fix: preserve Luna context-limit error code

[deep1725]

Summary

• Preserve metadata.error_code: context_limit for the Luna model server's explicit maximum-length rejection.
• Keep the server's existing error-text redaction intact. Generic runtime failures are not classified as context limits.
• Document the code so clients can distinguish unsupported inputs from transient evaluator errors.

Scope

• User-facing/API changes: one safe metadata value on confirmed context-limit failures.
• Internal changes: classify the existing RuntimeError produced by a failed scorer response.
• Out of scope: changing the model context window or retrying/truncating input.

Risk and Rollout

• Risk level: low; successful results and other failures are unchanged.
• Rollout requires publishing the updated Galileo evaluator and rebuilding consumers that pin its version. Roll back by restoring the previous package version.

Testing

• make lint and make typecheck passed for the Galileo evaluator using public PyPI.
• No tests added or run locally; automated CI results will be reported separately.

Checklist

• Updated documentation for the metadata code.
• Identified evaluator package release and consumer image update as required follow-up work.

Related PRs

Repo	PR	Purpose	Status
agentcontrol/agent-control	#284	Safe Luna context-limit metadata	Open
rungalileo/ml-deploy	#159	Explicit exclusions and valid latency evidence	Open
rungalileo/orbit	#2681	Opt-in H100 128K ceiling and exact 100K GPU probe	Draft
rungalileo/galileo-infra-automation	#1140	H100 validation Helm context and served-context check	Draft

Context validation rollout

Release the Wizard image that permits the opt-in H100 128K ceiling.
Validate the production Gemma bundle with the exact 100K GPU probe before enabling the Infra configuration.
Confirm the selected chart uses that image and the live model catalog advertises 131,072 tokens.
Then dispatch a fresh Aegis evaluation with the fixed worker and evaluator.
The two context PRs remain drafts until live GPU validation succeeds.
Historical evaluation attempts keep their original evidence.