* test(#3900): a unix socket under the repo root must not break the overlay (failing first)
* fix(#3900): the overlay walker skips non-regular files (sockets, FIFOs, device nodes)
The walker classified every not-a-directory entry as a file, so a unix
socket anywhere under the repo root (an MCP server, a language-server
daemon, a watcher) made copyFileSync throw ENXIO — the reporter
measured 32 local test failures across 6 files, all in before() hooks,
none about the code under test; a FIFO would block until a writer
appeared. The hard-link path does not save it: the overlay sits under
tmpfs, so cross-device EXDEV always falls through to the copy.
One guard at the classification site (srcStat.isFile() before the
copy/link branches) — sockets, FIFOs, and device nodes are not
repository content; CI never saw this because a fresh clone has no
working-tree daemons, which is exactly why it ate local debugging
time. Also adds opts.root (test-only root override, default REPO_ROOT)
so the regression tests build the overlay from a tiny synthetic root
in milliseconds instead of paying a full repo walk.
* chore(#3900): changeset fragment (pr number backfilled after PR creation)
* fix(#3900): review fold-ins — sibling cold-fixture guard, changeset format, root doc
buildColdInstallTree enumerated the live repo's hooks/ with a NAME-only
filter — the same class #3900 documents: a daemon socket or FIFO inside
hooks/ would reach cpSync and throw ENXIO (or block). Dirent type guard
added. The changeset gains the mandated bold user-visible lead, and
opts.root is documented at its definition (mirroring cold-runtime-lib-
fixture's repoRoot convention).
* chore(#3900): backfill changeset PR number (4062)
---------
Co-authored-by: sim <sim@local>
90fff40f5a