feat(mosaic): durable pre-update snapshot + verify net + restore CLI (#791 PR2) (#811)

packages/mosaic/framework/install.sh

@@ -109,6 +109,225 @@ restore_snapshot() {
 }
 cleanup_snapshot() { [[ -n "$SNAPSHOT_DIR" && -d "$SNAPSHOT_DIR" ]] && rm -rf "$SNAPSHOT_DIR"; SNAPSHOT_DIR=""; }
 
+# ─── durable operator-config snapshot (#791 PR2) ─────────────────────────────
+# A SECOND, independent safety layer, distinct from SNAPSHOT_DIR above:
+#   • SNAPSHOT_DIR is ephemeral (/tmp, deleted on success) and mirrors the WHOLE
+#     target for CRASH rollback if the sync aborts mid-write.
+#   • DURABLE_SNAPSHOT_DIR is RETAINED, holds only the operator-owned surface, and
+#     lives OUTSIDE the framework tree and any repo. It exists for the failure the
+#     crash-rollback cannot see: a sync that finishes "successfully" yet a
+#     manifest/logic bug let it modify an operator file. verify_operator_surface()
+#     (post-sync) heals from it; `mosaic restore` recovers from it days later.
+# Path convention is mirrored in packages/mosaic/src/commands/restore.ts — keep
+# the two in sync (there is no shared code across the bash/TS boundary).
+DURABLE_SNAPSHOT_DIR=""
+backup_root() { printf '%s/mosaic/backups' "${XDG_STATE_HOME:-$HOME/.local/state}"; }
+
+# Relative paths that a migration INTENTIONALLY removes from the target (e.g. the
+# legacy bin/ tree). Such a path is operator-classified by the manifest (unknown⇒
+# operator), so the durable snapshot captures it — but its post-migration absence
+# is correct, NOT a manifest bug. run_migrations() records each removal here so
+# verify_operator_surface() does not "heal" it back and silently undo the
+# migration (which would then be skipped forever once the version is stamped).
+MIGRATION_REMOVED_PATHS=()
+
+# True (0) if $1 (a path relative to TARGET_DIR) equals or lives under a path a
+# migration deliberately removed this run.
+is_migration_removed() {
+  local rel="$1" removed
+  for removed in ${MIGRATION_REMOVED_PATHS[@]+"${MIGRATION_REMOVED_PATHS[@]}"}; do
+    [[ -n "$removed" ]] || continue
+    [[ "$rel" == "$removed" || "$rel" == "$removed"/* ]] && return 0
+  done
+  return 1
+}
+
+# True (0) if any parent directory of $1 (relative to TARGET_DIR) is a symlink.
+# Restoring THROUGH a symlinked ancestor would let cp write snapshot contents —
+# possibly secrets — outside the target (CWE-59), so the verify net refuses it.
+has_symlinked_parent() {
+  local rel="$1" dir p seg
+  dir="$(dirname "$rel")"
+  [[ "$dir" == "." ]] && return 1
+  p="$TARGET_DIR"
+  local IFS='/'
+  for seg in $dir; do
+    [[ -n "$seg" ]] || continue
+    p="$p/$seg"
+    [[ -L "$p" ]] && return 0
+  done
+  return 1
+}
+
+# Emit (NUL-delimited, into file $1) the operator-owned relative paths that exist
+# under TARGET_DIR, classified via the shared manifest (deny-wins; unknown⇒
+# operator). Returns non-zero if the filesystem walk itself failed — we must
+# NEVER snapshot from a truncated scan (a `< <(find …)` process substitution
+# would hide that error; capture-then-check does not — cf. #791 blocker-D1).
+enumerate_operator_files() {
+  local out="$1" scan abs rel
+  scan="$(mktemp)"
+  if ! find "$TARGET_DIR" -type f -print0 > "$scan"; then
+    rm -f "$scan"
+    return 1  # OP-SCAN-GUARD
+  fi
+  : > "$out"
+  while IFS= read -r -d '' abs; do
+    rel="${abs#"$TARGET_DIR"/}"
+    # Not operator config: version marker and any VCS metadata.
+    case "$rel" in .framework-version|.git|.git/*) continue ;; esac
+    manifest_is_framework "$rel" || printf '%s\0' "$rel" >> "$out"
+  done < "$scan"
+  rm -f "$scan"
+}
+
+# Retain only the newest MOSAIC_BACKUP_RETENTION (default 5) snapshots. The
+# pre-update-<UTC-ts> names sort lexicographically = chronologically, so a
+# reverse sort is newest-first. Pruning failures are non-fatal (they only leave
+# extra old backups); the enclosing find's status is still honored, not swallowed.
+prune_durable_snapshots() {
+  local root keep list d i=0
+  root="$(backup_root)"
+  keep="${MOSAIC_BACKUP_RETENTION:-5}"
+  [[ "$keep" =~ ^[0-9]+$ ]] && (( keep >= 1 )) || keep=5
+  list="$(mktemp)"
+  if ! find "$root" -maxdepth 1 -type d -name 'pre-update-*' > "$list"; then
+    rm -f "$list"; return 0
+  fi
+  # Newest-first ordering needs `sort` (`-o` writes back in place — no `mv`
+  # dependency); if it is somehow unavailable, leave the backups untouched rather
+  # than risk pruning in an undefined order.
+  if ! LC_ALL=C sort -r -o "$list" "$list" 2>/dev/null; then
+    rm -f "$list"; return 0
+  fi
+  while IFS= read -r d; do
+    [[ -n "$d" ]] || continue
+    i=$((i + 1))
+    (( i > keep )) && rm -rf "$d"
+  done < "$list"
+  rm -f "$list"
+}
+
+# Take the durable pre-update snapshot BEFORE any mutation. Fail-OPEN: the durable
+# snapshot is a recovery bonus on top of the manifest (which already keeps the
+# sync out of operator paths) and the crash-rollback — so an un-writable backup
+# location warns and continues rather than blocking the upgrade. Everything it
+# creates is private (umask 077 + explicit 0700 dirs / 0600 files): the snapshot
+# mirrors operator config, which may hold secrets, and must never be world-readable.
+make_durable_snapshot() {
+  is_existing_install || return 0
+  local root ts dir list rel src dst count=0 old_umask
+  root="$(backup_root)"
+  # Fail-open if we cannot even stamp a timestamp: the durable snapshot is a
+  # recovery bonus and must never be the thing that aborts an upgrade.
+  ts="$(date -u +%Y%m%dT%H%M%SZ 2>/dev/null || true)"
+  if [[ -z "$ts" ]]; then
+    warn "Durable snapshot skipped: no UTC timestamp available (upgrade continues)."
+    return 0
+  fi
+  # umask 077 makes every dir/file the snapshot creates private from birth (it
+  # mirrors operator config, which may hold secrets). It is PROCESS-global, so we
+  # save and restore it around exactly this block — otherwise every later sync
+  # copy and new framework dir would inherit 0600/0700 instead of 0644/0755.
+  old_umask="$(umask)"
+  umask 077
+  if ! mkdir -p "$root"; then
+    umask "$old_umask"
+    warn "Durable snapshot skipped: cannot create backup dir $root (upgrade continues; operator files remain manifest-protected)."
+    return 0
+  fi
+  chmod 700 "$root" 2>/dev/null || true
+  dir="$root/pre-update-$ts"
+  if [[ -e "$dir" ]]; then           # same-second re-run: disambiguate
+    local n=1; while [[ -e "$dir-$n" ]]; do n=$((n + 1)); done; dir="$dir-$n"
+  fi
+  if ! mkdir -p "$dir"; then
+    umask "$old_umask"
+    warn "Durable snapshot skipped: cannot create $dir (upgrade continues)."
+    return 0
+  fi
+  chmod 700 "$dir"
+  list="$(mktemp)"
+  if ! enumerate_operator_files "$list"; then
+    umask "$old_umask"
+    warn "Durable snapshot skipped: could not enumerate operator files (upgrade continues)."
+    rm -f "$list"; rmdir "$dir" 2>/dev/null || true
+    return 0
+  fi
+  while IFS= read -r -d '' rel; do
+    src="$TARGET_DIR/$rel"; dst="$dir/$rel"
+    [[ -f "$src" ]] || continue
+    mkdir -p "$(dirname "$dst")"
+    if ! cp "$src" "$dst"; then
+      warn "Durable snapshot: could not copy operator file '$rel' (skipped)."
+      continue
+    fi
+    chmod 600 "$dst" 2>/dev/null || true
+    count=$((count + 1))
+  done < "$list"
+  rm -f "$list"
+  # Tighten every dir the copy created (mkdir -p honors umask, but be explicit).
+  find "$dir" -type d -exec chmod 700 {} + 2>/dev/null || true
+  umask "$old_umask"   # UMASK-RESTORE-NORMAL — restore before the upgrade proper resumes (see above)
+  DURABLE_SNAPSHOT_DIR="$dir"
+  ok "Durable pre-update snapshot: $count operator file(s) saved to $dir (recover with: mosaic restore --list)"
+  prune_durable_snapshots
+}
+
+# Post-sync safety net: a keep-mode upgrade must NEVER modify an operator file.
+# Compare every file in the durable snapshot to its current target counterpart;
+# any that changed (or vanished) was touched by a framework bug — restore it from
+# the snapshot and warn loudly. This does NOT abort: the framework itself synced
+# correctly; we only heal the operator collateral. Runs after the restore trap is
+# disarmed so its corrective copies can't spuriously trip a full rollback, and
+# every step is guarded so `set -e` cannot exit silently mid-heal (cf. blocker-D2).
+verify_operator_surface() {
+  [[ -n "$DURABLE_SNAPSHOT_DIR" && -d "$DURABLE_SNAPSHOT_DIR" ]] || return 0
+  local scan snap rel cur healed=0
+  scan="$(mktemp)"
+  if ! find "$DURABLE_SNAPSHOT_DIR" -type f -print0 > "$scan"; then
+    rm -f "$scan"
+    warn "Post-upgrade verify skipped: could not enumerate the pre-update snapshot at $DURABLE_SNAPSHOT_DIR."
+    return 0
+  fi
+  while IFS= read -r -d '' snap; do
+    rel="${snap#"$DURABLE_SNAPSHOT_DIR"/}"
+    cur="$TARGET_DIR/$rel"
+    # A migration may legitimately delete an operator-classified path (e.g. legacy
+    # bin/). Its absence is intended — do not heal it back, or the migration is
+    # silently undone and never re-runs once the version is stamped (#791 PR2).
+    is_migration_removed "$rel" && continue   # MIGRATION-SKIP-GUARD
+    if [[ ! -e "$cur" ]] || ! cmp -s "$snap" "$cur"; then
+      # Never restore THROUGH a symlink: an operator path swapped for a link would
+      # otherwise let cp write snapshot contents (possibly secrets) outside the
+      # target (CWE-59). Refuse a symlinked parent; drop a symlinked leaf and write
+      # a real file in its place.
+      if has_symlinked_parent "$rel"; then
+        warn "Operator path '$rel' has a symlinked parent under $TARGET_DIR; refusing to restore through it (possible tampering) — recover it manually from $DURABLE_SNAPSHOT_DIR."
+        continue
+      fi
+      [[ -L "$cur" ]] && rm -f "$cur"   # SYMLINK-LEAF-GUARD
+      # Guard mkdir too: under set -e (trap already disarmed) a bare failure would
+      # exit the whole installer before the recovery pointer below is emitted.
+      if ! mkdir -p "$(dirname "$cur")"; then
+        warn "Operator file '$rel' was modified by the upgrade but could NOT be auto-restored (parent dir unavailable) — recover it manually from $DURABLE_SNAPSHOT_DIR."
+        continue
+      fi
+      if cp "$snap" "$cur"; then
+        chmod 600 "$cur" 2>/dev/null || true
+        warn "Operator file was modified by the upgrade and has been restored from the pre-update snapshot: $rel"
+        healed=$((healed + 1))
+      else
+        warn "Operator file '$rel' was modified by the upgrade but could NOT be auto-restored — recover it manually from $DURABLE_SNAPSHOT_DIR."
+      fi
+    fi
+  done < "$scan"
+  rm -f "$scan"
+  if (( healed > 0 )); then
+    warn "$healed operator file(s) were unexpectedly changed by this upgrade and were restored from the pre-update snapshot. A keep-mode upgrade must never modify operator files — this indicates a framework manifest bug; please report it (#791)."
+  fi
+}
+
 # Reconcile contract files after sync: framework-owned overwrite (backup-once),
 # user-seeded seed-if-absent.
 reconcile_framework_files() {
@@ -326,6 +545,10 @@ run_migrations() {
   # Remove bin/ directory — all executables now live in the npm CLI.
   # Scripts that were in bin/ are now in tools/_scripts/.
   if [[ "$from_version" -lt 2 ]]; then
+    # bin/ and the rails symlink are operator-classified by the manifest (unknown⇒
+    # operator) and thus captured in the durable snapshot; record them as
+    # intentional removals so the post-sync verify net does not restore them.
+    MIGRATION_REMOVED_PATHS+=("bin" "rails")
     if [[ -d "$TARGET_DIR/bin" ]]; then
       ok "Removing legacy bin/ directory (executables now in npm CLI)"
       rm -rf "$TARGET_DIR/bin"
@@ -383,6 +606,9 @@ fi
 # not for a validation failure that has touched nothing yet (#791 blocker-1).
 if [[ "$INSTALL_MODE" == "keep" ]]; then
   manifest_load
+  # Durable, operator-scoped backup taken BEFORE any mutation (#791 PR2). Kept
+  # outside the framework tree; recovered later via `mosaic restore`. Fail-open.
+  make_durable_snapshot
 fi
 
 # Snapshot before any destructive file operation; restore on interrupt/failure.
@@ -421,6 +647,10 @@ run_migrations
 
 # File-system phase complete and consistent — clear the restore trap.
 trap - ERR INT TERM
+# Post-sync safety net: heal any operator file a manifest bug let the sync touch,
+# using the durable pre-update snapshot (#791 PR2). Runs with the trap disarmed so
+# a corrective copy can't spuriously trigger a full rollback.
+verify_operator_surface   # VERIFY-NET (#791 PR2)
 cleanup_snapshot
 
 # Testability / minimal-install hook: stop after the file-system phase, before any

packages/mosaic/framework/tools/quality/scripts/test-upgrade-durable-snapshot.sh

@@ -0,0 +1,313 @@
+#!/usr/bin/env bash
+# test-upgrade-durable-snapshot.sh — the #791 PR2 regression gate.
+#
+# PR1 gave keep-mode upgrades two protections: the manifest (a keep-sync only
+# ever writes framework-owned paths — operator config is structurally untouched)
+# and an EPHEMERAL /tmp snapshot that rolls the whole target back if the sync
+# CRASHES mid-write. PR2 adds a third, independent layer for the case neither
+# covers: a "successful" upgrade that a manifest/logic bug silently let touch an
+# operator file. That layer is a DURABLE, operator-scoped pre-update snapshot:
+#
+#   Part 1 (scope):      before any mutation, the installer copies exactly the
+#                        operator-owned files that exist into a retained backup
+#                        under $XDG_STATE_HOME/mosaic/backups/pre-update-<ts>/ —
+#                        framework files are NOT captured.
+#   Part 2 (perms):      the backup root, snapshot dir and every nested dir are
+#                        0700; every backed-up file is 0600 (never world-readable,
+#                        even though operator config may hold secrets).
+#   Part 3 (no leak):    a secret seeded into credentials.json is copied into the
+#                        snapshot (proving coverage) but its value never appears
+#                        on stdout/stderr — the snapshot reports counts/paths only.
+#   Part 4 (retention):  only the newest MOSAIC_BACKUP_RETENTION snapshots survive;
+#                        older ones are pruned.
+#   Part 5 (verify net): if the upgrade DID modify an operator file (injected here
+#                        with a cp shim that scribbles on SOUL.md while a framework
+#                        file is copied), the post-sync verify restores that file
+#                        from the durable snapshot and warns loudly. The control —
+#                        the same installer with the verify call stripped — leaves
+#                        the corruption in place, proving the net is load-bearing.
+#
+# Usage:  bash test-upgrade-durable-snapshot.sh
+set -uo pipefail
+
+FW="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"   # packages/mosaic/framework
+INSTALL="$FW/install.sh"
+ORIG_PATH="$PATH"
+FRAMEWORK_VERSION="$(grep -m1 '^FRAMEWORK_VERSION=' "$INSTALL" | cut -d= -f2)"
+
+# Control installers must live INSIDE $FW: install.sh derives SOURCE_DIR from its
+# own path and sources tools/_lib/manifest.sh relative to it, so a copy anywhere
+# else aborts before the sync. Each control is a shipped installer with one guard
+# line stripped (keyed off a `# <MARKER>` anchor), proving that guard load-bearing.
+# All controls share the .install-*.tmp.sh glob so one trap sweeps them on exit.
+VERIFYCTRL="$FW/.install-verifynet-control.tmp.sh"
+rm -f "$FW"/.install-*.tmp.sh
+trap 'rm -f "$FW"/.install-*.tmp.sh' EXIT
+
+# mk_control <marker-regex> <name> — echo a control installer path ($FW-local) that
+# is $INSTALL with every line matching /<marker-regex>/ deleted.
+mk_control() {
+  local path="$FW/.install-$2.tmp.sh"
+  sed "/$1/d" "$INSTALL" > "$path"
+  printf '%s' "$path"
+}
+
+pass=0; fail=0
+chk() { if eval "$2"; then echo "  ✓ $1"; pass=$((pass + 1)); else echo "  ✗ $1"; fail=$((fail + 1)); fi; }
+
+SECRET='SUPER-SECRET-TOKEN-do-not-log-pr2'
+SOUL_ORIG='# persona'
+# A framework file the sync copies (source ships it; the seeded target omits it,
+# so the bytes differ and cp is attempted). The Part-5 shim keys off this path.
+POISON_REL='guides/E2E-DELIVERY.md'
+
+# Seed a recognized keep-mode install holding four operator-owned files across
+# the identity file, an operator subtree, memory, and the credentials carve-out.
+seed_home() {
+  local H="$1"
+  mkdir -p "$H/agents" "$H/tools/_lib" "$H/memory"
+  printf '%s\n' "$SOUL_ORIG" > "$H/SOUL.md"    # recognized install → keep mode
+  printf 'MODEL=opus\n'      > "$H/agents/coder0.conf"
+  printf '# operator memory\n' > "$H/memory/note.md"
+  printf 'TOKEN=%s\n' "$SECRET" > "$H/tools/_lib/credentials.json"
+  echo 3 > "$H/.framework-version"
+  # Deliberately NO guides/E2E-DELIVERY.md so the sync copies it (framework file,
+  # bytes differ) — that copy is where the Part-5 corruption shim fires.
+}
+
+# A pre-v2 (legacy) keep-mode install: SOUL.md marks it recognized, and a bin/
+# tree with NO .framework-version makes installed_framework_version() report 1, so
+# the v1→v2 migration (which deletes bin/) runs. bin/ is unknown⇒operator, so the
+# durable snapshot captures it — the verify net must NOT heal the intended removal.
+seed_home_v1() {
+  local H="$1"
+  mkdir -p "$H/agents" "$H/tools/_lib" "$H/memory" "$H/bin"
+  printf '%s\n' "$SOUL_ORIG" > "$H/SOUL.md"
+  printf 'TOKEN=%s\n' "$SECRET" > "$H/tools/_lib/credentials.json"
+  printf '#!/bin/sh\necho legacy\n' > "$H/bin/tool.sh"; chmod +x "$H/bin/tool.sh"
+  # Deliberately NO .framework-version and NO guides/E2E-DELIVERY.md (see seed_home).
+}
+
+# A cp shim that, while the framework POISON file is being copied during sync,
+# swaps the operator credentials file for a symlink pointing at an attacker-
+# readable file OUTSIDE the target — simulating post-snapshot tampering (CWE-59).
+# The durable snapshot already holds the real credentials (it is taken before any
+# sync), so the verify net must restore a REAL file in place WITHOUT following the
+# link (which would write the snapshot's secret out through it). $EXFIL_TARGET is
+# expanded at shim-write time from the caller's environment.
+#
+# PORTABILITY (why this shim, not the real `cp`): the CWE-59 leak this exercises is
+# `cp` writing THROUGH a symlinked destination. GNU/BSD cp — what a real operator
+# runs `mosaic update` under — follows the dest symlink and leaks. busybox cp (the
+# Alpine CI image) REPLACES a symlinked dest instead of following it, so under the
+# CI harness the leak vector simply does not exist and the negative control could
+# never reproduce it. This shim therefore emulates the real-target GNU cp behavior
+# PORTABLY: when the destination is a symlink it writes the source bytes through the
+# link via redirection (which follows symlinks on every coreutils, busybox included);
+# otherwise it delegates to the host's real cp unchanged. Both the shipped-case and
+# the negative control run through this identical shim, so the ONLY difference
+# between them remains the SYMLINK-LEAF-GUARD — the control stays load-bearing and
+# non-tautological. It does NOT touch install.sh (approved) or the real assertions:
+# with the guard present the symlinked leaf is dropped BEFORE this cp runs, so the
+# dest is a real file and the delegate path is taken exactly as on a GNU host.
+make_symlink_leaf_shim() {
+  local dir="$1" home="$2"
+  cat > "$dir/cp" <<SHIM
+#!/usr/bin/env bash
+dest="\${@: -1}"
+src="\${@:(-2):1}"
+case "\$dest" in
+  */$POISON_REL)
+    rm -f "$home/tools/_lib/credentials.json"
+    ln -s "$EXFIL_TARGET" "$home/tools/_lib/credentials.json"
+    ;;
+esac
+# Coreutils-agnostic emulation of GNU cp's follow-through-dest-symlink behavior.
+if [[ -L "\$dest" && -f "\$src" ]]; then
+  cat "\$src" > "\$dest"
+  exit \$?
+fi
+exec env PATH="$ORIG_PATH" cp "\$@"
+SHIM
+  chmod +x "$dir/cp"
+}
+
+# A cp shim that, while the framework POISON file is being copied during sync,
+# also appends garbage to the operator SOUL.md — simulating a manifest bug that
+# writes outside the framework lane. The framework copy itself still succeeds
+# (real cp runs), so the sync completes 0 and the post-sync verify is what must
+# catch and undo the operator-file damage. The snapshot's own cp only ever
+# targets operator files (never guides/…), so it is never corrupted by this shim.
+make_corrupt_shim() {
+  local dir="$1" home="$2"
+  cat > "$dir/cp" <<SHIM
+#!/usr/bin/env bash
+dest="\${@: -1}"
+case "\$dest" in
+  */$POISON_REL) printf 'CORRUPTION-mid-sync\n' >> "$home/SOUL.md" 2>/dev/null || true ;;
+esac
+exec env PATH="$ORIG_PATH" cp "\$@"
+SHIM
+  chmod +x "$dir/cp"
+}
+
+# Run one keep-mode, sync-only upgrade with $XDG_STATE_HOME redirected to a
+# throwaway dir (so the real ~/.local/state is never touched). Optional args:
+#   $2 shim-maker (default none), $3 MOSAIC_BACKUP_RETENTION (default unset).
+# Echoes: "<exit>\t<out>\t<state-dir>\t<home>".
+#   $4 seeder (default seed_home) — swap in seed_home_v1 for the migration case.
+run_snap() {
+  local installer="$1" shim_maker="${2:-}" retention="${3:-}" seeder="${4:-seed_home}" H STATE OUT SHIM rc pathpre
+  H=$(mktemp -d); STATE=$(mktemp -d); OUT=$(mktemp); pathpre="$ORIG_PATH"
+  "$seeder" "$H"
+  if [[ -n "$shim_maker" ]]; then
+    SHIM=$(mktemp -d); "$shim_maker" "$SHIM" "$H"; pathpre="$SHIM:$ORIG_PATH"
+  fi
+  set +e
+  env PATH="$pathpre" XDG_STATE_HOME="$STATE" \
+      ${retention:+MOSAIC_BACKUP_RETENTION="$retention"} \
+      MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 \
+      bash "$installer" >"$OUT" 2>&1
+  rc=$?
+  set -e 2>/dev/null || true
+  [[ -n "$shim_maker" ]] && rm -rf "$SHIM"
+  printf '%s\t%s\t%s\t%s\n' "$rc" "$OUT" "$STATE" "$H"
+}
+
+# Resolve the single pre-update-* snapshot dir under a state dir (newest if many).
+snap_dir() {
+  find "$1/mosaic/backups" -maxdepth 1 -type d -name 'pre-update-*' 2>/dev/null \
+    | LC_ALL=C sort -r | head -1
+}
+
+echo "── Part 1/2/3: durable snapshot scope, perms, no-leak ──────────────────"
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$INSTALL")
+SNAP="$(snap_dir "$STATE")"
+
+chk "upgrade succeeds"                        "[ '$rc' -eq 0 ]"
+chk "exactly one pre-update snapshot created" "[ \$(find '$STATE/mosaic/backups' -maxdepth 1 -type d -name 'pre-update-*' | wc -l) -eq 1 ]"
+chk "snapshot: SOUL.md captured"              "[ -f '$SNAP/SOUL.md' ]"
+chk "snapshot: operator subtree captured"     "[ -f '$SNAP/agents/coder0.conf' ]"
+chk "snapshot: memory captured"               "[ -f '$SNAP/memory/note.md' ]"
+chk "snapshot: credentials carve-out captured" "[ -f '$SNAP/tools/_lib/credentials.json' ]"
+chk "snapshot: SOUL.md bytes preserved"       "[ \"\$(cat '$SNAP/SOUL.md')\" = '$SOUL_ORIG' ]"
+chk "snapshot: framework file NOT captured"   "[ ! -e '$SNAP/CONSTITUTION.md' ] && [ ! -e '$SNAP/$POISON_REL' ]"
+
+# Part 2 — permissions (0700 dirs, 0600 files); never world-readable.
+chk "perms: backup root is 0700"              "[ \$(stat -c '%a' '$STATE/mosaic/backups') -eq 700 ]"
+chk "perms: snapshot dir is 0700"             "[ \$(stat -c '%a' '$SNAP') -eq 700 ]"
+chk "perms: nested dir is 0700"               "[ \$(stat -c '%a' '$SNAP/agents') -eq 700 ]"
+chk "perms: credentials backup is 0600"       "[ \$(stat -c '%a' '$SNAP/tools/_lib/credentials.json') -eq 600 ]"
+chk "perms: SOUL.md backup is 0600"           "[ \$(stat -c '%a' '$SNAP/SOUL.md') -eq 600 ]"
+
+# Part 3 — the secret is backed up but never emitted to stdout/stderr.
+chk "no-leak: secret IS in the backup file"   "grep -q '$SECRET' '$SNAP/tools/_lib/credentials.json'"
+chk "no-leak: secret NOT on stdout/stderr"    "! grep -q '$SECRET' '$OUT'"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+echo "── Part 4: retention prune (MOSAIC_BACKUP_RETENTION) ───────────────────"
+# Pre-seed four dated snapshots, then take one real snapshot with retention=2:
+# only the two newest (the fresh real one + the newest pre-seeded) must survive.
+IFS=$'\t' read -r rc OUT STATE H < <(
+  H=$(mktemp -d); STATE=$(mktemp -d); OUT=$(mktemp)
+  seed_home "$H"
+  mkdir -p "$STATE/mosaic/backups"
+  for ts in 20200101T000000Z 20210101T000000Z 20220101T000000Z 20230101T000000Z; do
+    mkdir -p "$STATE/mosaic/backups/pre-update-$ts"
+  done
+  set +e
+  env PATH="$ORIG_PATH" XDG_STATE_HOME="$STATE" MOSAIC_BACKUP_RETENTION=2 \
+      MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 \
+      bash "$INSTALL" >"$OUT" 2>&1
+  rc=$?
+  set -e 2>/dev/null || true
+  printf '%s\t%s\t%s\t%s\n' "$rc" "$OUT" "$STATE" "$H"
+)
+chk "retention: upgrade succeeds"             "[ '$rc' -eq 0 ]"
+chk "retention: pruned to exactly 2 snapshots" "[ \$(find '$STATE/mosaic/backups' -maxdepth 1 -type d -name 'pre-update-*' | wc -l) -eq 2 ]"
+chk "retention: newest pre-seeded survives"   "[ -d '$STATE/mosaic/backups/pre-update-20230101T000000Z' ]"
+chk "retention: oldest pre-seeded pruned"     "[ ! -d '$STATE/mosaic/backups/pre-update-20200101T000000Z' ]"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+echo "── Part 5: post-sync verify restores an operator file (+ control) ──────"
+# Shipped installer: the cp shim corrupts SOUL.md mid-sync; verify must restore it.
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$INSTALL" make_corrupt_shim)
+chk "verify: upgrade still succeeds"          "[ '$rc' -eq 0 ]"
+chk "verify: SOUL.md restored to original"    "[ \"\$(cat '$H/SOUL.md')\" = '$SOUL_ORIG' ]"
+chk "verify: no corruption remains in SOUL.md" "! grep -q 'CORRUPTION-mid-sync' '$H/SOUL.md'"
+chk "verify: loud restore warning emitted"    "grep -qi 'restored from the pre-update snapshot' '$OUT'"
+chk "verify: secret still not leaked"         "! grep -q '$SECRET' '$OUT'"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+# Control: strip the verify call → the corruption must SURVIVE (net is load-bearing).
+sed '/# VERIFY-NET/d' "$INSTALL" > "$VERIFYCTRL"
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$VERIFYCTRL" make_corrupt_shim)
+chk "control: SOUL.md corruption survives"    "grep -q 'CORRUPTION-mid-sync' '$H/SOUL.md'"
+chk "control: no restore warning emitted"     "! grep -qi 'restored from the pre-update snapshot' '$OUT'"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+echo "── Part 6: verify net honors an intentional migration removal (+ control) ─"
+# BLOCKER regression: on a pre-v2 install, bin/ is operator-classified so the durable
+# snapshot captures it — but the v1→v2 migration deletes bin/ ON PURPOSE. The verify
+# net must SKIP that removal (is_migration_removed), or it heals bin/ back and the
+# migration is silently undone forever once the version is stamped.
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$INSTALL" "" "" seed_home_v1)
+chk "migration: upgrade succeeds"                 "[ '$rc' -eq 0 ]"
+chk "migration: legacy bin/ stays removed"        "[ ! -e '$H/bin' ]"
+chk "migration: operator SOUL.md untouched"       "[ \"\$(cat '$H/SOUL.md')\" = '$SOUL_ORIG' ]"
+chk "migration: version stamped to $FRAMEWORK_VERSION" "[ \"\$(cat '$H/.framework-version')\" = '$FRAMEWORK_VERSION' ]"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+# Control: strip the MIGRATION-SKIP-GUARD → the verify net restores bin/ from the
+# snapshot, silently undoing the migration (proves the guard is load-bearing).
+MIGCTRL="$(mk_control 'MIGRATION-SKIP-GUARD' migration-control)"
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$MIGCTRL" "" "" seed_home_v1)
+chk "control: bin/ wrongly restored by verify"    "[ -e '$H/bin/tool.sh' ]"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+echo "── Part 7: verify net never restores a secret through a symlink (+ control) ─"
+# HIGH (CWE-59) regression: an attacker who swaps an operator file for a symlink
+# AFTER the durable snapshot must not cause the verify net's restore to write the
+# snapshot's secret out THROUGH that link. The shipped net drops a symlinked leaf and
+# writes a real file in its place, leaving the external target untouched.
+EXFIL_DIR=$(mktemp -d); EXFIL_TARGET="$EXFIL_DIR/stolen"
+printf 'ATTACKER-PLACEHOLDER\n' > "$EXFIL_TARGET"
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$INSTALL" make_symlink_leaf_shim)
+chk "symlink-leaf: upgrade succeeds"              "[ '$rc' -eq 0 ]"
+chk "symlink-leaf: secret NOT written through link" "! grep -q '$SECRET' '$EXFIL_TARGET'"
+chk "symlink-leaf: credentials.json is a real file" "[ -f '$H/tools/_lib/credentials.json' ] && [ ! -L '$H/tools/_lib/credentials.json' ]"
+chk "symlink-leaf: credentials.json restored intact" "grep -q '$SECRET' '$H/tools/_lib/credentials.json'"
+chk "symlink-leaf: secret not leaked to stdout/stderr" "! grep -q '$SECRET' '$OUT'"
+rm -rf "$STATE" "$H" "$EXFIL_DIR"; rm -f "$OUT"
+
+# Control: strip the SYMLINK-LEAF-GUARD → cp follows the swapped-in link and writes
+# the snapshot secret out through it (proves the guard is load-bearing).
+EXFIL_DIR=$(mktemp -d); EXFIL_TARGET="$EXFIL_DIR/stolen"
+printf 'ATTACKER-PLACEHOLDER\n' > "$EXFIL_TARGET"
+LEAFCTRL="$(mk_control 'SYMLINK-LEAF-GUARD' symlinkleaf-control)"
+IFS=$'\t' read -r rc OUT STATE H < <(run_snap "$LEAFCTRL" make_symlink_leaf_shim)
+chk "control: secret leaked through the symlink"  "grep -q '$SECRET' '$EXFIL_TARGET'"
+rm -rf "$STATE" "$H" "$EXFIL_DIR"; rm -f "$OUT"
+
+echo "── Part 8: snapshot umask 077 does not leak into synced files (+ control) ──"
+# SHOULD-FIX regression: umask 077 is process-global. Scoped to the snapshot it keeps
+# backups 0600; leaked past it, every later cp/mkdir inherits 0600/0700. A freshly-
+# synced framework file must be 0644 (per the ambient 022 umask) while the backup of
+# a secret stays 0600.
+IFS=$'\t' read -r rc OUT STATE H < <(umask 022; run_snap "$INSTALL")
+SNAP="$(snap_dir "$STATE")"
+chk "umask: upgrade succeeds"                     "[ '$rc' -eq 0 ]"
+chk "umask: synced framework file is 0644"        "[ \$(stat -c '%a' '$H/$POISON_REL') -eq 644 ]"
+chk "umask: backup of a secret stays 0600"        "[ \$(stat -c '%a' '$SNAP/tools/_lib/credentials.json') -eq 600 ]"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+# Control: strip the UMASK-RESTORE-NORMAL line → umask 077 leaks past the snapshot,
+# so the newly-synced framework file is created 0600 (proves the restore matters).
+UMASKCTRL="$(mk_control 'UMASK-RESTORE-NORMAL' umask-control)"
+IFS=$'\t' read -r rc OUT STATE H < <(umask 022; run_snap "$UMASKCTRL")
+chk "control: leaked umask makes synced file 0600" "[ \$(stat -c '%a' '$H/$POISON_REL') -eq 600 ]"
+rm -rf "$STATE" "$H"; rm -f "$OUT"
+
+echo ""
+echo "RESULT: $pass passed, $fail failed"
+[ "$fail" -eq 0 ]

packages/mosaic/framework/tools/quality/scripts/test-upgrade-manifest-guard.sh

@@ -29,6 +29,13 @@ INSTALL="$FW/install.sh"
 pass=0; fail=0
 chk() { if eval "$2"; then echo "  ✓ $1"; pass=$((pass + 1)); else echo "  ✗ $1"; fail=$((fail + 1)); fi; }
 
+# Redirect the #791 PR2 durable pre-update snapshot ($XDG_STATE_HOME/mosaic/backups)
+# into a throwaway so a keep-mode upgrade under test never writes into the real
+# ~/.local/state. This test asserts operator-surface fidelity, not backup content.
+export XDG_STATE_HOME
+XDG_STATE_HOME="$(mktemp -d)"
+trap 'rm -rf "$XDG_STATE_HOME"' EXIT
+
 SECRET='SUPER-SECRET-TOKEN-do-not-log-3f9a'
 
 # Seed a throwaway MOSAIC_HOME with an operator sentinel per ownership class.
@@ -206,7 +213,7 @@ fi
 #    does not silently depend on rsync being installed. (Provide the coreutils the
 #    installer needs on the stripped PATH.)
 FBIN=$(mktemp -d)
-for t in bash cp find mktemp rm mkdir chmod cmp sed grep cat dirname basename stat sha256sum awk tr; do
+for t in bash cp find mktemp rm mkdir chmod cmp sed grep cat dirname basename stat sha256sum awk tr date sort; do
   p=$(command -v "$t" 2>/dev/null) && ln -s "$p" "$FBIN/$t"
 done
 run_matrix "rsync-absent" env "PATH=$FBIN"

packages/mosaic/src/cli.ts

@@ -17,6 +17,7 @@ import { registerConfigCommand } from './commands/config.js';
 import { registerFleetCommand } from './commands/fleet.js';
 import { registerMissionCommand } from './commands/mission.js';
 import { registerUninstallCommand } from './commands/uninstall.js';
+import { registerRestoreCommand } from './commands/restore.js';
 // prdy is registered via launch.ts
 import { registerLaunchCommands } from './commands/launch.js';
 import { registerAuthCommand } from './commands/auth.js';
@@ -406,6 +407,10 @@ registerStorageCommand(program);
 
 registerUninstallCommand(program);
 
+// ─── restore ─────────────────────────────────────────────────────────────────
+
+registerRestoreCommand(program);
+
 // ─── telemetry ───────────────────────────────────────────────────────────────
 
 registerTelemetryCommand(program);

packages/mosaic/src/commands/restore.spec.ts

@@ -0,0 +1,466 @@
+/**
+ * Tests for `mosaic restore` (#791 PR2 Task 12).
+ *
+ * The durable pre-update snapshot (install.sh: make_durable_snapshot) writes the
+ * operator-owned surface to $XDG_STATE_HOME/mosaic/backups/pre-update-<ts>/ with
+ * 0700 dirs / 0600 files. `mosaic restore` is the recovery counterpart:
+ *   • --list (default)   enumerate snapshots by timestamp — dry-run, never mutates.
+ *   • --from <ts>        restore that snapshot over MOSAIC_HOME, confirmation-gated.
+ * It reports counts and relative paths ONLY — a snapshot may contain secrets
+ * (credentials.json), so no file content is ever printed (secrev invariant).
+ */
+
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import {
+  mkdtempSync,
+  rmSync,
+  mkdirSync,
+  writeFileSync,
+  readFileSync,
+  statSync,
+  lstatSync,
+  symlinkSync,
+} from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir, homedir } from 'node:os';
+import { Command } from 'commander';
+import {
+  resolveBackupRoot,
+  listSnapshots,
+  resolveSnapshotDir,
+  planRestore,
+  applyRestore,
+  runRestore,
+  registerRestoreCommand,
+} from './restore.js';
+
+const SECRET = 'SUPER-SECRET-TOKEN-do-not-log-restore';
+
+function seedSnapshot(root: string, ts: string, files: Record<string, string>): string {
+  const dir = join(root, `pre-update-${ts}`);
+  for (const [rel, content] of Object.entries(files)) {
+    const abs = join(dir, rel);
+    mkdirSync(join(abs, '..'), { recursive: true });
+    writeFileSync(abs, content);
+  }
+  return dir;
+}
+
+describe('mosaic restore (#791 PR2)', () => {
+  let tmp: string;
+  let backups: string;
+
+  beforeEach(() => {
+    tmp = mkdtempSync(join(tmpdir(), 'mosaic-restore-'));
+    backups = join(tmp, 'state', 'mosaic', 'backups');
+    mkdirSync(backups, { recursive: true });
+    vi.clearAllMocks();
+  });
+
+  afterEach(() => {
+    rmSync(tmp, { recursive: true, force: true });
+  });
+
+  describe('resolveBackupRoot', () => {
+    it('honors XDG_STATE_HOME', () => {
+      expect(resolveBackupRoot({ XDG_STATE_HOME: '/x/state' })).toBe('/x/state/mosaic/backups');
+    });
+    it('falls back to ~/.local/state', () => {
+      expect(resolveBackupRoot({})).toBe(join(homedir(), '.local', 'state', 'mosaic', 'backups'));
+    });
+  });
+
+  describe('listSnapshots', () => {
+    it('returns [] when the backup root does not exist', () => {
+      expect(listSnapshots(join(tmp, 'nope'))).toEqual([]);
+    });
+
+    it('lists pre-update snapshots newest-first with file counts, ignoring other dirs', () => {
+      seedSnapshot(backups, '20240101T000000Z', { 'SOUL.md': 'a' });
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'b', 'agents/x.conf': 'c' });
+      mkdirSync(join(backups, 'unrelated-dir'), { recursive: true });
+
+      const snaps = listSnapshots(backups);
+      expect(snaps.map((s) => s.timestamp)).toEqual(['20260101T000000Z', '20240101T000000Z']);
+      expect(snaps[0]!.fileCount).toBe(2);
+      expect(snaps[1]!.fileCount).toBe(1);
+    });
+  });
+
+  describe('resolveSnapshotDir', () => {
+    it('resolves by bare timestamp and by full pre-update-<ts> name', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'a' });
+      expect(resolveSnapshotDir(backups, '20260101T000000Z')).toBe(dir);
+      expect(resolveSnapshotDir(backups, 'pre-update-20260101T000000Z')).toBe(dir);
+    });
+    it('returns undefined for an unknown timestamp', () => {
+      expect(resolveSnapshotDir(backups, '19990101T000000Z')).toBeUndefined();
+    });
+  });
+
+  describe('planRestore', () => {
+    it('walks nested dirs and returns every relative file path', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', {
+        'SOUL.md': 'a',
+        'agents/x.conf': 'b',
+        'tools/_lib/credentials.json': 'c',
+      });
+      expect(planRestore(dir).sort()).toEqual(
+        ['SOUL.md', 'agents/x.conf', 'tools/_lib/credentials.json'].sort(),
+      );
+    });
+  });
+
+  describe('applyRestore', () => {
+    it('restores byte-exact content, creates parent dirs, and sets 0600', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', {
+        'SOUL.md': 'original-soul',
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      // A diverged operator file that restore must overwrite.
+      writeFileSync(join(home, 'SOUL.md'), 'CORRUPTED');
+
+      const n = applyRestore(dir, home, planRestore(dir));
+      expect(n).toBe(2);
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('original-soul');
+      expect(readFileSync(join(home, 'tools/_lib/credentials.json'), 'utf8')).toBe(
+        `TOKEN=${SECRET}\n`,
+      );
+      expect(statSync(join(home, 'SOUL.md')).mode & 0o777).toBe(0o600);
+      expect(statSync(join(home, 'tools/_lib/credentials.json')).mode & 0o777).toBe(0o600);
+    });
+  });
+
+  describe('runRestore', () => {
+    it('--list prints timestamps and counts, mutating nothing', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'a', 'agents/x.conf': 'b' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+
+      const code = await runRestore({
+        list: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(0);
+      const out = log.mock.calls.flat().join('\n');
+      expect(out).toContain('20260101T000000Z');
+      expect(out).toMatch(/2\b/); // the file count is surfaced
+      log.mockRestore();
+    });
+
+    it('--from restores the snapshot over MOSAIC_HOME byte-exact (yes bypasses prompt)', async () => {
+      seedSnapshot(backups, '20260101T000000Z', {
+        'SOUL.md': 'restored-soul',
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'STALE');
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        yes: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(0);
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('restored-soul');
+      expect(readFileSync(join(home, 'tools/_lib/credentials.json'), 'utf8')).toBe(
+        `TOKEN=${SECRET}\n`,
+      );
+    });
+
+    it('--from with an unknown timestamp fails without mutating', async () => {
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'KEEP');
+      const err = vi.spyOn(console, 'error').mockImplementation(() => {});
+
+      const code = await runRestore({
+        from: '19990101T000000Z',
+        yes: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(1);
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('KEEP');
+      err.mockRestore();
+    });
+
+    it('--dry-run with --from reports the plan but mutates nothing', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'snap' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'UNCHANGED');
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        dryRun: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(0);
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('UNCHANGED');
+      log.mockRestore();
+    });
+
+    it('--from prompts and applies the restore when the operator confirms', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'confirmed-soul' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'STALE');
+      vi.spyOn(console, 'log').mockImplementation(() => {});
+      const confirm = vi.fn().mockResolvedValue(true);
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+        confirm,
+      });
+
+      expect(code).toBe(0);
+      expect(confirm).toHaveBeenCalledOnce();
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('confirmed-soul');
+    });
+
+    it('--from aborts without mutating when the operator declines', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'snap' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'KEEP');
+      vi.spyOn(console, 'log').mockImplementation(() => {});
+      const confirm = vi.fn().mockResolvedValue(false);
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+        confirm,
+      });
+
+      expect(code).toBe(0);
+      expect(confirm).toHaveBeenCalledOnce();
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('KEEP');
+    });
+
+    it('MOSAIC_ASSUME_YES=1 bypasses the confirmation prompt', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'env-yes' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'STALE');
+      vi.spyOn(console, 'log').mockImplementation(() => {});
+      const confirm = vi.fn().mockResolvedValue(false);
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state'), MOSAIC_ASSUME_YES: '1' },
+        confirm,
+      });
+
+      expect(code).toBe(0);
+      expect(confirm).not.toHaveBeenCalled();
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('env-yes');
+    });
+
+    it('--list reports gracefully when no snapshots exist', async () => {
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+
+      const code = await runRestore({
+        list: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'empty-state') },
+      });
+
+      expect(code).toBe(0);
+      expect(log.mock.calls.flat().join('\n')).toMatch(/No pre-update snapshots/);
+    });
+
+    it('never prints a secret value found inside a backed-up file', async () => {
+      seedSnapshot(backups, '20260101T000000Z', {
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+      const err = vi.spyOn(console, 'error').mockImplementation(() => {});
+
+      await runRestore({
+        list: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+      await runRestore({
+        from: '20260101T000000Z',
+        yes: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      const all = [...log.mock.calls, ...err.mock.calls].flat().join('\n');
+      expect(all).not.toContain(SECRET);
+      log.mockRestore();
+      err.mockRestore();
+    });
+  });
+
+  // Regression coverage for the codex code+security review of PR2 (#791):
+  // CWE-22 traversal via --from, and CWE-59 symlink write-through in applyRestore.
+  describe('security hardening', () => {
+    it.each([
+      '../../etc',
+      'pre-update-/../../tmp/poison',
+      'pre-update-../evil',
+      '20260101T000000Z/../../../tmp',
+      'not-a-timestamp',
+      '2026-01-01',
+    ])('resolveSnapshotDir rejects traversal / malformed selector %j', (bad) => {
+      // Even if a matching directory exists on disk, a non-timestamp selector
+      // must not resolve — the only accepted shape is <8>T<6>Z[-n].
+      expect(resolveSnapshotDir(backups, bad)).toBeUndefined();
+    });
+
+    it('runRestore --from a traversal selector fails closed without copying', async () => {
+      // Plant a real dir one level ABOVE the backup root. The naive resolver
+      // `join(root, from)` with `from='../poison'` would reach it (backups is
+      // .../mosaic/backups, so `../poison` == .../mosaic/poison) and import it.
+      const outside = join(tmp, 'state', 'mosaic', 'poison');
+      mkdirSync(outside, { recursive: true });
+      writeFileSync(join(outside, 'x'), 'attacker');
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      const err = vi.spyOn(console, 'error').mockImplementation(() => {});
+
+      const code = await runRestore({
+        from: '../poison',
+        yes: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(1);
+      expect(statSync(home).isDirectory()).toBe(true);
+      // Nothing from `outside` was imported.
+      expect(() => statSync(join(home, 'x'))).toThrow();
+      err.mockRestore();
+    });
+
+    it('applyRestore refuses to write a secret through a symlinked leaf (CWE-59)', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', {
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(join(home, 'tools', '_lib'), { recursive: true });
+      // Attacker points the operator credentials file at a file they can read.
+      const exfil = join(tmp, 'exfil-target');
+      writeFileSync(exfil, 'original-attacker-content');
+      symlinkSync(exfil, join(home, 'tools', '_lib', 'credentials.json'));
+
+      expect(() => applyRestore(dir, home, planRestore(dir))).toThrow();
+      // The secret was NOT written through the link into the attacker's file.
+      expect(readFileSync(exfil, 'utf8')).toBe('original-attacker-content');
+    });
+
+    it('applyRestore refuses to write through a symlinked ancestor (CWE-59)', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', {
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(join(home, 'tools'), { recursive: true });
+      // Attacker replaces the `tools/_lib` ancestor with a symlink out of the root.
+      const exfilDir = join(tmp, 'exfil-dir');
+      mkdirSync(exfilDir, { recursive: true });
+      symlinkSync(exfilDir, join(home, 'tools', '_lib'));
+
+      expect(() => applyRestore(dir, home, planRestore(dir))).toThrow();
+      // Nothing was written into the attacker-controlled directory.
+      expect(() => statSync(join(exfilDir, 'credentials.json'))).toThrow();
+    });
+
+    it('runRestore surfaces a symlink violation as exit 1 without leaking the secret', async () => {
+      seedSnapshot(backups, '20260101T000000Z', {
+        'tools/_lib/credentials.json': `TOKEN=${SECRET}\n`,
+      });
+      const home = join(tmp, 'home');
+      mkdirSync(join(home, 'tools', '_lib'), { recursive: true });
+      const exfil = join(tmp, 'exfil-target');
+      writeFileSync(exfil, 'attacker');
+      symlinkSync(exfil, join(home, 'tools', '_lib', 'credentials.json'));
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+      const err = vi.spyOn(console, 'error').mockImplementation(() => {});
+
+      const code = await runRestore({
+        from: '20260101T000000Z',
+        yes: true,
+        mosaicHome: home,
+        env: { XDG_STATE_HOME: join(tmp, 'state') },
+      });
+
+      expect(code).toBe(1);
+      expect(readFileSync(exfil, 'utf8')).toBe('attacker');
+      const all = [...log.mock.calls, ...err.mock.calls].flat().join('\n');
+      expect(all).not.toContain(SECRET);
+      log.mockRestore();
+      err.mockRestore();
+    });
+
+    it('applyRestore replaces a diverged regular file in place with 0600 (not a symlink)', () => {
+      const dir = seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'restored' });
+      const home = join(tmp, 'home');
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, 'SOUL.md'), 'stale');
+
+      const n = applyRestore(dir, home, planRestore(dir));
+      expect(n).toBe(1);
+      expect(lstatSync(join(home, 'SOUL.md')).isSymbolicLink()).toBe(false);
+      expect(readFileSync(join(home, 'SOUL.md'), 'utf8')).toBe('restored');
+      expect(statSync(join(home, 'SOUL.md')).mode & 0o777).toBe(0o600);
+    });
+  });
+
+  describe('registerRestoreCommand', () => {
+    it('registers `restore` with the expected flags', () => {
+      const program = new Command();
+      program.exitOverride();
+      registerRestoreCommand(program);
+      const cmd = program.commands.find((c) => c.name() === 'restore');
+      expect(cmd).toBeDefined();
+      const longs = cmd!.options.map((o) => o.long);
+      expect(longs).toEqual(
+        expect.arrayContaining(['--list', '--from', '--dry-run', '--yes', '--mosaic-home']),
+      );
+    });
+
+    it('runs the list action end-to-end via the parsed command', async () => {
+      seedSnapshot(backups, '20260101T000000Z', { 'SOUL.md': 'a' });
+      const prevXdg = process.env['XDG_STATE_HOME'];
+      process.env['XDG_STATE_HOME'] = join(tmp, 'state');
+      const log = vi.spyOn(console, 'log').mockImplementation(() => {});
+      try {
+        const program = new Command();
+        program.exitOverride();
+        registerRestoreCommand(program);
+        await program.parseAsync(['restore', '--list', '--mosaic-home', join(tmp, 'home')], {
+          from: 'user',
+        });
+        expect(log.mock.calls.flat().join('\n')).toContain('20260101T000000Z');
+      } finally {
+        log.mockRestore();
+        if (prevXdg === undefined) delete process.env['XDG_STATE_HOME'];
+        else process.env['XDG_STATE_HOME'] = prevXdg;
+      }
+    });
+  });
+});

packages/mosaic/src/commands/restore.ts

@@ -0,0 +1,313 @@
+/**
+ * restore.ts — top-level `mosaic restore` command (#791 PR2 Task 12)
+ *
+ * Recovery counterpart to the durable pre-update snapshot taken by install.sh
+ * (make_durable_snapshot). Before a keep-mode upgrade mutates anything, the
+ * installer copies the operator-owned surface to
+ *   $XDG_STATE_HOME/mosaic/backups/pre-update-<UTC-ts>/   (0700 dirs / 0600 files)
+ * This command lets the operator inspect and roll back to those snapshots:
+ *
+ *   mosaic restore                 # == --list: enumerate snapshots (dry-run)
+ *   mosaic restore --list
+ *   mosaic restore --from <ts>     # restore that snapshot over MOSAIC_HOME
+ *   mosaic restore --from <ts> --dry-run
+ *
+ * SECREV INVARIANT: a snapshot may contain secrets (e.g. tools/_lib/credentials.json).
+ * This command reports counts and RELATIVE PATHS only — it never reads a backed-up
+ * file into any logged string. Restored files are written back 0600 (owner-only),
+ * matching the snapshot's own private posture. The path convention here mirrors
+ * install.sh `backup_root()`; keep the two in sync (no shared code across the boundary).
+ */
+
+import {
+  existsSync,
+  readdirSync,
+  statSync,
+  lstatSync,
+  readFileSync,
+  openSync,
+  writeSync,
+  fchmodSync,
+  closeSync,
+  constants,
+} from 'node:fs';
+import { createInterface } from 'node:readline';
+import { homedir } from 'node:os';
+import { join, dirname, relative } from 'node:path';
+import type { Command } from 'commander';
+import { DEFAULT_MOSAIC_HOME } from '../constants.js';
+import { assertCanonicalContainment, ensureManagedDirectory } from '../fleet/secure-file.js';
+
+// ─── types ───────────────────────────────────────────────────────────────────
+
+export interface SnapshotInfo {
+  /** The UTC stamp after the `pre-update-` prefix, e.g. "20260716T232225Z". */
+  readonly timestamp: string;
+  /** Absolute path to the snapshot directory. */
+  readonly dir: string;
+  /** Number of files captured in the snapshot. */
+  readonly fileCount: number;
+}
+
+export interface RestoreOptions {
+  list?: boolean;
+  from?: string;
+  dryRun?: boolean;
+  yes?: boolean;
+  mosaicHome: string;
+  /** Environment source (injectable for tests); defaults to process.env. */
+  env?: NodeJS.ProcessEnv;
+  /**
+   * Confirmation gate (injectable for tests); defaults to an interactive
+   * readline prompt. Returns true to proceed with the overwrite.
+   */
+  confirm?: (question: string) => Promise<boolean>;
+}
+
+const SNAPSHOT_PREFIX = 'pre-update-';
+
+/**
+ * The exact shape install.sh `make_durable_snapshot()` stamps: `<8>T<6>Z` UTC,
+ * with an optional `-<n>` same-second collision suffix. `--from` is matched
+ * against this — nothing containing a path separator or `..` can pass, so a
+ * selector can never escape the backup root (CWE-22).
+ */
+const SNAPSHOT_TS_RE = /^\d{8}T\d{6}Z(?:-\d+)?$/;
+
+// ─── pure helpers ─────────────────────────────────────────────────────────────
+
+/** Resolve the durable-snapshot root, mirroring install.sh `backup_root()`. */
+export function resolveBackupRoot(env: NodeJS.ProcessEnv = process.env): string {
+  const stateHome = env['XDG_STATE_HOME'] || join(homedir(), '.local', 'state');
+  return join(stateHome, 'mosaic', 'backups');
+}
+
+/** Recursively collect every file under `dir` as a path relative to `dir`. */
+export function planRestore(dir: string): string[] {
+  const out: string[] = [];
+  const walk = (cur: string): void => {
+    for (const entry of readdirSync(cur, { withFileTypes: true })) {
+      const abs = join(cur, entry.name);
+      if (entry.isDirectory()) {
+        walk(abs);
+      } else if (entry.isFile()) {
+        out.push(relative(dir, abs));
+      }
+    }
+  };
+  if (existsSync(dir)) walk(dir);
+  return out;
+}
+
+/** Enumerate snapshots newest-first (the `pre-update-<ts>` names sort chronologically). */
+export function listSnapshots(root: string): SnapshotInfo[] {
+  if (!existsSync(root)) return [];
+  let entries: string[];
+  try {
+    entries = readdirSync(root);
+  } catch {
+    return [];
+  }
+  return entries
+    .filter((name) => name.startsWith(SNAPSHOT_PREFIX))
+    .map((name) => join(root, name))
+    .filter((dir) => {
+      try {
+        return statSync(dir).isDirectory();
+      } catch {
+        return false;
+      }
+    })
+    .sort()
+    .reverse()
+    .map((dir) => ({
+      timestamp: dir.split('/').at(-1)!.slice(SNAPSHOT_PREFIX.length),
+      dir,
+      fileCount: planRestore(dir).length,
+    }));
+}
+
+/**
+ * Resolve a snapshot dir from a `--from` selector. Accepts ONLY a strict
+ * generated identifier — a bare `<ts>` or the full `pre-update-<ts>` name — and
+ * builds exactly `join(root, 'pre-update-' + ts)`. A selector containing `/`,
+ * `..`, or anything but the timestamp shape is rejected (returns undefined), so
+ * `--from` can never traverse outside the backup root (CWE-22). The resolved dir
+ * must be a real, non-symlink directory (lstat, not stat), so a symlinked
+ * snapshot entry can't redirect the restore either.
+ */
+export function resolveSnapshotDir(root: string, from: string): string | undefined {
+  const ts = from.startsWith(SNAPSHOT_PREFIX) ? from.slice(SNAPSHOT_PREFIX.length) : from;
+  if (!SNAPSHOT_TS_RE.test(ts)) return undefined;
+  const dir = join(root, `${SNAPSHOT_PREFIX}${ts}`);
+  try {
+    if (lstatSync(dir).isDirectory()) return dir;
+  } catch {
+    /* absent or inaccessible */
+  }
+  return undefined;
+}
+
+/**
+ * Copy each `relPaths` entry from the snapshot back into `mosaicHome`, forcing
+ * 0600 on the restored file (owner-only — the operator surface may hold secrets).
+ * Returns the number of files restored. Never reads a file's content into a
+ * logged string.
+ *
+ * SYMLINK-SAFE (CWE-59): a snapshot may hold secrets, so we must never let a
+ * tampered destination redirect the write. Every destination path is contained
+ * within `mosaicHome` (assertCanonicalContainment) and every ancestor is proven
+ * to be a real, non-symlink directory (ensureManagedDirectory) before we write.
+ * The leaf itself is opened O_NOFOLLOW, so if it was swapped for a symlink the
+ * open fails closed (ELOOP) rather than writing the secret through the link.
+ */
+export function applyRestore(
+  snapDir: string,
+  mosaicHome: string,
+  relPaths: readonly string[],
+): number {
+  let restored = 0;
+  for (const rel of relPaths) {
+    const src = join(snapDir, rel);
+    const dst = join(mosaicHome, rel);
+    // Fail closed if the target path escapes the managed root or any ancestor is
+    // a symlink; create missing ancestors as private (0700) real directories.
+    assertCanonicalContainment(mosaicHome, dst);
+    ensureManagedDirectory(mosaicHome, dirname(dst));
+    // O_NOFOLLOW: refuse to follow a symlink at the leaf (secret exfil guard).
+    // O_CREAT|O_TRUNC: create a fresh 0600 file, or overwrite a diverged real one.
+    const fd = openSync(
+      dst,
+      constants.O_WRONLY | constants.O_CREAT | constants.O_TRUNC | constants.O_NOFOLLOW,
+      0o600,
+    );
+    try {
+      fchmodSync(fd, 0o600); // enforce 0600 even when the file pre-existed
+      writeSync(fd, readFileSync(src));
+    } finally {
+      closeSync(fd);
+    }
+    restored += 1;
+  }
+  return restored;
+}
+
+// ─── orchestration ────────────────────────────────────────────────────────────
+
+async function promptConfirm(question: string): Promise<boolean> {
+  const rl = createInterface({ input: process.stdin, output: process.stdout });
+  try {
+    return await new Promise<boolean>((resolve) => {
+      rl.question(`${question} [y/N] `, (ans) => resolve(ans.trim().toLowerCase() === 'y'));
+    });
+  } finally {
+    rl.close();
+  }
+}
+
+/**
+ * Run `mosaic restore`. Returns a process exit code (0 ok, 1 error) rather than
+ * calling process.exit, so it stays unit-testable.
+ */
+export async function runRestore(opts: RestoreOptions): Promise<number> {
+  const env = opts.env ?? process.env;
+  const root = resolveBackupRoot(env);
+
+  // Default action (and explicit --list): enumerate, never mutate.
+  if (opts.list || !opts.from) {
+    const snaps = listSnapshots(root);
+    if (snaps.length === 0) {
+      console.log(`No pre-update snapshots found under ${root}.`);
+      return 0;
+    }
+    console.log(`Pre-update snapshots under ${root} (newest first):\n`);
+    for (const s of snaps) {
+      console.log(`  ${s.timestamp}  —  ${s.fileCount} file(s)`);
+    }
+    console.log(`\nRestore one with:  mosaic restore --from <timestamp>`);
+    return 0;
+  }
+
+  // --from <ts>: restore over the operator surface.
+  const snapDir = resolveSnapshotDir(root, opts.from);
+  if (!snapDir) {
+    console.error(`No snapshot matching '${opts.from}' under ${root}.`);
+    console.error(`Run 'mosaic restore --list' to see available timestamps.`);
+    return 1;
+  }
+
+  const relPaths = planRestore(snapDir);
+  const ts = snapDir.split('/').at(-1)!.slice(SNAPSHOT_PREFIX.length);
+
+  if (opts.dryRun) {
+    console.log(
+      `[dry-run] Would restore ${relPaths.length} file(s) from snapshot ${ts} into ${opts.mosaicHome}:`,
+    );
+    for (const rel of relPaths) console.log(`  ${rel}`);
+    console.log('[dry-run] No changes made.');
+    return 0;
+  }
+
+  const assumeYes = opts.yes || env['MOSAIC_ASSUME_YES'] === '1';
+  if (!assumeYes) {
+    console.log(
+      `About to restore ${relPaths.length} operator file(s) from snapshot ${ts} into ${opts.mosaicHome}.`,
+    );
+    console.log('This OVERWRITES those files with their pre-update contents.');
+    const ok = await (opts.confirm ?? promptConfirm)('Proceed?');
+    if (!ok) {
+      console.log('Restore cancelled. No changes made.');
+      return 0;
+    }
+  }
+
+  let n: number;
+  try {
+    n = applyRestore(snapDir, opts.mosaicHome, relPaths);
+  } catch (err) {
+    // A containment/symlink violation is a fail-closed security stop, not a
+    // routine error — surface it without leaking file contents and abort.
+    console.error(
+      `Restore aborted: a destination path under ${opts.mosaicHome} is unsafe to write ` +
+        `(symlink or escapes the managed root). No files were restored. (${(err as Error).message})`,
+    );
+    return 1;
+  }
+  console.log(`Restored ${n} operator file(s) from snapshot ${ts} into ${opts.mosaicHome}.`);
+  return 0;
+}
+
+// ─── commander registration ───────────────────────────────────────────────────
+
+export function registerRestoreCommand(program: Command): void {
+  program
+    .command('restore')
+    .description('List or restore durable pre-update snapshots of your operator config (#791)')
+    .option('--list', 'List available snapshots by timestamp (default action)')
+    .option('--from <timestamp>', 'Restore the snapshot with this timestamp over MOSAIC_HOME')
+    .option('--dry-run', 'With --from: show what would be restored without changing anything')
+    .option('--yes, -y', 'Skip the confirmation prompt (also: MOSAIC_ASSUME_YES=1)')
+    .option(
+      '--mosaic-home <path>',
+      'Override MOSAIC_HOME directory',
+      process.env['MOSAIC_HOME'] ?? DEFAULT_MOSAIC_HOME,
+    )
+    .action(
+      async (opts: {
+        list?: boolean;
+        from?: string;
+        dryRun?: boolean;
+        yes?: boolean;
+        mosaicHome: string;
+      }) => {
+        const code = await runRestore({
+          list: opts.list,
+          from: opts.from,
+          dryRun: opts.dryRun,
+          yes: opts.yes,
+          mosaicHome: opts.mosaicHome,
+        });
+        if (code !== 0) process.exit(code);
+      },
+    );
+}