providers/sdk: budget pricing cannot reach feed rows for native vendor-endpoint constructors (kimi/zai/minimax) — add a base-URL-authoritative pricing key

[luizParreira]

OpenAIProvider::kimi / ::zai / ::minimax differ only by base_url; provider() reports "openai" and the native model ids carry no vendor prefix (kimi-k2.5, glm-5, MiniMax-M2.5). Budget pricing candidates therefore never reach the feeds' vendor-keyed rows (("moonshotai","kimi-k2.5") etc.), and the static table files those models under slugs — so a feed-only native-endpoint model prices to None and can never trip max_cost_usd (documented fail-open; no wrong price is ever produced).

A model-id prefix map (kimi-* → moonshotai, …) was considered and rejected: a wrong vendor guess silently prices a run at another vendor's rate.

The localized, authoritative fix: the vendor IS known at construction from base_url (the BASE_URL_KIMI/BASE_URL_ZAI/BASE_URL_MINIMAX constants exist, and code already branches on base_url.contains("api.z.ai")). Add an additive, defaulted trait method on LlmProvider:

fn pricing_key(&self) -> Option<(String, String)> { None }

overridden by OpenAIProvider only when its base_url is one of those endpoints (e.g. ("moonshotai", "kimi-k2.5")), threaded into the budget the same way the cost estimator was (TurnContext + the contextless checks). Re-keys nothing in the static table; changes no audit/wire provenance.

Adjacent (same underlying gap — provenance does not name the serving backend): wrapper provenance for FallbackProvider/ModelRouter, documented as a residual on PR #368.

Refs: PR #368 (source-major pricing lookup), issue #356.

[luizParreira]

Scope addition — the provenance gap cuts BOTH directions: (1) native vendor-endpoint constructors (kimi/zai/minimax) cannot REACH their feed rows (original scope); (2) the budget layer INFERS OpenRouter from model.contains('/'), so a slash-qualified model served by a non-OpenRouter OpenAI-compatible backend (e.g. Groq's openai/gpt-oss-120b) prices at OpenRouter's listed rate — a misprice, not fail-open (Groq $0.15/$0.60 vs OpenRouter $0.036/$0.18 per million). A provider-supplied pricing_key() resolves both: the constructor knows its true backend, so the budget stops guessing from model-id syntax. Until then, PR #368 documents the Groq-style ambiguity as a residual.