fix(moa): replace deprecated gemini-3-pro-preview + forward max_tokens

[teknium1]

Salvage of #6636 (@devorun) onto current main.

Summary

Stops MoA from 404-looping on OpenRouter after google/gemini-3-pro-preview was pulled upstream, and forwards max_tokens on both reference and aggregator calls (was missing — the API silently accepted the param-less request but user-set limits were ignored).

Reported via Norbert's Discord cron failure (paste.rs/Q5WEB): 6 retries × 4 frontier models × 600s idle timeout = one dead cron job + $4.63 burned.

Changes

• tools/mixture_of_agents_tool.py: google/gemini-3-pro-preview → google/gemini-2.5-pro; add max_tokens to both chat.completions.create() calls
• tests/tools/test_mixture_of_agents_tool.py: convert the catalog-snapshot test to an invariant (non-empty, well-formed vendor/model slugs) — the old test was a change-detector that would flip red on the next churn anyway
• scripts/release.py: AUTHOR_MAP entry for @devorun

Validation

tests/tools/test_mixture_of_agents_tool.py .................. 3 passed

Does NOT fix the bigger cost problem (MoA still calls 3+ frontier models when the agent decides to use it, and cron currently ignores hermes tools config so users can't disable it there). Those are separate follow-ups — next PR wires cron to _get_platform_tools() so hermes tools actually gates cron toolsets.

Closes #6636.