Stabilize persistent rust-ci-full failures

[starr-openai]

Why

The last 10 completed rust-ci-full runs on main all failed. Downloaded JUnit artifacts showed the same 23 final test failures in all 10 runs, plus related Windows and compact parity failures recurring in most runs.

What

• give remote-aware RMCP stdio test servers the absolute cwd required by explicit remote MCP environments
• make the thread inventory detail assertion path-format-independent on Windows
• widen the remote-control auth recovery test timeout so loaded CI workers can observe the wakeup
• carry forward the prior narrow full-ci candidates for code mode output history, deferred code mode guidance, compact parity normalization, local network-denial placement, and PowerShell literal output

Evidence

• JUnit window: rust-ci-full runs 26621289198, 26620612135, 26619249101, 26617558236, 26615321582, 26609276281, 26607181426, 26607124222, 26606958861, 26605931156
• RMCP log anchor: run 26621289198, job 78449098722, stderr: remote stdio MCP server \rmcp` requires an absolute cwd`
• Local validation: cd codex-rs && just fmt
• Intended validation: branch-triggered rust-ci-full run.