fix: test-script review findings (v3.5.13)

PWA:
- The SCP80 "Source" switch sticks: the choice is stored in
  `params.source`, switching keeps both values (the server honors the
  explicit source when both are present, and the form validates only the
  selected one).
- The SW check field is empty by default (placeholder "default: 9000
  (91?? when polling)"), so the server defaults apply - previously the
  pre-filled 9000 defeated the polling STATUS default and a card that
  announced a command made the step fail with "expected 9000".
- Item text checks offer contains/exact only (the mask mode was rejected
  by the server validation).
- The SCP80 counter is written back to the preset even when the run is
  observed after a page reload or was started elsewhere (resolved by
  ICCID or preset name from the run snapshot).

Server:
- Step-entry mutations happen under `_TEST_LOCK`
  (`_test_entry_update`/`_test_finish_entry`): the status endpoint
  serializes the state with json.dumps, so entries must not change while
  it iterates them.
- The pending-command drains (unexpected command, error, stop) hold
  `_CARD_LOCK` like every other card conversation.
- `_int` accepts plain decimals with leading zeros and 0x hex.
- The scripted TERMINAL RESPONSE text string uses the CR-set tag `8D`
  (consistent with the other TR TLVs; both are legal).
- A script-driven menu selection mirrors `server.menu_active`.

Tests: frontend +4 (source switch/round-trip, status SW default, render
checks, item modes), Python +3 (integer parsing, menu_active, source
selection).  Help/docs unaffected beyond api.md's `source` note.

583 frontend / 470 Python green.

frontend/sw.js

@@ -1,4 +1,4 @@
-const CACHE = 'simple-v265';
+const CACHE = 'simple-v266';
 const URLS = [
   'index.html',
   'help.html',