diff --git a/docs/changes/2026-07-08-backup-restore-mutation-audit.md b/docs/changes/2026-07-08-backup-restore-mutation-audit.md
new file mode 100644
index 0000000..b518749
--- /dev/null
+++ b/docs/changes/2026-07-08-backup-restore-mutation-audit.md
@@ -0,0 +1,93 @@
+# Backup/restore data-safety mutation audit (round 2) — 7 gates pinned, 1 gap closed
+
+- **Type:** test-quality audit + one test added (code change: `internal/api/handlers_backups_test.go`)
+- **Date:** 2026-07-08
+- **Area:** `internal/api` (`handlers_backups.go`), `internal/backupjob`
+- **Task:** continues the #82 test-quality integrity audit onto the on-demand
+  backup/restore surface, which postdates the 18-gate round-1 audit
+  ([mutation audit `4626ab5`](2026-07-07-test-quality-mutation-audit.md)). The
+  backup/restore endpoints (`7a7c0d5` / `f2fc57c` / `fc748d3`) were not in that pass.
+
+## Method
+
+Same as round 1: apply a one-line mutation to a fail-open gate in the source, run the
+package tests, confirm the **specifically-named** test reddens with an *assertion*
+failure (`--- FAIL: <subtest>`), then revert. A build break (`declared and not used`,
+`undefined`) is not a valid verdict, so mutations are operator-flips that keep every
+operand referenced (`!=`→`==`, drop a `!`, a literal→`true`, or `if false && <orig>` to
+disable a gate without orphaning its variables). Airtightness: each mutation is re-run
+with `-run` scoped to the intended subtest and `grep -- "--- FAIL: <subtest>"`, so a
+reddening sibling can't be mistaken for the gate under test. Oracle: WSL Fedora-44,
+go1.26.4.
+
+## Fail-open gates mutation-verified (all CAUGHT at the named subtest)
+
+| # | Gate (file:line) | What it guards | Mutation | Subtest that reddened |
+|---|---|---|---|---|
+| A | `handlers_backups.go:282` enqueueBackup stopped-gate | RWO double-mount / torn archive while the world is up | `!=`→`==` | `TestBackupNow/starting_server_->_409_not_stopped` |
+| B | `:151` restore stopped-gate | restore Job can't mount a live world's RWO PVC | `!=`→`==` | `TestRestoreBackup/starting_server_->_409_not_stopped` |
+| C | `:136` restore former-owner match | a fresh claimant resurrecting the previous owner's world | `!=`→`==` | `TestRestoreBackup/current_owner_who_is_not_former_owner_->_403` |
+| D | `:116` restore cross-server guard | restoring server A's backup onto server B | `!=`→`==` | `TestRestoreBackup/restore_by_backup_id_cross-server_->_403` |
+| E | `:214` backup owner-or-admin authz | a stranger backing up someone else's world | drop `!` | `TestBackupNow/non-owner_->_403,_no_backup` |
+| F | `:26` list cross-user scope | a user seeing other tenants' backups | `p.IsAdmin()`→`true` | `TestListBackups/user_sees_only_own_former-owned_present_backups` |
+
+## The gap this audit found — and closed
+
+**Restore's owner-or-admin gate (`handlers_backups.go:85`) was not pinned by any test.**
+It is the twin of gate E, but the two are *not* symmetric. Disabling it
+(`if false && !a.isOwnerOrAdmin(p, rec)`, which keeps `rec` referenced so the package
+still builds) reddened **nothing** — `go test ./internal/api/` stayed `ok`. The same
+disable applied to backup's L214 (gate E) reddened `non-owner` immediately, proving the
+technique valid and the asymmetry real.
+
+Root cause: the former-owner gate at L136 backstops every non-owner case the suite
+exercised (a stranger and a wrong-backup current owner both fail L136 *and* L85, so
+L136's 403 masks a broken L85). The one case only L85 catches went untested: a
+**superseded former owner** — a user who owned a server, took this backup
+(`FormerOwner=them`), then released it to a *new* owner. They still pass L136 (they *are*
+the former owner) but must be stopped by L85, or they could roll the new owner's live
+server back onto their old world (cross-tenant clobber). The handler comment names this
+the "must re-claim first" rule (`handlers_backups.go:77-79`).
+
+**Fix (code):** added `TestRestoreBackup/former owner after release -> 403, no restore`,
+the mirror of the existing L136 test. Verified both directions: green on the clean tree,
+and it is the sole subtest that reddens when L85 is disabled — so it now pins the owner
+gate specifically, not L136. No production code changed; `handlers_backups.go` is a pure
+test addition away from where it was.
+
+## Enumeration — covered vs. scoped (so "the gates" means all of them)
+
+- **Fail-open data-safety gates — all pinned:** A–F above, plus L85 (now closed). 7/7.
+- **Accountability, not fail-open (verified non-vacuous):** the `os_user` attribution at
+  `handlers_backups.go:254` — a supplied operator name overrides the default `break-glass`
+  audit actor. Mutating `u != ""`→`u == ""` reddens
+  `TestInternalBackup/os_user_body_attributes_the_audit_to_the_operator`, so the
+  attribution test isn't vacuous. A failure here degrades the audit actor; it grants no
+  bypass, so it is out of the fail-open bucket.
+- **Contract/behavioral (tested, out of mutation scope):** nil `Backuper`/`Restorer` → 503;
+  a failed backup/restore → 500 **not** audited; `backup_ref` never serialized to the wire;
+  the internal-face actor defaults to `break-glass`. Each has a direct test; none is a
+  fail-open safety gate.
+- **`internal/backupjob` (glanced, not mutated):** orchestration only — each backup gets a
+  unique Job name (`BackupJobName` + random suffix) so a repeat "立即备份" tap can't collide
+  with a just-finished Job still inside its TTL; `ErrAlreadyExists` is a defensive no-op;
+  `Backup` returns once the Job is created (the async 202 is honest). Unit-tested against a
+  fake `Jobs`; the controller-runtime `k8sjobs.go` and the `jobspec.go` Pod shape (weak SA
+  with its token un-mounted, config Secret mounted for the self-recorded row, read-only
+  world mount) are integration-verified per the package doc — not fail-open API gates.
+
+## Coverage nuance (documented, not a gate failure)
+
+The stopped-gate is `if info.Ready || info.DesiredState != DesiredStopped`. The `!=`→`==`
+mutation pins the `DesiredState` operand (both A and B reddened), but `info.Ready` is not
+*independently* pinned: no test sets `Ready=true` together with `DesiredState=Stopped` —
+the stopping-but-still-up race. Low risk because in practice `Ready` drops as
+`DesiredState` leaves `Stopped`, but the belt-and-suspenders `Ready` operand rides on
+coverage of the operand beside it rather than its own case.
+
+## Verdict
+
+The backup/restore data-safety surface is a coherent unit, and this closes it: **7/7
+fail-open gates pinned** (6 pre-existing, 1 added this round), one accountability gate
+shown non-vacuous, one coverage edge documented. Not extended to every handler — that
+would be an unbounded "continue the audit."
diff --git a/docs/changes/INDEX.md b/docs/changes/INDEX.md
index c2ed069..950d4e8 100644
--- a/docs/changes/INDEX.md
+++ b/docs/changes/INDEX.md
@@ -250,3 +250,5 @@ primary record.
 | 5a7cd5a | 2026-07-07 | docs(changes): fold 346ec68 cloudflare-edge walkthrough into its detail doc |
 | 4626ab5 | 2026-07-07 | docs(changes): mutation-audit the ledger's "unit-tested" safety claims |
 | 729bd7b | 2026-07-07 | docs(changes): index the mutation audit and two lagging ledger rows |
+| c67a4d3 | 2026-07-08 | docs(changes): close §B4 with the S3 archive backend deferred by design |
+| 85b8a92 | 2026-07-08 | test(api): pin restore's owner gate against a superseded former owner |