fix(batch): re-create the fs-preload temp file if an OS cleaner removed it (#951)

[hakanbaysal]

What / Why / How

Fixes the stale-preload half of #951.

The cm-fs-preload-<pid>.js FS-tracking snippet is written once at module load
into the OS temp dir. macOS $TMPDIR periodic purges (and systemd-tmpfiles on
Linux) can delete it while the MCP server process is still alive. The server
then keeps injecting NODE_OPTIONS=--require <missing file> into every batch
command, so all node/npm/npx children die at preload with MODULE_NOT_FOUND
until the affected server PID is killed. Non-node commands are unaffected,
which makes the failure look like a broken npm rather than a stale temp file.

Fix: extract the preload write into ensureFsPreload() — write-if-missing,
return the path — and call it at the ctx_batch_execute injection site so a
deleted file is restored before the prefix references it. The module-load write
is kept (same startup behavior and failure mode); the added cost is a single
existsSync per batch call.

The /reload-plugins worker-accumulation aggravation described in #951 is a
host lifecycle concern and intentionally out of scope here.

Affected platforms

• All platforms

Test plan

• New test: delete the preload file externally, call ensureFsPreload(),
  assert the same path is restored with the real tracker content
  (__CM_FS__), not an empty placeholder.
• npx vitest run tests/core/server.test.ts → 500 passed (incl. the 25
  runBatchCommands cases).
• npx tsc --noEmit → clean.

Checklist

• Tests added/updated (TDD: red → green)
• npm test passes (server suite; full build requires Node >=22.5)
• npm run typecheck passes
• Docs updated if needed — n/a
• No Windows path regressions (uses existing join(tmpdir(), ...) path)
• Targets next branch