UR-Nexus 1.87.4

• Extended model-native thinking and performance controls across OpenAI, Gemini,
  Ollama, NVIDIA and OpenRouter. Both model pickers preview effort, supported
  on/off modes, token budgets (B) and serving/routing preferences (S) before
  confirmation. /thinking, /effort and /fast use the same contracts.
• Added current OpenAI/Gemini model contracts, native reasoning-off none,
  exact custom effort names, Ollama thinking.values, and NVIDIA schema-derived
  levels/defaults. Mandatory reasoning cannot be disabled; unknown controls
  are omitted. Gemini 2.5 uses validated per-model native token budgets.
• Added OpenAI Fast/Ultrafast/Flex and Gemini Priority/Flex where documented,
  plus explicit OpenRouter endpoint preferences. Premium pricing is labelled;
  Flex retains its lower-cost tier and receives a longer default timeout.
  Native controls preserve the selected model and package-manager installation.
• Automatically use Responses for all GPT-6 and supported Responses-only Pro
  models. Corrected older GPT-5 models and snapshot aliases to expose their
  documented reasoning-off modes and defaults. Fixed Gemini endpoint discovery
  incorrectly disabling tool support and
  stopped coupling OpenAI reasoning effort to answer verbosity. Cold compatible
  server discovery now resolves native controls before shaping the first request.
  Updated docs and added full-query, streaming, keyboard and Bun-free native
  regressions.
• Updated root Zod to the upstream-fixed release and backported early collection
  size rejection for two legacy Zod consumers. Release auditing verifies both
  backports by installed source hashes and behavior; their raw advisories remain
  documented with a review deadline.