Add Copilot auto routing via /models/session + /models/session/intent - #1

What

Makes github-copilot/auto follow Copilot's own auto selection — including the per-turn intent router that VS Code uses — rather than a static availability pick.

The flow (reverse-engineered from microsoft/vscode-copilot-chat + @vscode/copilot-api, verified against the lib's _mixinHeaders):

1. POST /models/session — per-conversation: returns the candidate pool + a Copilot-Session-Token (attaching it to chat requests is what puts them in the auto billing/rate-limit pool + discount).
2. POST /models/session/intent — the router: given the prompt, returns ranked candidate_models (reasoning vs. fast). The only call that sees the prompt.

Wired into opencode hooks: provider.models (inject Auto), chat.message (capture prompt), chat.params (route + override model id), chat.headers (attach session token), event (re-eval on compaction).

Routing behavior (mirrors VS Code)

• Cache-stable cadence: route on turn 0, stick for the rest of the conversation (keeps the prompt KV cache warm), re-evaluate only after session.compacted — matching VS Code's automode router (turnCount > 0 skip + invalidateRouterCache).
• Within-family: the cloned auto model has a single fixed endpoint (npm + url), so routing stays in the anchored family (a reasoning tier vs. a fast tier). A guard anchors on a family with ≥2 tiers so the router always has a choice. (VS Code re-instantiates the endpoint per turn and can cross families; opencode can only rewrite the model id, hence the constraint.)
• Default selection mirrors VS Code's _selectDefaultModel — first same-family model in available_models — instead of the undocumented selected_model.
• Reasoning effort: needs_reasoning raises the routed model's effort (reasoning_effort: high for OpenAI/Copilot, an extended-thinking budget for Anthropic); fast turns stay at the model default.

UI

• Picker/status bar read Auto; a toast surfaces each (re)route (e.g. Auto → gpt-5.4 · needs_reasoning).

Headers / gating

CAPI headers verified against @vscode/copilot-api's _mixinHeaders. The Copilot-Integration-Id / Editor-Version / Editor-Plugin-Version lines are left commented so the minimal header set can be tested first; the honest non-impersonating floor is Copilot-Integration-Id: code-oss (what a plain OSS build sends). Note the intent router is a Team-Internal/Panel-only experiment in shipped VS Code, so this tracks internal VS Code behavior.

Tests

bun test (7 pass), bun run typecheck, bun run build all green.