Record the provider-reported model on usage events when a router resolves an alias

[FabioLeitao]

Problem

With OpenAI-compatible routers and aliases (OpenRouter auto, LiteLLM aliases, local gateways), the model Zero requested is often not the model that served the request. The response says which one did: the stream carries model, and Codex Responses carries response.model. Zero drops that value. So zero usage report and the session log can only say "requested auto", never which model actually answered or how usage split across models over time.

Behavior I'm proposing

• Decode the provider-reported model into Usage.ResolvedModel when the stream provides it (OpenAI-compatible chunks and Codex response.model). When a stream reports usage across several chunks, the last non-empty value wins.
• Persist it as an optional resolvedModel key on usage events, through the existing single writer usage.EventUsagePayload. It is omitted when empty, so older readers and payloads are unaffected.
• No change to pricing or to the existing model key. Cost still comes from the session's configured model through the registry, exactly as today. resolvedModel is attribution, not billing.

Why it belongs in Zero

Zero already persists usage for local accounting. This makes that record truthful when routing is involved, at no cost when it is not. It is provider-agnostic: any router that echoes model benefits, and nothing is provider-specific.

Alternatives considered

• Replacing model with the resolved value would change pricing and break escalation attribution, which is why it is a separate key.
• Logging it only in debug output would leave zero usage report unable to use it.

Status

Implemented and tested on my fork: feat/resolved-model-usage on FabioLeitao/zero (448cc575), one commit on main (99721c76), 11 files, +129/-1, mostly tests. make fmt-check, go vet ./... and go test -race on providers, usage, zeroruntime and cli pass. I will open an issue or PR only if the core team wants it.