fix: clean up partial refresh profiler overrides

tests/README.md

@@ -34,8 +34,11 @@ for five minutes around the apply; the changed ceiling must still appear within
 Control-phase requests pass the temporary bearer token through curl's stdin configuration,
 keeping it out of process arguments. Each POST uses its own securely allocated response file,
 removed after reading the change identifier.
+The drop-in is written in a private staging file and published atomically with `link` only after the
+write succeeds. A partial write fails setup and cleanup removes the staging file. Publication refuses
+an existing or competing destination, including symlinks, without claiming or removing it.
 Cleanup quiesces the timer and instrumented pass, emits only measurement records from the journal, removes
-the drop-in, reloads systemd and restores the timer’s prior active state before config/runtime restoration.
+the owned drop-in, reloads systemd and restores the timer’s prior active state before config/runtime restoration.
 The job fails if journal collection or measurement export fails, no profiled refresh completes
 successfully, or cleanup fails. Begin events from unfinished helpers are retained to locate stalled
 operations, including a pass stopped by cleanup.

tests/e2e-refresh-window.sh

@@ -1,6 +1,6 @@
 # shellcheck shell=bash
 # Reserved control-phase lifecycle; overrides only ExecStart, retaining the real service policy.
-REFRESH_PROFILE_DROPIN="" REFRESH_PROFILE_SINCE="" REFRESH_PROFILE_TIMER_ACTIVE=""
+REFRESH_PROFILE_STAGING="" REFRESH_PROFILE_DROPIN="" REFRESH_PROFILE_SINCE="" REFRESH_PROFILE_TIMER_ACTIVE=""
 refresh_profile_start() {
     local dir="${REFRESH_PROFILE_UNIT_ROOT:-/run/systemd/system}/rigforge-api-refresh.service.d"
     local script
@@ -13,12 +13,14 @@ refresh_profile_start() {
     systemctl is-active --quiet rigforge-api-refresh.timer && REFRESH_PROFILE_TIMER_ACTIVE=1
     REFRESH_PROFILE_SINCE=$(date -u '+%Y-%m-%d %H:%M:%S')
     local dropin="$dir/rigforge-profile.conf"
-    (
-        umask 077
-        set -o noclobber
-        printf '[Service]\nExecStart=\nExecStart=/bin/bash %s %s\n' "$script" "$RIGFORGE" >"$dropin"
-    ) || return 1
+    # Stage privately: failed writes must never expose an incomplete service policy.
+    REFRESH_PROFILE_STAGING=$(mktemp "$dir/.rigforge-profile.XXXXXX") || return 1
+    printf '[Service]\nExecStart=\nExecStart=/bin/bash %s %s\n' "$script" "$RIGFORGE" >|"$REFRESH_PROFILE_STAGING" || return 1
+    # POSIX link uses the exact destination (unlike ln's directory handling) and never replaces it.
+    link "$REFRESH_PROFILE_STAGING" "$dropin" || return 1
     REFRESH_PROFILE_DROPIN="$dropin"
+    rm -f "$REFRESH_PROFILE_STAGING" || return 1
+    REFRESH_PROFILE_STAGING=""
     systemctl daemon-reload
 }
 refresh_profile_state() {
@@ -29,8 +31,11 @@ refresh_profile_state() {
         -p ExecMainStartTimestamp -p ExecMainExitTimestamp -p ExecMainStatus
 }
 refresh_profile_finish() {
-    [ -n "$REFRESH_PROFILE_DROPIN" ] || return 0
     local log rc=0
+    if [ -n "$REFRESH_PROFILE_STAGING" ]; then
+        if rm -f "$REFRESH_PROFILE_STAGING"; then REFRESH_PROFILE_STAGING=""; else rc=1; fi
+    fi
+    [ -n "$REFRESH_PROFILE_DROPIN" ] || return "$rc"
     refresh_profile_state || rc=1
     # Quiesce both units before collecting: the timer could otherwise dispatch during cleanup.
     systemctl stop rigforge-api-refresh.timer || rc=1

tests/test-refresh-profile.sh

@@ -97,7 +97,90 @@ printf 'existing override\n' >"$owned"
 if refresh_profile_start; then exit 1; fi
 refresh_profile_finish
 check test "$(cat "$owned")" = 'existing override'
-rm "$owned"
+command rm "$owned"
+# A failed partial write must remain owned until cleanup removes it.
+saved_umask=$(umask)
+umask 027
+exec 9>"$CASE_DIR/caller-fd"
+printf() {
+    if [[ "$1" == '[Service]'* ]]; then
+        builtin printf '[Service]\nExecStart=\n'
+        return 7
+    fi
+    builtin printf "$@"
+}
+if refresh_profile_start; then exit 1; fi
+unset -f printf
+check test "$(umask)" = 0027
+case $- in *C*) exit 1 ;; esac
+builtin printf 'caller descriptor preserved\n' >&9
+exec 9>&-
+check test "$(cat "$CASE_DIR/caller-fd")" = 'caller descriptor preserved'
+umask "$saved_umask"
+check test -z "$REFRESH_PROFILE_DROPIN"
+check test ! -e "$owned"
+staged="$REFRESH_PROFILE_STAGING"
+check test "$(cat "$staged")" = "$(builtin printf '[Service]\nExecStart=')"
+rm() { return 7; }
+if refresh_profile_finish; then exit 1; fi
+check test "$REFRESH_PROFILE_STAGING" = "$staged"
+check test -e "$staged"
+unset -f rm
+refresh_profile_finish >"$CASE_DIR/partial-write-collected"
+check test ! -e "$owned"
+check test ! -e "$staged"
+check test -z "$REFRESH_PROFILE_STAGING"
+check test -z "$REFRESH_PROFILE_DROPIN"
+# A competitor arriving after the initial existence check must not be claimed or removed.
+mkdir() {
+    command mkdir "$@" || return
+    builtin printf 'competing override\n' >"$owned"
+}
+set -o noclobber
+if refresh_profile_start 2>"$CASE_DIR/collision-error"; then exit 1; fi
+unset -f mkdir
+case $- in *C*) : ;; *) exit 1 ;; esac
+set +o noclobber
+check test "$(umask)" = "$saved_umask"
+check test -z "$REFRESH_PROFILE_DROPIN"
+refresh_profile_finish
+check test "$(cat "$owned")" = 'competing override'
+command rm "$owned"
+# A raced symlink to a nonregular target must also remain untouched.
+mkdir() {
+    command mkdir "$@" || return
+    ln -s /dev/null "$owned"
+}
+if refresh_profile_start 2>"$CASE_DIR/symlink-collision-error"; then exit 1; fi
+unset -f mkdir
+check test -z "$REFRESH_PROFILE_DROPIN"
+refresh_profile_finish
+check test -L "$owned"
+check test "$(readlink "$owned")" = /dev/null
+command rm "$owned"
+# Staging allocation failure must fail setup without claiming any override.
+mktemp() { return 7; }
+if refresh_profile_start; then exit 1; fi
+unset -f mktemp
+check test -z "$REFRESH_PROFILE_STAGING"
+check test -z "$REFRESH_PROFILE_DROPIN"
+refresh_profile_finish
+check test ! -e "$owned"
+# Failure to unlink staging after publication must not prevent live-override cleanup.
+rm() {
+    if [ "${2:-}" = "$REFRESH_PROFILE_STAGING" ]; then return 7; fi
+    command rm "$@"
+}
+if refresh_profile_start; then exit 1; fi
+staged="$REFRESH_PROFILE_STAGING"
+check test -n "$REFRESH_PROFILE_DROPIN"
+if refresh_profile_finish >"$CASE_DIR/staging-unlink-failed"; then exit 1; fi
+check test ! -e "$owned"
+check test -e "$staged"
+unset -f rm
+refresh_profile_finish
+check test ! -e "$staged"
+check test -z "$REFRESH_PROFILE_STAGING"
 # A window without a completed profiled refresh cannot report successful diagnostics.
 refresh_profile_start
 journalctl() { :; }