fix: configure Copilot thinking limits

[danielsuit]

Summary

• advertise Subconscious Copilot models as thinking-capable
• preserve the 5,000,000 input and 65,536 output token declarations required for Copilot agent prompt rendering
• migrate profiles containing the interim undersized 12,288 / 4,096 defaults while preserving custom budgets
• normalize gateway origins, /v1 URLs, and full chat-completions URLs to one correct endpoint
• add regression coverage for model metadata and endpoint normalization

Tests

• /opt/homebrew/opt/node@22/bin/node --test (18 passed)

Companion: api-gateway PR #421.