security: address silent-failure review findings on top of batch 1

Second-pass fixes caught by a self-review after the initial commit — each
one would have undermined a fix from the previous commit.

- mfaPolicy now goes through `verifyJwtAndLoadUser` too. Without this,
  a JWT stolen before a password reset still satisfied `require_mfa`
  until its natural 24h expiry, defeating the whole point of the
  password_version bump.
- Drop the `?? keys[0]` fallback in OIDC JWKS key selection. When the
  token carries a `kid` that is not in the current JWKS, refuse
  outright instead of picking an arbitrary key and letting the
  signature check produce a generic failure — the real failure mode
  deserves a specific error code.
- Tighten OAuth DCR custom-scheme rule so `javascript:`, `data:`,
  `vbscript:`, `file:`, `blob:`, `about:`, `chrome:` are all rejected.
  Previously the catch-all "not http/https" check admitted them; the
  authorize flow later 302s the browser to whatever is registered,
  which with a `javascript:` URI would execute attacker script on
  redirect. Also require the private-use scheme body to be reverse-DNS
  (contain a dot), matching RFC 8252 §7.1.
- permanentDeleteFile / emptyTrash only delete the trip_files row when
  the on-disk unlink actually succeeded. Previously Promise.all
  swallowed individual unlink failures and DELETE ran unconditionally,
  so a permission / ENOSPC failure would orphan bytes on disk.
- restoreFromZip also invalidates the permissions cache in the outer
  catch. If extraction threw before the DB swap even started, the
  cache wasn't stale, but belt-and-braces is cheap and guarantees no
  failed-restore path leaves stale cache behind.

server/src/services/backupService.ts

@@ -260,6 +260,13 @@ export async function restoreFromZip(zipPath: string): Promise<RestoreResult> {
   } catch (err: unknown) {
     console.error('Restore error:', err);
     if (fs.existsSync(extractDir)) fs.rmSync(extractDir, { recursive: true, force: true });
+    // Belt-and-braces: the inner `finally` already drops the permissions
+    // cache after a successful swap, but if the extraction/copy step
+    // itself threw before the DB swap even started, the cache wasn't
+    // stale anyway. Invalidating here too costs nothing and guarantees
+    // we never serve cached permissions that don't match the DB state
+    // we leave the process in after a failed restore.
+    try { invalidatePermissionsCache(); } catch { /* best-effort */ }
     throw err;
   }
 }