fix(core): isolate Codex OAuth input-limit metadata

[hbw00111]

Summary

Codex OAuth turns failed before sending a request because public OpenAI API metadata supplied a 922,000-token input limit alongside a smaller OAuth window (272,000 or 372,000). Exclude that API input limit from Codex OAuth metadata, including refreshed catalogs and shipped fallbacks, so pinned and discovered windows remain usable. Public API metadata and explicit inventory/user input limits retain their existing behavior and conflict validation.

Fixes #5285

Verification

• Full production build, workspace typechecks, lint, format, Desktop/UI Knip, ASF headers and renderer architecture checks passed.
• Core suite: 843 passed. Related Runtime context-budget, model-discovery and subscription-fetch suites: 48 passed.
• Before the fix, the new pinned-window and discovered-window regressions failed with the reported error; both pass after rebuilding the fix.
• Compiled Runtime probe now resolves gpt-5.5, gpt-5.4 and discovered terra/luna to 272,000, and sol to 372,000, without throwing.
• Not exercised: authenticated ChatGPT OAuth turns, Windows packaged Desktop, or the full repository test matrix.

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex and native Codex subagents reproduced the issue, implemented and reviewed the fix, added regression tests, and ran local validation on behalf of hbw00111. This automated draft is awaiting independent human review and a human decision to advance it; AI review does not replace that review.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No