Expose authoritative output-token capacity in the Codex model catalog

[jroth1111]

Problem

The Codex client model catalog exposes input-context metadata such as context_window, but does not expose the model registry output ceiling. Clients that need to reserve output headroom for proactive compaction therefore cannot derive a complete capacity contract from /v1/models.

This metadata must remain distinct from request translation: Codex currently rejects max_output_tokens when it is sent as a request parameter on this path. The catalog may advertise capacity without forwarding that field upstream.

Proposed behavior

• Add max_output_tokens to Codex client model entries when the registry has a positive MaxCompletionTokens or OutputTokenLimit.
• Apply the field consistently to template-backed and synthesized model entries.
• Omit the field when capacity is unknown; do not invent a default.
• Do not add or translate any request parameter.
• Add focused tests for known, fallback, and unknown output limits.

Why this belongs upstream

The source data already exists in registry.ModelInfo; exposing it in the model catalog provides a provider-neutral, machine-readable capacity signal for any Codex-compatible client without adding client-specific routing or hard-coded model margins.

Privacy

This report is based solely on public code behavior and synthetic request shapes. It contains no credentials, account identifiers, local paths, logs, or conversation content.