fix: MCP pool stability and auth schema edge cases

[Fahremback]

This splits out the non-controversial fixes from #1148 as requested. Includes the client-pool timeout/dedup logic + tests, and defensive schema updates for auth.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[codebuff-team]

The MCPClientPool addition (common/src/mcp/client-pool.ts) is the strong part of this PR: clean dedup-on-connect logic, timeout handling via withTimeout, and a leak-avoidance path that closes a stale connection if it resolves after a timeout fires. The four tests in client-pool.test.ts cover the real edge cases — concurrent dedup, retry-after-failure, partial/full close, and timeout-then-retry — and they read as if they'd actually catch a regression, not just pad coverage. This part is ready to be ported.

Two things worth tightening before/at port time:

1. This diff adds the pool but doesn't wire it into any existing MCP connection path. That's fine if the integration lives in the parent PR (feat: Hermes architecture, Agentic UI, and CLI optimizations #1148) and this really is just the isolated primitive, but as a standalone diff there's no way to see it's actually exercised by real callers. Worth saying explicitly in the PR body which call site will consume this.

2. cli/src/utils/auth.ts: relaxing name/email from z.string() to z.string().nullish() is a real behavior change with no test and no explanation of what response was failing validation. "Defensive" schema loosening on auth data is exactly the kind of change that can silently paper over a real upstream bug (e.g., a provider returning malformed user info) rather than fix it. Please include a repro or a comment on which auth flow produces null/missing name or email, and ideally a regression test in the same style as the pool tests.

Net: the pool code is a clean, mergeable-quality contribution. The auth.ts hunk is small but needs justification.