Add Gemini 4 Argon with bounded live eval coverage

[ProfSynapse]

Nexus can now use gemini-4-argon through the Google provider, with the documented thinking levels, sampling restrictions, output limits, timeout, and introductory pricing. The existing Google default stays in place. Responses requiring Google's separate long-decode continuation protocol surface as incomplete instead of silently appearing complete.

Live testing exposed Google's rejection of the existing tool-schema encoding. Send native parametersJsonSchema so the released Nexus tool catalog retains its JSON Schema constraints and works with Argon.

The eval harness now supports bounded output and tool rounds, consumes current chat runtime events, fails on aborted/failed turns, records per-response usage, and evaluates chat context against the current runtime contract. The committed five-scenario configuration uses low thinking, 2,048 generated tokens per request, four tool rounds, and no retries. Skill guidance and generated mirrors document the verification findings.

Validation:

• Production build and 62 focused regressions passed after rebasing onto current main.
• Full local suite before rebase: 6,141 tests passed, 89 skipped.
• Three gated live API smoke tests passed.
• Five of five live-model harness scenarios passed using scripted tools. The final measured run cost approximately $0.14; this excludes earlier exploratory runs and is a small integration sample, not a broad model benchmark.

Official references checked: model documentation, latest-model integration guidance, pricing, thinking, long decode, and function declarations. The model ID and limits were also checked against the authenticated API.