Add Kunavo provider

[cooldk]

Adds Kunavo — a unified AI gateway to Claude, Gemini, GPT and media models behind one OpenAI-compatible API, with transparent per-token pricing 30–70% under providers' list prices.

• providers/kunavo/provider.toml — OpenAI-compatible, env = ["KUNAVO_API_KEY"], api = https://api.kunavo.com/v1
• 10 model TOMLs using base_model inheritance (anthropic / google / openai upstreams) with Kunavo's own [cost] — model ids match GET /v1/models exactly
• logo.svg — currentColor mono

Compatibility verified today against the live endpoint: streaming ✓, tool calling (streamed tool_call deltas, finish_reason=tool_calls) ✓, stream_options.include_usage ✓, GET /v1/models lists all 10 ids ✓ — tested on both a Gemini and a Claude model.

Machine-readable catalog: https://kunavo.com/llms.txt · Docs: https://kunavo.com/docs/quickstart

Also serves the native Anthropic Messages API on the same key (/v1/messages); this PR registers the OpenAI-compatible surface only.

[github-actions]

Action items

• [high] [violation] providers/kunavo/models/*.toml (all 10 new model files) - Check: Reasoning models must declare reasoning_options reflecting the provider's actual API surface. Why: Each file uses base_model pointing at metadata where reasoning = true (e.g. anthropic/claude-sonnet-5, anthropic/claude-fable-5, openai/gpt-5.5, openai/gpt-5.3-codex, openai/gpt-5.4-mini, google/gemini-2.5-pro, google/gemini-2.5-flash), and that boolean is inherited through the merge — but none of the Kunavo TOMLs define reasoning_options. The schema refinement Must set reasoning_options when reasoning is true (packages/core/src/schema.ts:282) and the AGENTS.md checklist make this a hard merge blocker; bun validate will reject every file. Comparable aggregators serving these same models (OpenRouter, Kilo, Venice) all declare reasoning_options for them. Action: Add a reasoning_options array to each of the 10 files — reasoning_options = [] if Kunavo exposes no verified reasoning control, or the actual effort/toggle/budget_tokens controls Kunavo's API accepts (with a TOML comment documenting the exact request syntax for niche controls, per AGENTS.md).

• [low] [possible mistake] providers/kunavo/models/*.toml (all 10 [cost] blocks) - Check: Per-token pricing must be supported by a direct provider citation that states what it supports, and cache dimensions should be declared where the provider charges for them. Why: The PR body lists a catalog/docs URL but does not map it to the specific input/output/cache_read values. The Claude models declare cache_read while omitting cache_write, which every comparable Claude provider entry includes (OpenRouter, Kilo, Venice all publish cache_write); the Gemini/GPT models omit cache pricing entirely. These material factual values cannot be reviewed against an unmapped source. Action: Cite Kunavo's pricing source for the per-token costs (e.g., state explicitly that https://kunavo.com/llms.txt provides the pricing used) and confirm whether cache_write (Claude) and cache_read (Gemini/GPT) apply — add them wherever Kunavo charges for those dimensions.