Make /model provider-model-effort aware and stabilize agents

[skyguan92]

Summary

• make /model target-aware with a 3-step provider -> model -> effort flow and persist defaults per provider target
• restore persisted provider/model/effort selections at startup without letting launcher env vars permanently override them
• stabilize delegated sessions by keeping spawned teammates on the host-managed provider route, forwarding PATH, and serializing git worktree mutations
• route direct Codex/OpenAI fetch paths through a proxy-aware retry helper so Codex web search and related calls recover from transient socket-closed failures

Testing

• PATH=/Users/jguan/.bun/bin:$PATH /Users/jguan/.bun/bin/bun test src/services/api/codexShim.test.ts src/services/api/openaiShim.test.ts
• PATH=/Users/jguan/.bun/bin:$PATH /Users/jguan/.bun/bin/bun test src/utils/worktree.test.ts src/utils/swarm/spawnUtils.test.ts src/services/api/fetchWithProxyRetry.test.ts src/utils/model/providerTargets.test.ts src/utils/managedEnvConstants.test.ts
• PATH=/Users/jguan/.bun/bin:$PATH /Users/jguan/.bun/bin/bun run build
• smoke: web-search-ok
• smoke: agent-ok