packages/mosaic/framework/tools/wake/test-wake-store-ack.sh

@@ -306,19 +306,57 @@ echo "== T9: burn-before-enqueue — a failed durable write must NOT advance obs
   [ "$s1" = "1" ] || fail_msg "T9: baseline allocation should be 1, got '$s1'"
   obs_before="$("$STORE" cursors | sed -n 's/^observed_seq=//p')"
   [ "$obs_before" = "1" ] || fail_msg "T9: observed_seq should be 1 before the forced failure, got '$obs_before'"
-  # Force the durable pending write to FAIL: make STATE_DIR read-only so the
-  # atomic-write mktemp inside it cannot create its temp file => the durable
-  # pending write fails and the enqueue must abort. The already-existing cursor
-  # FILE stays owner-writable (a read-only DIR blocks new-entry creation, not a
-  # truncating write to an existing file), so a regression that committed the
-  # observed_seq cursor BEFORE/independent of the durable write (the arrow-#1
-  # burn) would still advance it here — the RED signal this test catches. The
-  # lock file + cursor files already exist, so lock acquisition + cursor reads
-  # still work; only the durable pending WRITE fails.
-  chmod 500 "$STATE_DIR"
+  # Force the durable PENDING write to FAIL via a PRIVILEGE-INVARIANT trigger:
+  # bind-mount a throwaway file OVER $STATE_DIR/pending.jsonl (_atomic_write's
+  # exact rename TARGET for the durable write — see _wake-common.sh) so the
+  # atomic write's `mv -f tmp target` hits EBUSY ("Device or resource busy"):
+  # the kernel refuses to rename ANY file onto an ACTIVE MOUNT POINT. That is
+  # a structural VFS constraint, not a DAC permission check — root does NOT
+  # bypass it (no uid can rename over a live mountpoint short of an explicit
+  # umount first). This replaces the old `chmod 500 "$STATE_DIR"` injection: a
+  # CI container running as root BYPASSES DAC checks entirely, so the
+  # "read-only" dir let mktemp/the write through and every assertion below
+  # went spuriously green. The old chmod also blocked the CURSOR's own
+  # mktemp (same dir), so it could never have caught a premature cursor
+  # commit either — narrower coverage than its comment claimed. Here
+  # observed_seq is left an ordinary writable file, never touched by the
+  # mount, so a regression that commits the cursor before/independent of the
+  # durable write is still genuinely caught by the assertions below.
+  #
+  # The mount + forced enqueue run inside a private mount+user namespace
+  # (`unshare --mount --user --map-root-user`) so the mechanism is IDENTICAL
+  # whether this harness runs unprivileged (dev) or as root (CI): creating a
+  # user namespace needs no pre-existing privilege, and the busy-mountpoint
+  # rule inside it is the real kernel rule, unaffected by uid in or out of a
+  # namespace. The bind mount is entirely private to that namespace and
+  # vanishes the instant it exits, so $STATE_DIR/pending.jsonl is a plain
+  # writable file again for every assertion below — no manual cleanup step.
   rc=0
-  "$STORE" enqueue --class actionable --locators '{"n":2}' >/dev/null 2>&1 || rc=$?
-  chmod 700 "$STATE_DIR"
+  if ! command -v unshare >/dev/null 2>&1; then
+    fail_msg "T9: 'unshare' unavailable — cannot construct the privilege-invariant fault injection in this sandbox"
+    rc=99
+  else
+    ro_src="$TMP_ROOT/t9-ro-src"
+    mkdir -p "$ro_src"
+    : >"$ro_src/pending.jsonl"
+    # shellcheck disable=SC2016 # deliberate: $1/$2/$3 expand in the NESTED
+    # bash (inside the unshare'd namespace), not this outer single-quoted one.
+    unshare --mount --user --map-root-user bash -c '
+      set -u
+      ro_src="$1"
+      state_dir="$2"
+      store="$3"
+      if ! mount --bind "$ro_src/pending.jsonl" "$state_dir/pending.jsonl" 2>/dev/null; then
+        exit 98
+      fi
+      mount -o remount,ro,bind "$state_dir/pending.jsonl" 2>/dev/null || true
+      "$store" enqueue --class actionable --locators "{\"n\":2}" >/dev/null 2>&1
+    ' _ "$ro_src" "$STATE_DIR" "$STORE"
+    rc=$?
+    if [ "$rc" -eq 98 ]; then
+      fail_msg "T9: could not bind-mount over pending.jsonl (mount unavailable in this sandbox) — fault injection not constructed"
+    fi
+  fi
   [ "$rc" -ne 0 ] || fail_msg "T9: an enqueue whose durable write fails must EXIT NON-ZERO (fail loud)"
   # THE ARROW-#1 ASSERTION: the observed_seq cursor did NOT advance — no seq was
   # burned by a write that never reached the store. (RED if the cursor bumps