feat(presets): add 'optimal' preset (multi-provider speed-first routing)

[Destynova2]

Summary

• Adds a new shipped preset optimal.toml alongside existing cheap/fast/perf/medium/local.
• Targets the best price / perf / quality ratio for Claude Code users with Anthropic Max, validated by real benchmarks (2026-04-26).
• Wires 6 providers (Anthropic OAuth Max, OpenRouter, DeepSeek direct, Groq, Mercury, xAI), 8 virtual models, and 9 routing tiers.
• xAI Grok provider stub is included (enabled = false) with commented [[models.mappings]] blocks ready to uncomment after grob secrets add xai.

Routing strategy

Slot	Primary	Fallbacks
trivial / background	Groq gpt-oss-20b (960 t/s mesuré)	Anthropic Haiku → OR Haiku → Mercury 2
default	DeepSeek V4-Flash (81% SWE-V, 1M ctx)	V3.2 → Groq gpt-oss-120b → Sonnet 4.6
think	Claude Opus 4.7 (87.6% SWE-V, 1549 LMArena coding)	V4-Pro (80.6%) → GPT-5.5 → Sonnet 4.6
search	Claude Sonnet 4.6 (tool-use, OAuth free)	OR Sonnet → Gemini 3.1 Pro
Rust / k8s / Terraform	Anthropic only (quality on unsafe/HCL/YAML)	—
long context (>150k)	OR / DeepSeek (1M+ ctx)	—

Tier matchers cover max_tokens_below = 500 (trivial), min_input_tokens thresholds (30k / 150k), file-pattern + keyword combos for Rust, Terraform, Kubernetes, CI/CD YAML, and min_messages = 50 for big-history conversations.

Bench data baked into the preset (2026-04-26)

Model	t/s effective	RTT 200 tok
Groq openai/gpt-oss-20b	960	364ms
Groq openai/gpt-oss-120b	472	643ms
Groq llama-3.1-8b-instant	793	377ms
Mercury 2 (Inception Labs)	155 (reasoning overhead)	1250ms
DeepSeek V4-Flash (OR)	~83	743ms

Mercury's vendor-claimed 1109 t/s drops to ~155 t/s effective because the model is reasoning-heavy. It stays as a last-resort fallback.

Test plan

• python3 -c "import tomllib; tomllib.loads(open('presets/optimal.toml').read())" — TOML parses
• grob preset info optimal — preset lists 6 providers, 8 models, all router slots wired
• grob preset apply optimal --dry-run — emits valid normalized TOML
• No hardcoded version strings (grep -E '\bv[0-9]+\.[0-9]+\.[0-9]+\b' clean)
• Docs lint passes in CI (markdownlint + lychee)
• Manual smoke after merge: grob preset apply optimal --reload then a small request hits Groq via tier:trivial

🤖 Generated with Claude Code