feat: configurable upstreams for local and self-hosted models

[armenr]

Adds a first-class [upstreams] table so an alias can point at any Anthropic-Messages-compatible endpoint — the headline use being a local model, which is what modelmux couldn't do before (its upstreams were hardcoded to the two cloud URLs).

What

• [upstreams] in routes.toml: name = { base = "...", auth = "..." }. anthropic (passthrough) and openrouter (bearer) stay built in; add your own on top.
• auth: passthrough | passthrough:ENV | bearer:ENV | none.
• Upstream names are no longer a fixed enum; loadConfig validates each model's upstream against the merged table and fails loud on an unknown one.
• No auth leak: Claude Code's inbound auth is never forwarded to a none/bearer upstream, so your Anthropic token stays out of a local process.

Local models

Recent Ollama (v0.14+) speaks the Anthropic Messages API natively, so no shim:

[upstreams]
local = { base = "http://localhost:11434", auth = "none" }
[models]
flagship = "local:qwen3-coder:30b"   # subagents on local Qwen, orchestrator stays Claude

(OpenAI-format runners like LM Studio/vLLM: put LiteLLM in front, point base at it.)

Verified

• bun run check green: eslint + tsc + 78 tests (+7), incl. an end-to-end integration test (custom local upstream routes correctly, Claude auth not leaked).
• Hand-driven in the real runtime: a [upstreams] local config parses/lists; a local model without a declared upstream fails loud with a clear message.
• README gains a Local & self-hosted models section.