Codex MCP rejects explicit project_path from a projectless cwd

[maverickalo]

Summary

In TraceDecay v0.0.73, the Codex MCP server can fail before dispatching a
read-only tool with an explicit project_path when Codex launched the server
from a projectless working directory.

Reproduction

1. Register and index a repository at /path/to/repo.

2. Confirm tracedecay projects context /path/to/repo resolves it.

3. From an unindexed directory such as /Users/example, run:

   tracedecay tool tracedecay_context \
     --args '{"project_path":"/path/to/repo","task":"verify selection"}'

4. The call fails with:

   no TraceDecay index found at '/Users/example'
   

5. The outer CLI selector succeeds from the same directory:

   tracedecay tool --project /path/to/repo tracedecay_context \
     --args '{"task":"verify selection"}'

The same failure occurs through the Codex plugin because its generated MCP
configuration launches tracedecay serve without --path.

Expected behavior

Read-only selectors such as project_path, project_id, and
project_selector should resolve through the global registry before the
server requires an active project derived from its process working directory.

Current workaround

Pin the Codex MCP launch to a registered bootstrap project:

{
  "args": ["serve", "--path", "/path/to/repo"]
}

Once bootstrapped, explicit read-only project selectors work. This workaround
is not ideal for genuinely projectless Codex tasks and is overwritten by
plugin refreshes.