fix(server): avoid parsing shadowed max-token aliases

[Graffioh]

Summary

(low prio fix)

• Parse only the highest-priority max-token alias.
• Preserve the existing alias precedence while ignoring malformed shadowed values.

Example

{
  "max_tokens": 100,
  "max_completion_tokens": "invalid"
}

Previously the unused max_completion_tokens value was eagerly parsed and rejected. The request now uses max_tokens: 100.

Testing

• Added one focused server unit test.
• git diff --check

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}