fix(auxiliary): forward max_tokens to all providers, not just Anthropic/NIM (#60388)

[kyssta-exe]

Description

_build_call_kwargs() previously only included max_tokens in the API request when the endpoint was Anthropic-compat or NVIDIA NIM. For all other providers (OpenRouter, OpenAI direct, custom/local, Copilot, Nous), the value was silently dropped — so reference_max_tokens in MoA presets, title_generation caps, and any other caller-supplied max_tokens had no effect.

The call_llm function already has a retry mechanism that strips max_tokens when a provider rejects it (e.g. ZAI vision models with error 1210, OpenAI GPT-5 models that require max_completion_tokens), so forwarding max_tokens unconditionally is safe: providers that accept it respect the cap, providers that reject it fall through to the existing retry path (lines 6524-6540 of auxiliary_client.py).

Testing

• Existing test suite covers provider fallback behavior when max_tokens is rejected
• The retry mechanism at lines 6524-6540 handles ZAI vision error 1210, OpenAI GPT-5 max_completion_tokens requirements, and generic unsupported_parameter errors

Closes #60388

[alt-glitch]

This was generated by AI during triage.

Related to #60391 (also fixes #60388): this PR is the single-file _build_call_kwargs fix; #60391 makes the same agent/auxiliary_client.py change plus adds MoA per-slot max_tokens (moa_loop.py, moa_config.py, tests) and was opened earlier. Subset/superset competing PRs for the same issue — not a duplicate. Also related: #60008 (gateway pinned/rehydrate superset), #58261 (gemini-native-only). Maintainer to pick the canonical fix.

[kyssta-exe]

Stale — 6-7 days without merge activity. Can resubmit if still needed.