feat(presets): add EU sovereign presets (eu-eco/pro/max)

[Destynova2]

Summary

Adds 3 strict-EU sovereign routing presets (eu-eco, eu-pro, eu-max) for users who need GDPR data residency. All three use the same providers (Scaleway FR + Nebius eu-north1 Helsinki); they differ only in routing strategy / model preference.

Why three tiers

Different users have different cost/quality budgets. Rather than picking one balance, ship three so users can grob preset apply eu-eco|eu-pro|eu-max and benchmark them.

Preset	Cost (4M tok/day, 22d)	Estimated SWE-V	Strategy
eu-eco	€15-25/month	~75-78%	Mid-size models only (gpt-oss-120b default, Qwen3-Next-80B think). No 405B/397B heavyweights.
eu-pro	€40-60/month	~82-85%	Balanced — Hermes-4-405B in think, Qwen3.5-397B in code, Qwen3-Next-80B-Thinking in default.
eu-max	€80-100/month	~85-87%	Preemptive — Qwen3.5-397B every default request, Hermes-4-405B every think.

Quality estimates are vs Opus 4.7 at 87.6% SWE-V. The trade for eu-max is purely cost: 4-5× the eu-eco bill for ~10 benchmark points.

Routing slots (all three presets)

default-model, think-model, background-model, trivial-model, code-model (new), search-model, vision-model (new).

The new code-model slot is wired by tier matchers for Rust / Terraform / Kubernetes / CI YAML keywords + file patterns — those auto-route to the EU coding-tuned models (Qwen3.5-397B-A17B on Scaleway ChatCode endpoint).

Why xAI Grok was dropped

Earlier versions of this branch included xAI Grok at https://eu-west-1.api.x.ai/v1. After research:

• xAI publishes a regional endpoint that fail-closes if EU cannot serve, but at-rest residency is unspecified (x.ai/legal/data-processing-addendum says "contact sales").
• No announced xAI EU datacenter — the eu-west-1 endpoint is most likely partner-cloud (Memphis is xAI's only confirmed datacenter region).
• For users who genuinely need strict-EU sovereignty, this is a contract gap.

The benefit Grok added was the 2M context window. Without it, max EU context drops to ~256k (Scaleway devstral-2-123b / Qwen3.5-397B), which still covers >99% of agentic Claude Code workflows — sessions exceeding 256k normally trigger a compact.

Sister presets

This complements the optimal preset (PR #295) which is the global multi-provider speed-first config. Different audiences:

• optimal — best raw price/perf/quality with US providers in the mix (Anthropic Max + Groq + DeepSeek + OpenRouter)
• eu-* — GDPR-strict EU customers who can't use US providers regardless of speed or cost

Test plan

• grob preset info eu-eco / eu-pro / eu-max — all parse, providers/models/router slots wired
• grob preset apply <name> --dry-run — emits valid normalized TOML
• No hardcoded version strings (CI guard)
• xAI Grok refs scrubbed from all three (grep -lE 'xai_eu|grok-' presets/eu-*.toml returns nothing)
• Docs lint passes in CI (markdownlint + lychee)
• Manual smoke after merge: grob preset apply eu-pro --reload then small request hits Nebius/Scaleway

🤖 Generated with Claude Code