DLSS 5 Autopilot 1.8.1

Answers to what the first day of 1.8.0 reported, and one the shared results
found that the tool had not been giving.

Updating from 1.8.0 or older with [ update now ]: the old updater can
end with a box saying "Security validation failure: failed to obtain
executable path for parent process". The update is already in place when
it appears - close it and start dlss5-autopilot.exe again. That box is a
bug in the updater being replaced (#136), and 1.8.1's own updater does not
have it.

What changes for you

• Enshrouded and other Vulkan games that carry NVIDIA's crash library are
  no longer read as DirectX 12, so they get a route that loads (#130).
• The trial exe of a game is no longer picked over the full game (#131).
• Ticking multi-frame generation no longer stops the whole install when
  its project changes shape (#141), and a broken download in the cache is
  thrown away instead of failing every install after it (#140).
• On the optiscaler route, OptiScaler's own "update available" note no
  longer points at a mainline build with no neural rendering (#51).
• The left side is places, not steps; video and RTX Remix have pages of
  their own, and the app opens on your games.
• Wrong diagnoses corrected: "update the driver" on new drivers (#127),
  "OptiScaler never loaded" (#110), a DLSS game read as having none (#119).

On driver 616.64 and newer, try the standalone route

Many of the failures reported since 1.8.0 trace to one change: from 616.64
the driver routes neural rendering through its own runtime, and the
renodx-dlss5 add-on the feeder route loads faults there - 4.6 and 4.7 on
every evaluate, 4.55 in some games. The tool pinned 4.55 on these drivers
and, where that faulted too, said to roll the driver back.

The shared results point to a way round that needs no rollback: one game
failed three times on the feeder route on driver 616.92 and then worked
twice on the standalone route, reported by the same person with the same
RTX 5080 and driver. Standalone runs its own feed and never goes through
renodx-dlss5. So on these drivers, for a 64-bit D3D11 or D3D12 game:

• the install page says so on the routes that load renodx-dlss5 (feeder,
  native, bridge), before anything is written;
• the diagnosis names the standalone route when it finds this fault, ahead
  of the rollback;
• the standalone route is no longer warned off with a fault record that
  belongs to the add-on it does not load.

It is one game, so it is offered as the thing to try rather than made the
default: standalone presents through a window of its own and is still
marked experimental. Games that are not offered standalone - 32-bit,
DirectX 9 and 10, Vulkan and OpenGL - are not sent looking for it. Every
result shared from the tool goes into the same list, and that is what will
say whether it holds.

Wrong answers, corrected

• "Update the graphics driver" on a driver that is new, not old
  (#127). Failed to find NVSDK_NGX_..._EvaluateFeature_C was always read
  as "your driver predates DLSS 5", and on 616.92 that is not what it
  means: on 616.64 the same line appears in sessions that deliver frames,
  beside a successful hook of the plain call. The driver's version now
  decides - older than 616.56 is still told to update, a driver that
  carries DLSS 5 is not.
• The bridge's substitute contract was never switched on (#127). For a
  game with no DLSS of its own the install writes synth_after into
  dlss5-bridge.cfg, and from bridge 1.4.0 on the bridge replaces a
  settings file that does not start with its version line with its
  defaults - substitute off. The file now starts with # dlss5-bridge keep,
  the bridge's own way of keeping a file across its versions, so what the
  install writes is what the bridge reads. A bridge install on such a game
  from an earlier version has to be installed again; "did it work?" says so
  when it finds one.
• A game with its own DLSS read as a game without it (#119). NBA 2K27
  keeps Streamline under data\streamline, and folders named data were
  skipped to keep the search short - so the game was offered the route for
  games with no DLSS. A skipped folder is now entered when one of its own
  subfolders is named like a place runtimes live, and of its subfolders
  only those are followed.
• "OptiScaler never loaded" about an install that was working (#110).
  One of the OptiScaler builds this tool installs ships with its log
  switched off, so there was nothing to read, and "no log" was taken to
  mean "not loaded". The install switches the log on, and when there is
  no log the diagnosis reads OptiScaler.ini: with logging off it says so,
  and only with logging on does a missing log mean the game has not run or
  OptiScaler did not load.
• A game listed twice - Steam reports the library folder, the Rockstar
  launcher the game's subfolder inside it, and both were listed. One
  executable is one entry now, the one with an install or a hand-set
  graphics API kept.
• A bug report that dropped the line the diagnosis reads (#127). The
  report's ReShade.log excerpt lost its oldest lines first, and the add-on's
  hook lines are written at the start of a session; the excerpt keeps them.
• The optiscaler route recommended without saying what it needs
  (#116). An FSR or XeSS runtime on disk is not an upscaler the player can
  turn on. The route now says, before the install, that the game's own FSR
  or XeSS has to be switched on in its settings, and that the feeder route
  is the one to use when the game has no such setting.

Scrolling, and the dropdowns that changed under it

On the pages that scroll - the first page and the install settings -
the mouse wheel only worked in the bare strip beside the controls: the page
is a child window laid over its scroll area, and moving onto any control
dropped the wheel. And a wheel turn over a dropdown changed its selection:
scrolling past route or dlss5 add-on could pick a different one
without a click. The wheel now scrolls the page from anywhere on it and
leaves dropdowns alone; the log and the game list keep their own wheels.
Closing the remix list no longer takes the main window's wheel with it.

The window

• The scrollbars were Windows' light-grey default on a dark window, with
  grip marks in the middle of the thumb, and looked broken. They are thin
  and in the window's own colours.
• The dropdowns were the one proportional font in a monospaced window.
• The route description was one block of grey text. The description, the
  conditions and the driver warning are now told apart by colour, and a
  wrapped warning stays indented under its own mark.
• The checkboxes drew a light-grey frame round themselves; the aim-for-fps
  box turned white on the routes it does not apply to; the feeder build
  note was cut off at 150% and 200%; the game list's column headings were
  centred over columns whose text is left-aligned.
• After an update the app opened on the architecture page, as if it had
  never been run: the saved library belongs to the version that wrote it. With "scan
  library at start" on, anyone who has scanned before opens on the game
  library, and it is read again there.
• The left side holds places instead of numbered steps: games (the
  library), install, and video and youtube and rtx remix on
  pages of their own - they were two cards on the architecture page, which
  is gone: its filter is a dropdown beside the search on the games page. The remix page
  names the games in your library that have a mod, and which of them it can
  download and install for you.
• The windows the app opens - the remix list, the bug report - carry its
  icon instead of Tk's blank one.
• The name in the top-left corner goes home: to the game library, or to the
  start page before any game has been scanned. Ctrl+H and Esc do the same.

Shared results close themselves

A result shared from the tool is read into the compatibility list and then
labelled and closed, with a note saying where it went. It still counts -
the list reads closed results too - and the issue list shows the reports
that need someone.

Under the hood

• _tools/gui_lint.py walks every widget on every page and on every
  route but remix, at 100%, 150% and 200%, and reports clipped text,
  anything past the window's edge, colours outside the palette (disabled
  and read-only states included), a second font family, and overlaps.
• _tools/brain.py reads the shared results the way a maintainer needs
  them: success by route and by driver, games where one route rescued what
  another could not, and games failing on everything tried.