exec: non-required MCP servers let the first turn start with no tools #26

Problem

codex exec (0.162.0+fork.ca04973b4) starts a turn while a stdio MCP server that is not marked required is still handshaking. The first request of the turn carries none of that server's tools. Each exec / exec resume restarts the servers, so the first one or two calls of every turn are affected.

A model that names tools it read in instructions (for example Claude Haiku 5.5) then emits mcp__<server>__<tool> or a bare tool name with no namespace. Codex answers function_call_output: "unsupported call: <name>" and the call never reaches the server. Models that poll list_mcp_resources until tools appear hide the problem, so behaviour depends on timing.

Reproduction

A fake stdio MCP server that sleeps 25s before answering initialize, not marked required: the first request lists only the other servers' tools, and the model reports the tool unavailable. With required = true the same run succeeds.

Requested changes

1. exec should wait for every enabled server's startup (bounded by startup_timeout_sec), not only required ones.
2. Resolve mcp__<server>__<tool> names, and bare tool names that are unique across servers, when they arrive without a namespace.
3. Return "server still starting" instead of "unsupported call" when the named server has not finished its handshake.

Downstream workaround: msp-agent renders required = true on every [mcp_servers.*] (Versent/msp-agent#33). Tracked as Vogt WI-1514.