feat: add MCP latency benchmark script and Windows investigation (#11)

[Kunall7890]

Summary

Windows validation for MCP latency benchmark — addresses #11.

What I found

After installing @nhonh/react-debugger@2.1.2 globally and running
the benchmark, I discovered that the mcp subcommand is an interactive
installer (creates a local /mcp directory), not a stdio JSON-RPC
server. The CLI only exposes -v, -h, and -y flags — there is no
--stdio transport available in the current published version.

This means Phase A (the MCP stdio server) has not been published to
npm yet, so latency data cannot be collected at this time.

Windows-specific bugs found and fixed in the script

#	Error	Root cause	Fix applied
1	spawn EINVAL	.cmd files need shell: true on Windows	Added shell: true to spawn options
2	spawn ENOENT	npm global binaries are .cmd wrappers — plain path fails	resolveBin() now prefers .cmd path from where output

These bugs are Windows-only and would not appear on macOS/Linux.

Files added

• scripts/bench-mcp-latency.mjs — reusable benchmark runner, works on
  Windows + Linux, auto-detects OS and writes the correct output file
• evidence/phase0-latency-windows.json — run metadata + environment info
• evidence/phase0-latency-windows-notes.md — full investigation writeup

Ready to re-run

Once react-debugger mcp --stdio (or equivalent) is available in a
published version, re-running:

node scripts/bench-mcp-latency.mjs --runs 50

on Windows will produce valid evidence/phase0-latency-windows.json
timing data automatically. No script changes needed.

Environment

Key	Value
OS	Windows 11 Home Single Language (10.0.26200)
Node	v24.1.0
Chrome	149.0.7827.103
Package	@nhonh/react-debugger@2.1.2