docs(codex-plugin): document type='http' workaround for plugin-add bug

[Aryamanz29]

Summary

Documents a known issue with Codex CLI's codex plugin add command — it drops the type field from the bundled .mcp.json when writing into ~/.codex/config.toml. Without type = \"http\" Codex tries to launch the URL as a subprocess and tool discovery silently returns 0 tools.

Why this matters

Empirically reproduced on a fresh install:

1. codex plugin marketplace add https://github.com/atlanhq/agent-toolkit ✅
2. codex plugin add atlan@atlan ✅ (plugin installed)
3. ~/.codex/config.toml ends up with:

   [mcp_servers.atlan]
   url = \"https://mcp.atlan.com/mcp\"
   enabled = true
   # ↑ note: missing `type = \"http\"`

4. Codex session opens → user prompts "Find tables..." → agent says "I don't see an Atlan callable surfaced through tool discovery yet" and falls back to spawning codex exec child processes as a workaround.

What this PR changes

Adds a callout in codex-plugin/README.md right after the codex plugin add instructions explaining:

• The exact symptom (zero tools loaded, codex exec workarounds)
• What [mcp_servers.atlan] should look like
• That Codex must be fully quit and relaunched after editing config.toml (mid-session edits don't take effect)

What this PR is NOT

• The plugin's own codex-plugin/.mcp.json is already correct — it specifies \"type\": \"http\". This is a Codex CLI bug, not a plugin source issue.
• No code changes. README-only.

Follow-up

File an upstream issue against openai/codex so the workaround can eventually be removed. Until then, this README note saves the next user 30+ minutes of debugging.

🤖 Generated with Claude Code

[Aryamanz29]

Closing — wrong place for the fix. The plugin's .mcp.json is already correct; the bug is in Codex CLI's codex plugin add command. Documenting a manual edit-your-config step in our plugin README is misleading. Right answer is an upstream issue at openai/codex (and meanwhile, point users at the install path that works).