[andyfan1094]

Summary

Allow each provider platform (openai / claude / grok / gemini) to register multiple keyed endpoints - e.g. a second sub2api server, a spare group, or a dedicated image-generation key - instead of being limited to one key and one shared base URL per platform.

• New optional endpoints[] on ProviderProfile: each entry carries its own name, apiKeyEnv (credential ref), optional baseURL override (may point at a different gateway), optional wire-protocol override, and its own model catalog.
• Route naming: the legacy top-level key keeps the bare platform route (sub2api-openai); every named endpoint registers as sub2api-openai-<name> (name validated against [A-Za-z0-9_-]+). Endpoint credential refs are stored as SUB2API_<PLATFORM>_<NAME>_API_KEY.
• The pi-ai bridge emits one hand-declared profile per endpoint (display name gains a (name) suffix when a platform has several endpoints); endpoints without key + models are skipped, so partial/empty forms never surface dead routes.
• Settings page: each provider card renders an editor per extra endpoint (key field with stored-state checkmark, base URL override, model catalog with models.dev auto-fill, per-endpoint model discovery and usage lookup). Saving is fully backward compatible.
• Global image tools: tool model refs accept an optional #<endpoint> segment (provider: openai#backup) so analyze/generate can pin a specific endpoint; resolution falls back to whichever endpoint carries the model when omitted.
• Discovery / usage probes accept an endpoint parameter, use that endpoint's stored credential, and probe the endpoint's own base URL when it overrides the global one.
• Docs updated in both READMEs (config example + rules).

Backward compatibility

Existing single-key configs keep working untouched: no endpoints[] means exactly one route per platform, same route names, same settings shape. The settings schema treats all new fields as optional.

Testing

• npm run typecheck and npm run build pass.
• Smoke-tested the compiled bridge directly: legacy config produces identical routes; multi-endpoint config expands to sub2api-openai, sub2api-openai-backup, sub2api-openai-other-host; keyless/model-less entries are skipped; baseURL/protocol overrides land in the pi-ai profile; image-tool refs resolve to the named endpoint.