Refuse max_tokens above the real 4096 Ollama cap

[jmjava]

Summary

• Q-SLM-20 leftover Add tests for premium routing, local failover, and apply review #7: inspect_payload advertised MAX_TOKENS=8192 while chat() silently min()-clamped num_predict to MAX_TOKENS_CAP=4096.
• The gate and client now share the 4096 cap. Requests at or below the cap POST that exact num_predict; 8192 raises max_tokens_too_large and never hits Ollama.
• Leftover Add Halo host profile, C4 diagrams, and public roadmap #5 (MAX_BYTES) is untouched. No A8/A9, downstairs GPU, or Halo 4 work.

Test plan

• test_chat_num_predict_equals_requested captures POST options.num_predict and asserts it equals the requested value
• test_chat_max_tokens_above_cap_raises_without_post asserts max_tokens=8192 raises max_tokens_too_large and does not POST
• test_max_tokens_8192_is_too_large asserts the payload gate refuses 8192 and accepts 4096
• CI tests job green

Made with Cursor