Symptom

get_node on a Server.LLM node after a failed inference returned an empty meta.error field even though the node’s state was ERROR. Errors on LLM nodes were silently discarded.

Root cause

krill-sdk 0.0.48 (and 0.0.49, which was published from a different krill-oss PR before the interface-refactor landed) implements updateMetaWithError as a 29-arm exhaustive when (meta) { ... else -> meta }. LLMMetaData was never added to that chain, so every call fell through to else -> meta, returning the original metadata unchanged. ServerNodeManager called updateMetaWithError(node.meta, cause) unconditionally, so the node entered ERROR state but its meta.error stayed empty.

The krill-oss fix (PR #148 — abstract NodeMetaData.withError() replacing the when-chain) was committed to krill-oss but shipped to Maven Central under a version that had already been published (0.0.49 was taken by a prior PR), so the bug persisted in the artifact.

Fix

• Bumped krill-sdk in gradle/libs.versions.toml from 0.0.48 → 0.0.49.
• Added server/.../NodeMetaDataServerExt.kt: internal extension NodeMetaData.applyError(error) that handles LLMMetaData via copy(error = error) and delegates all other types to the SDK’s updateMetaWithError.
• Changed ServerNodeManager.setErrorState to call node.meta.applyError(cause) instead of updateMetaWithError(node.meta, cause).

NodeMetaDataServerExt.kt is explicitly marked for removal once krill-sdk ships abstract NodeMetaData.withError() (tracked in krill-oss#146).

Prevention

• Any new NodeMetaData subtype must be added to the SDK’s updateMetaWithError dispatch immediately, or the error message will silently vanish. The krill-oss fix (abstract interface method, compile-time enforced) eliminates this class of bug permanently once it ships.
• When a dependency version bump is blocked because the desired artifact was already published from a different commit, do not assume the published artifact contains the expected fix — inspect the jar (javap -c) before relying on it.
• Add a NodeMetaDataServerExtTest (or equivalent) covering every server-side node type that goes through error-state promotion, so a missing arm fails a test rather than silently returning stale metadata.