The Lcn, LcnXX and Checkskin prompts used 'if answer is None / else',
so answering No fell into the else branch and ran the scan or check
anyway. Guard with 'elif answer:'. LcnXX also reopened its prompt with
self.Lcn as callback, so the answer ran the wrong handler. Checkskin
additionally connected its eTimer to the return value of
check_module_skin() - the function was already executed and returns
None, so the timer fired into a TypeError; run the check directly and
drop the timer.

- A missing comma in the generic pattern list concatenated two regexes
  into one broken pattern that could never match.
- A dead server made make_request return None and '"x" in None' raise
  a confusing TypeError; fail cleanly with a clear message instead.
- The pattern loop kept scanning after a successful match, so a page
  matching several patterns appended duplicate server entries to the
  config; stop after the first hit.

- removenow: package filenames without an underscore lost their last
  character (ipk[:-1]), so opkg remove targeted a nonexistent package;
  strip the extension instead.
- The grid page count used npics // 20 + 1, creating a phantom empty
  page whenever an item count is an exact multiple of 20 (all six grid
  screens share the formula).
- requests.get in runScriptWithConsole and retfile had no timeout and
  could freeze the GUI thread forever on a stalled connection.

check_vers compared version strings lexicographically, so an installed
2.10.x would be offered a downgrade to 2.9.x and 2.9.x would never see
a 2.10.x update. Use the shared version_tuple helper.

'Label: %s' % x if x else 'Unknown' binds as ('Label: %s' % x) if x
else 'Unknown', so any missing value printed a bare 'Unknown' line
without its field label. Format through a small helper instead, and
show the real boolean for the VTi/DMM image flags (False previously
displayed as 'Unknown').

- TransferBouquetTerrestrialFinal defined its inner RestoreTerrestrial
  function but never called it, so the terrestrial bouquet restore
  after a settings install was a no-op.
- addInTVBouquets compared the bouquet's full path against bouquets.tv
  lines that only contain the file name, so the already-present check
  never matched and a duplicate reference (to a possibly different
  hardcoded name) was appended on every run; compare and write the
  actual bouquet file name.
- RestoreTerrestrial checked for '#NAME' inside line.lower(), which can
  never match, so the bouquet was never renamed to Digitale Terrestre.
- Terrestrial.py's ServiceScan hook imported itself through
  .LCNScanner.Terrestrial, a path that does not exist from inside the
  LCNScanner package; the class is defined in the same module.

translate_utils.py is imported by nothing and cannot be imported at
all: it expects DEBUG, HEADERS and SYSTEM_DIR from the package, which
do not exist (leftover from the ForecaOne plugin). The translation
workflow uses its own copies in update_translations.py.

bissfeedautokey.sh and ipaudiopro_1.4.sh are byte-identical duplicates
of Bissfeedautokey.sh and Ipaudiopro_1.4.sh; the capitalized names are
the ones referenced by plugin.py. Multistalker_pro.sh and
multisalker_pro1_eliesat.sh differ in content and stay untouched.

Opening the Information screen imported stbinfo on the GUI thread; its
module-level StbInfo() runs pings, TCP probes and up to four public-IP
lookups, blocking the whole interface for 20+ seconds on first open
(worse when offline, since every lookup runs into its timeout).

- LSinfo now shows 'Collecting system information...' immediately,
  gathers the data in a daemon thread (including the heavy first
  import) and applies the text from an eTimer on the main thread. The
  poll stops on screen close so a late result cannot touch a dead
  widget.
- stbinfo tracks connectivity as a boolean and skips the public-IP
  lookups entirely when the box is offline - previously the offline
  case was the slowest. Lookup timeouts trimmed to 3s.

okRun1 wiped lamedb and every .tv/.radio file from /etc/enigma2 before
checking that the download or unzip had succeeded: urlretrieve had no
error handling, the unzip exit code was ignored, and a zip with an
unexpected layout made the cp copy nothing - any of these left the box
with no channel list at all.

The install now runs download -> verify -> backup -> wipe -> install:
- download via requests with a timeout and clear failure message,
  aborting before /etc/enigma2 is touched;
- unzip exit code checked; the payload must actually contain a lamedb
  or .tv file (channel lists at the zip root are now handled too, and
  the top-level folder is detected instead of walking into the deepest
  subdirectory);
- the current /etc/enigma2 is saved to /tmp/settings_backup.tar.gz
  before the wipe and restored automatically if the copy fails.

Six screens (LinuxsatPanel, LSskin, LSChannel, LulullaScript,
CiefpInstaller, ScriptInstaller) each carried a byte-identical copy of
the 20-tile grid machinery: title/skin/resolution setup, widget and
ActionMap wiring, paging (openTest/paintFrame), navigation
(key_left/right/up/down), sorting and the view-log prompt. Every fix
in this area had to be applied six times.

The engine now lives once in an LPGridScreen base class: subclasses
fill their menu lists and call initGrid(). Behavior is unchanged,
verified by driving every screen through full navigation wrap-around,
paging and sort/restore in a stubbed enigma2 environment.

Also folded in while moving the code:
- _view_log kept its eTimer in a local variable, so the timer could be
  garbage collected before firing and the log viewer silently never
  opened; the timer is now held on the screen instance.
- Dropped keyNumberGlobal, dead code referencing a "menu" widget that
  none of these screens define, with no key binding pointing at it.

plugin.py shrinks by ~1150 lines net.

Every remaining network operation ran on the enigma main thread and
froze the whole GUI: the addon catalog was fetched at panel open AND
again on every category click, provider pages were scraped
synchronously when a channel-list screen opened, package downloads
blocked until complete, and the settings install (download + unzip +
copy) locked the interface for its entire duration.

- New AsyncCall helper runs a blocking function in a daemon thread and
  delivers the result to a callback on the main thread via eTimer.
  The AsyncMixin (used by LPGridScreen and addInstall) cancels pending
  callbacks when the screen closes, so a late result never touches a
  dead widget.
- The catalog is fetched once per session (5-minute TTL) and warmed in
  the background when the panel opens, so a category click is served
  from cache instantly; the shared _openCategory lives in LPGridScreen.
- Provider page scraping, package downloads and the settings install
  now show a status text and run in the background; the console or
  result message appears when the work is done.

Verified in the stubbed enigma2 environment end-to-end: cache hit
counting, panel prefetch, category click -> addInstall slice, provider
page failure path, async package install -> opkg console command, and
a full settings install with a real zip archive through the real shell
(unzip, payload check, backup, wipe, install), plus grid navigation
regression on the refactored screens.

- Search (GREEN button, shown in the skins next to Exit): on the main
  panel it searches the entire addon catalog across every category and
  opens the matches as a normal installable list; on the category
  grids it jumps straight to the first matching tile, changing page if
  needed. Runs through the async helper, so a cold catalog never
  blocks the GUI.
- Installed addons are shown in green in the addon lists. The
  installed-package scan that message() already did on click is now a
  shared helper (get_installed_packages) and feeds one flag per list
  entry.
- The startup update check, present but disconnected since its
  onLayoutFinish line was commented out, is enabled again: it runs in
  the background at panel open and shows a single clear message
  (version, changelog, how to update) instead of the old
  modal-detection logic. Uses the fixed numeric version comparison.
- AsyncCall hardened against double delivery of a result.

Verified in the stubbed environment: update prompt on a newer remote
version, global search returning only matching catalog entries plus
the no-match path, grid search jumping to the correct tile and page,
installed-flag computation against the package database, and list
entry rendering with the installed color. All three LinuxsatPanel.xml
skins extended with the key_green label and re-validated as XML.

Lint (ruff, pylint, autopep8) and the translation updater only
triggered on main, so develop work got no checks until release. The
two auto-commit workflows also hardcoded 'origin main' in their
pull/push steps; they now use the triggering branch, and the autopep8
commit step is limited to push events so pull requests are lint-only.
Autotag stays main-only since tags mark releases.

Version bump across __init__.py, CONTROL/control, installer.sh and the
README badge. The installer.sh changelog (what the in-plugin update
check displays) summarizes the 3.0.0 line:

- Search all addons with the GREEN button
- Installed addons shown in green in the lists
- No more GUI freezes: all network work is asynchronous and the addon
  catalog is cached per session
- Settings install verifies the download and keeps an automatic backup
- Update check runs at panel start
- Grid engine unified in one base class, dead code removed and a long
  list of bug fixes (see the individual commits)

key_up's conditions almost never matched a real cursor state: pressing
UP anywhere but a page's first tile fell into a branch that jumped to
the last page, and because openTest() resets the index to the page
start, on page 2 the cursor just snapped to the first tile instead of
moving up a row. key_left's wrap had the same openTest problem, landing
on the first tile of the previous page instead of the last one.

All four direction keys are rewritten with plain row/column logic:
LEFT/RIGHT step one tile and wrap page-to-page at the edges; UP/DOWN
move one row and cross pages keeping the current column (clamped on a
partial last row).

Verified exhaustively in the stubbed environment: every key from every
position on grids of 6, 16, 20, 25, 37, 40 and 45 items keeps the
cursor inside the visible page, full LEFT/RIGHT cycles visit every
tile exactly once and wrap home, UP/DOWN are inverse operations within
a page, and the reported case (UP on page 2) moves one row up.

The last page pre-filled all 20 slots with the LSS.png placeholder and
then overwrote only the occupied ones, so a page with e.g. 14 entries
showed 6 leftover placeholder icons. openTest now draws the occupied
tiles and hides the unused pixmaps (re-showing them when a full page
is displayed again); the min/max entry computation for full and last
pages is unified in the process.

Verified for 6/16/20/25/34/37/45-item grids: the last page shows
exactly its own tiles, returning to a full page restores all 20, and
the navigation invariants still hold from every position.

The import inside check_and_install_requests exists only to test
whether the module is installed before falling back to opkg/apt-get,
so the unused-import warning is intentional. importlib.util.find_spec
is not an option while the Python 2 code paths remain.