Summary

Fixes #1271.

This updates /setup-gbrain for the Windows MSYS/Git Bash quirks reported against gbrain v0.18.2:

• document why the installer uses bun install --ignore-scripts on Windows shells, while bun link and gbrain init --pglite still run later
• use a Windows .exe fallback for the local stdio MCP registration path when command -v gbrain does not resolve the binary
• document the PGLite single-writer lock case where claude mcp list can report Failed to connect while the in-session mcp__gbrain__* tools work
• switch the local smoke test from stdin piping to gbrain put "$SLUG" --content ...

Tests / verification

• bun test test/setup-gbrain-windows-msys-quirks.test.ts test/setup-gbrain-path4-structure.test.ts ✅ 21 pass
• git diff --check -- setup-gbrain/SKILL.md setup-gbrain/SKILL.md.tmpl test/setup-gbrain-windows-msys-quirks.test.ts ✅
• bun run gen:skill-docs --host all --dry-run ✅ on the upstream-based branch before fork push rewrite; the fork-based branch reports stale generated host files because maxpetrusenkoagent/main is behind upstream and this token cannot push workflow-file updates without workflow scope. The PR diff against garrytan/gstack:main remains limited to setup-gbrain/SKILL.md, setup-gbrain/SKILL.md.tmpl, and the new regression test.
• bun run test:windows ⚠️ hit an unrelated existing browse/test/terminal-agent.test.ts source-guard failure in shard 5: the test expects spawnClaude( in the message handler, while current source uses maybeSpawnPty(...). Setup-gbrain focused tests still pass.
• Codex autoreview against the exact diff: CLEAN, no blocking findings.

Notes

No UI/API/deployment behavior changed. This is a skill-doc/template fix plus a source-level regression test for the Windows setup-gbrain contract.