fix(multiuser): close cross-user web-layer scoping holes found in review

The opt-in multi-user feature's only enforcement is web-layer scoping
(all sessions share one OS account). An adversarial review found 8 critical
+ 7 high cross-user holes that defeated it, plus mediums; all fixed here.
Single-user (flag-off) behavior stays byte-identical apart from documented
consistency deltas.

Ownership / confinement:
- DELETE /api/sessions (bulk) + /:id now owner-scope / findSessionOrFail
- quick-start, cron (create+fire), scheduled runs confine workingDir to the
  owner's space; case link/docker-link/docker-import confine the host path
- resolveCasePath no longer resolves linked cases for non-admins; foreign
  remote/docker cases are skipped (fall through to the caller's own local case)
- history, subagents/workflows, mux-sessions, orchestrator, cron run-history,
  away-digest, and remote/docker host reads are owner- or admin-scoped

Permission policy (section 6.3):
- non-granted users are downgraded at every spawn site incl. legacy
  /api/scheduled, PlanOrchestrator one-shots, remote launch, and the cron-fire
  gemini/codex bypass switches; resolveClaudeModeForUsername now fails closed

Auth / store:
- verify-first login throttle (a correct password is never locked out),
  /ws terminal subject to the change-password lockbox, cookie fast-path
  re-validates identity live, role/grant changes revoke sessions, admin delete
  runs the last-admin guard before any teardown
- users.json: distinguish missing (ENOENT) from corrupt/unreadable so a bad
  read can't overwrite all accounts; unique per-process temp write path

Event streams:
- debounced session:updated + batched task:updated, clipboard, and push
  notifications route by owner (fail closed); getLightState hides machine-wide
  globalStats from non-admins

Tests: two suites updated to assert the fixed (secure) behavior. tsc, eslint,
and test:ci all green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/web/routes/admin-routes.ts

@@ -134,8 +134,12 @@ export function registerAdminRoutes(app: FastifyInstance, ctx: SessionPort & Aut
     }
     try {
       const user = await updateUser(username, parsed.data);
-      // Disabling revokes the user's cookie sessions.
-      if (parsed.data.disabled) revokeUserSessions(ctx.authSessions, username);
+      // Security: revoke the target's cookie sessions on ANY successful update. role,
+      // disabled, and canBypassPermissions are all authorization-relevant, and the
+      // cookie snapshots role, so a stale cookie could otherwise retain old privileges
+      // (a demoted admin staying admin). Idempotent, affects only the target, and
+      // forces a re-auth that re-snapshots the new record.
+      revokeUserSessions(ctx.authSessions, user.username);
       audit(req, 'user.update', user.username, parsed.data);
       ctx.broadcast(SseEvent.AdminUsersChanged, {});
       return { success: true, data: { user: toPublicUser(user) } };
@@ -177,14 +181,18 @@ export function registerAdminRoutes(app: FastifyInstance, ctx: SessionPort & Aut
     const parsed = DeleteUserSchema.safeParse(req.body ?? {});
     const deleteSpace = parsed.success ? parsed.data.deleteSpace : false;
     try {
-      // Kill the user's live sessions first (normal kill flow, incl. docker/remote
-      // teardown), before removing the record.
+      // Security: validate BEFORE any teardown. deleteUser runs the authoritative
+      // existence + last-admin guard under lock with no side effects, so a refusal
+      // (409 LAST_ADMIN / 404 USER_NOT_FOUND) leaves the user's live sessions and
+      // cookies untouched. Only after it succeeds do we irreversibly kill sessions and
+      // revoke cookies. (owned is captured from the in-memory map, independent of the
+      // record, so it is safe to read before the delete.)
       const owned = [...ctx.sessions.values()].filter((s) => s.owner === username).map((s) => s.id);
+      await deleteUser(username); // throws LAST_ADMIN / USER_NOT_FOUND (no side effects)
       for (const id of owned) {
         await ctx.cleanupSession(id, true, 'admin_delete_user').catch(() => {});
       }
       revokeUserSessions(ctx.authSessions, username);
-      await deleteUser(username); // throws LAST_ADMIN / USER_NOT_FOUND
       if (deleteSpace) await deleteUserSpace(username);
       audit(req, 'user.delete', username, { deleteSpace, killedSessions: owned.length });
       ctx.broadcast(SseEvent.AdminUsersChanged, {});