provider: add Gondola

[Codonyat]

Adds Gondola as a provider.

Gondola is an OpenAI-compatible gateway that resells Venice AI inference below list price, metered per request and settled in USDC on Base. No subscription, no minimum. Venice is the supply, so the model ids are identical to the existing venice provider and Crush users can switch between the two without touching model names.

What is in here

• internal/providers/configs/gondola.json, 30 models
• the //go:embed + accessor + providerRegistry entry in internal/providers/providers.go (alphabetical section)
• InferenceProviderGondola in the const block and KnownProviders(), following provider: add atlas cloud #463

Defaults are kimi-k3 (large) and deepseek-v4-flash (small).

Coverage is frontier (Claude Opus 5, Claude Fable 5, GPT-5.6 Terra, Gemini 3.7 Flash, Grok 4.6), coding (Kimi K3, DeepSeek V4 Pro and Flash), cheap workhorses, and an uncensored line.

Where the numbers come from

Everything is generated from the public catalog at https://api.gondola-ai.com/v1/models, which needs no auth if you want to check it:

curl -s https://api.gondola-ai.com/v1/models | jq '.data[] | select(.id=="claude-opus-5")'

Two things worth flagging because they are easy to get wrong:

• default_max_tokens is the completion cap Gondola actually serves, not the base model's headline cap. Advertising a ceiling the gateway will not serve breaks harness compaction, so these are the real numbers.
• The cached fields follow the semantics in anthropic.json, where claude-opus-5 is in: 5, in_cached: 6.25, out_cached: 0.5. So cost_per_1m_in_cached is the cache write price and cost_per_1m_out_cached is the cache read price. They are populated only where Venice prices that tier, and left at 0 otherwise.

Auto-refresh

Gondola is a marketplace, so prices move with supply rather than being fixed by a vendor. I would rather this not go stale in your repo.

I am happy to follow the #463 shape and add cmd/gondola/main.go plus the Taskfile.yaml task and the update.yml step, so the config regenerates every 6 hours like the other providers. I left it out of this PR to keep the first one reviewable. Say the word and I will push it here.