fix(windows): run the better-sqlite3 boot-install via node, not npm.cmd+shell (#861 follow-up)

[ken-jo]

What / Why / How

Follow-up to #861 (e918eac). That fix stopped the Windows boot-install EPERM for the optional fetch deps in start.mjs by running npm's own CLI through node (shell:false honors cwd) instead of npm.cmd + shell:true (which DROPS cwd on Windows → npm resolves node_modules from the inherited cwd C:\Windows → mkdir C:\Windows\node_modules → EPERM). The identical npm.cmd cwd-drop still affected the native-dep install in hooks/ensure-deps.mjs — the reporter explicitly listed better-sqlite3 among the EPERMs, and that path was left unfixed.

With this PR applied, the better-sqlite3 boot-install uses the same cwd-honoring path, so it can no longer EPERM at C:\Windows\node_modules.

This PR absorbs e918eac's approach for the native-dep install:

• hooks/ensure-deps.mjs — resolve npm-cli.js beside process.execPath and run it via execFileSync(process.execPath, [npmCliJs, "install", …], { shell:false }) (honors cwd) + windowsHide:true; fall back to the npm.cmd shim only when npm-cli.js isn't beside node (so no host regresses); surface failures to stderr (was silently swallowed behind stdio:"pipe" + an empty catch).
• tests/hooks/ensure-deps.test.ts — pin the install branch to the node-CLI / shell:false path + the failure surfacing (source guard, matching this file's existing Windows install can leave better-sqlite3 native binding missing in v1.0.107 #408/Codex CLI 0.131.0: MCP startup_timeout_sec=30s default trips when codex prewarm + cold-start stack #634 patterns; the bug is Windows-only and doesn't reproduce on Linux CI).
• preserved: the ensureNativeCompat ABI-rebuild paths are untouched (they use npm.cmd too, but don't mkdir node_modules → no EPERM; left for a separate change).

Affected platforms

Agent / CLI scope:

• All platforms — hooks/ensure-deps.mjs runs on every adapter that boots the MCP server via start.mjs. No MCP/adapter behavior change; only the native-dep bootstrap install.

OS scope:

• Windows — the affected OS (npm.cmd drops cwd). Reachable on Windows + Node when better-sqlite3 is missing (codex/marketplace clones, broken installs) — the same condition that triggers the start.mjs turndown installs e918eac already fixed.
• Linux / macOS — unaffected: ensureDeps() returns early on Bun, and POSIX already runs npm with shell:false (cwd honored). The node-CLI path is equivalent there.

Scope notes:

• Only the "package not installed at all" install branch is changed. The probe and ABI-rebuild logic are untouched.

Root Cause

• e918eac (Windows: boot-time turndown/better-sqlite3 npm install fails EPERM (C:\Windows\node_modules) every session — shell:true drops cwd #861) fixed start.mjs but not hooks/ensure-deps.mjs, which still ran execSync("npm.cmd install better-sqlite3 …", { shell: true }). On Windows, shell:true/npm.cmd does not honor the cwd option (verified), so the install targets the inherited cwd → C:\Windows\node_modules → EPERM.
• ensureDeps() only early-returns on Bun (no modern-Node gate), so on Windows + Node it does run this install when better-sqlite3 is absent.

Validation

Check	Result
npm run build + npm run typecheck	passed (0 errors)
vitest run tests/hooks/ensure-deps.test.ts	26 passed (existing #206/#148/#408/#331 + 2 new #861-follow-up guards)
Windows (ssh my-window, Node v24.13)	node <npm-cli.js> install is-number launched from cwd=C:\Windows → installed into <target>\node_modules (installedInTarget:true, eperm:false); npm-cli.js found beside node

Behavioral evidence:

• BEFORE: execSync("npm.cmd install better-sqlite3", { shell:true }) — same cwd-drop class as the (now-fixed) turndown installs → EPERM at C:\Windows\node_modules when better-sqlite3 is missing.
• AFTER: node <npm-cli.js> install … (shell:false) installs into the plugin dir regardless of cwd — verified on real Windows from cwd=C:\Windows.

Cross-agent / cross-OS risk

Minimal. One install call site in hooks/ensure-deps.mjs; no new dependency; no MCP/adapter/hook-routing change. The node-CLI path is the same mechanism e918eac already shipped and validated for start.mjs. POSIX behavior is unchanged (already shell:false).

Checklist

• Tests added (source guard for the node-CLI/shell:false install + failure surfacing), integrated into the existing tests/hooks/ensure-deps.test.ts
• npm run build + npm run typecheck pass
• Generated bundles regenerate via CI (not committed)
• No new runtime dependency
• One PR, single logical fix (absorbs e918eac for the native-dep install)
• 3OS validation — Windows verified (mechanism); Linux/macOS unaffected (Bun early-return / POSIX shell:false)
• Targets next

[NgoQuocViet2001]

@ken-jo reviewed the Windows install path here.

The change matches the #861 invariant: the primary native-dep install now runs node <npm-cli.js> with shell:false, so cwd is honored instead of falling through the npm.cmd + shell path.

I also checked it locally on Windows:

• npx vitest run tests/hooks/ensure-deps.test.ts -> 26 passed
• smoke from parent cwd C:\Windows with cwd set to a temp target -> package installed into the temp target node_modules; nothing was written under C:\Windows\node_modules

Looks good to me.