validation at this exact head:

• focused Helmfile-to-nvcf-api ConfigMap regression test: pass
• full offline self-managed stack suite: pass
• nvcf-api Helm lint and template: pass
• focused Helmfile BDD wiring plus full short BDD suite: pass
• ShellCheck, Bash syntax, whitespace, DCO, targeted sensitive-string review, and offline changed-files TruffleHog scan: pass
• independent review: no Critical, Important, or Minor findings

Remaining merge gate: rerun the fresh live multi-region scenario from this exact head after reconciling the API workload and recreating any existing LLM function launch specification, then confirm the running worker uses Pylon 0.14.1. Public-catalog availability remains a separate release gate.

Lifecycle documentation correction added in b5e469f based on the fresh validation result: existing LLM function versions retain the worker-sidecar image metadata captured when the version is created. Replacing pods or redeploying the same version does not apply a new Pylon image. After updating the control plane, create and deploy a new function version.

This follow-up changes documentation only. The full docs check completed with Fern at 0 errors and one existing warning; whitespace, ASCII style, DCO, targeted sensitive-string review, and independent review all passed. No generated artifact update is required because the Fern navigation links directly to this Markdown source and the edit is outside generated catalog blocks.

The authoritative live rerun should use this exact head. No merge action was taken.

Follow-up corrective commit: f6b800914f2ce715638b72caf037ec5065de7eb0

Live lifecycle proof showed that the Spring Kubernetes remote ConfigMap has higher precedence than the API environment. An environment-only Pylon override therefore could not reliably replace the published API chart profile's older default.

This commit makes api.remoteConfig.configData.nvcf.sidecars.llm-router-client-image the stack contract:

• arbitrary api.remoteConfig.configData and api.env entries are preserved;
• the stack-owned request-router worker address retains precedence;
• an enabled LLM addon defaults Pylon to the effective sidecar registry/repository at 0.14.1;
• the deprecated api.env.NVCF_SIDECARS_LLM_ROUTER_CLIENT_IMAGE key is translated for one compatibility window and omitted from the rendered API environment;
• conflicting, empty, whitespace-only, and non-string Pylon values fail clearly;
• BDD fixtures, public docs, and the generated installation skill use the chart-native path.

Sanitized validation:

• focused render test: explicit canonical value, computed default, nested passthrough, stack precedence, legacy translation, same-value compatibility, conflict/empty failures, and actual nvcf-api.yaml remote ConfigMap output;
• full self-managed stack suite: pass;
• BDD short suite: pass across all four packages;
• API Helm lint, kubeconform validation (11/11 resources), and sidecar release-artifact test: pass;
• generated installation skill: synchronized, idempotent, and unit-tested;
• Fern/docs check: 0 errors with the existing advisory warning;
• ShellCheck, Bash syntax, whitespace, DCO, targeted public-safety review, and changed-files TruffleHog scan: pass;
• independent senior re-review: Critical 0, Important 0, Minor 0, Ready.

The full BDD lint lane reports only two existing ST1005 findings in an untouched DSL helper; the changed BDD code is clean under the configured linters.

Remaining merge gate: run the source-driven live multi-region scenario from this exact commit. After the control-plane update, create and deploy a new LLM function version; replacing pods or redeploying the same version does not change worker-sidecar metadata captured at version creation.

Live BDD fixture correction: 624d11225b01e9c8f710811ff26c6e277d092d7f

The source-driven run confirmed that the new function version selected Pylon 0.14.1, but its image name still contained unresolved local-fixture placeholders. The BDD environment step applies only the dotted-path values listed by each scenario; it does not interpolate arbitrary fixture content.

This narrowly scoped fix removes the redundant explicit Pylon leaf from both local self-managed fixtures and removes the matching feature-table overrides. Local BDD paths now exercise the stack's computed canonical remote-config default from the effective registry and repository.

Regression coverage copies and renders both real fixture variants after registry configuration, then asserts:

• api.remoteConfig.configData.nvcf.sidecars.llm-router-client-image is exactly nvcr.io/sample-org/sample-team/pylon:0.14.1;
• the rendered property contains no REPLACE_WITH_ placeholder.

Focused and full self-managed stack tests, focused and full short BDD tests, API Helm lint/validation/tests, docs/Fern, ShellCheck, Bash syntax, changed-code lint, DCO, public-safety review, and changed-files TruffleHog all passed. Independent re-review reported Critical 0, Important 0, Minor 0, Ready.

The final live source-driven rerun should use this exact commit and create/deploy a new function version after reconciling the control plane.

Focused live proof at head 624d112:

• The rendered configuration and live remote configuration were both placeholder-free.
• A newly created worker reached Ready 3/3 on Pylon 0.14.1 with running image digest prefix sha256:9befb43d30d5660ac.
• Sanitized metrics were registration=5, reverse=3, and regions=2.
• An authenticated chat.completion succeeded, with no no_eligible_candidates result observed.

This is focused proof only. The final fresh 2-scenario/63-step run from this exact head is now pending. No merge was performed.

Final sanitized live acceptance result for a validation tree combining exact #1321 head 40bdda431fe25ee1e54f2cc8aec302d322ff0cb8, exact #1322 head 26070673a0e2e819ec2f5161a91ae24b5dc55ace, and current main containing #1300:

• Passed 2/2 scenarios and 63/63 steps in 955.96 seconds.
• Verified the certificate chain and hostname, ALPN h2, the Region B SAN and certificate readback, and HTTPS Watch.
• Observed 3 canonical Deployment identities and 2 canonical StatefulSet identities from an explicit HTTPS remote seed, with no dashed-IP aliases.
• Observed registration=5 and reverse QUIC=3, healthy authenticated backends, and a successful authenticated chat.completion with the fixed response and no no_eligible_candidates result.
• Running artifacts were Pylon 0.14.1 at digest prefix sha256:9befb43d30d5660ac and Stargate 0.14.1 at digest prefix sha256:97c621e4d8c65c636. These were the pinned promoted artifacts, not local sideloads.

Task cleanup is complete and zero task-owned clusters remain. #1321 automated checks and CodeRabbit are green; required human review remains. No merge was performed.