feat(whisper): verify GPU backends and add Vulkan fallback

[IdrisGit]

Summary

• Add CUDA → Vulkan → CPU fallback for Whisper startup and inference failures.
• Verify the backend from whisper.cpp startup output before reporting GPU acceleration.
• Display the actual active CUDA or Vulkan backend in Settings.
• Require the complete bundled Linux CUDA runtime during installation.
• Prefer the bundled Linux CUDA libraries over inherited system CUDA paths.
• Update GPU documentation and localized UI copy.

Release dependency

This PR depends on the corresponding OW-whisper.cpp PR: OpenWhispr/whisper.cpp#7

The whisper.cpp PR must be merged and released first. After its release is published, this PR must update WHISPER_CPP_TAG and the matching SHA-256 digest in src/helpers/whisperCudaManager.js before it can ship.

Validation

• Confirmed CUDA and Vulkan locally with npm run dev.
• Passed 74 targeted tests.
• Passed typecheck, Vite production build, Prettier, and i18n JSON consistency checks.

Follow-ups for #2168 (boseq)

• Merged main. The fallback events carry fix(whisper): save and show why GPU acceleration fell back to CPU #2317's reasons. The Linux CUDA libs go through main's requiredLibraries, so an installed 0.0.10 pack shows "GPU pack needs re-downloading".
• When the bundled whisper-server is itself a GPU build (the AUR openwhispr-cuda package), Settings and the Home banner don't offer the pack. A pack failure that lands on that server doesn't say "switched to CPU".
• A Vulkan pin to a device that's gone is cleared before the backend check. A CUDA → Vulkan fallback updates open GPU cards. There's no re-download notice for a pack that failed and was then removed.
• Details: feat(whisper): verify GPU backends and add Vulkan fallback #1982 (comment)

Fixes #1415
Fixes #945
Fixes #2168

Related: #712, #126, #531.

[boseq]

Hey @IdrisGit, I pushed some commits on top of this. #2168 (the AUR openwhispr-cuda package on Arch's CUDA 13) has the same root cause as #1415, and this PR fixes it, so I wanted it mergeable again. Your commit is untouched: what I added is a merge of main plus ten commits.

• Merge with main. It conflicted with fix(whisper): save and show why GPU acceleration fell back to CPU #2317, which saves the reason for each GPU fallback. The fallback events now carry both reason and your fallbackBackend, and _fallbackAndRetry keeps the reason extraction.
• requiredLibraries instead of requiredFiles. Main already has requiredLibraries (fix(whisper): extract MSVC runtime DLLs beside win32 whisper-server #1687, for the Windows MSVC runtime), and since fix(gpu): tell users when an update needs a newer GPU pack #2501 a pack that lacks those libs shows as needing an update.
  • With a separate requiredFiles, an installed 0.0.10 Linux pack read as never downloaded. The card offered a fresh "Enable" instead of "GPU pack needs re-downloading".
  • The CUDA EULA stays in the release zip (Gemini Support? #7 puts it there) but isn't copied into the pack folder anymore. Happy to put it back if you'd rather keep it there.
• The backend check and the Vulkan device pin (After upgrading to 1.8.3, Windows GPU detection now fails for all previously-working versions of OpenWhispr #1606). A pin to a device that's gone makes whisper.cpp start on the CPU. The check failed Vulkan before the pin could be cleared, so the user stayed on CPU and Retry failed the same way. The pin is now settled first.
• fallbackToVulkan is always set. Restarts merge it over the last start's options, so leaving it out when false kept an old true after the Vulkan pack was removed.
• The card's 5 s status poll only ends "activating" again, as on main.
  • Taking the pack and failure state from the server mislabelled a failed pack once a bundled GPU build was running, and Remove then deleted the other pack.
  • A CUDA failure that lands on Vulkan now tells open GPU cards to re-read instead, so Settings still follows the server to the Vulkan pack. There's no CPU pop-up for it.
• A pack failure that lands on a bundled GPU build reports the backend the server actually runs on, so no "switched to CPU" pop-up shows then either.
• A reason for a GPU server that came up on the CPU. fix(whisper): save and show why GPU acceleration fell back to CPU #2317's reason code assumed the process had exited or timed out. When the server's log names no cause (only "no GPU found"), the card said "startup timed out after 30 s". Now it says the server started without initializing its GPU backend. A log line that names the cause (e.g. ggml_cuda_init: failed to initialize CUDA: …) still wins.
• No pack offer when the bundled server is already a GPU build. The AUR package swaps in a CUDA 13 build of whisper-server-linux-x64, and your stderr check reports it as CUDA. Once that server is running, Settings and the Home banner skip the offer. The banner checks again when the startup pre-warm comes up.
• No "needs re-downloading" notice for a pack that failed and then disappeared. The AUR launcher deletes the pack on every start, so at launch such a pack's enabled flag is dropped, the same as a delete would.
• Arabic strings for your new keys (Feat/arabic UI localization #1869 added Arabic), and fix(whisper): GPU card follow-ups — hex crash codes, live Retry in Settings, show the pack in use #2335's card tests now expect the per-backend wording.

Each fix has a test that fails without it, with two gaps: the launch-time flag drop runs in main.js startup code that no test covers, and the bundled-GPU fallback report is tested on the startup path but not the mid-transcription one. The CI checks pass locally. I also ran the 0.0.10 Linux binary on Arch (nvidia-utils 615.78, no CUDA 12 anywhere):

• On its own, it fails on libcudart.so.12.
• With the CUDA 12.9 libs from NVIDIA's pip wheels next to it and your LD_LIBRARY_PATH, it loads and starts.
• That machine has no GPU, so it logged ggml_cuda_init: failed to initialize CUDA: no CUDA-capable device is detected. A unit test feeds that same output to the app's start path and checks it falls back with that line as the reason.

This still needs OpenWhispr/whisper.cpp#7 merged and released, then the WHISPER_CPP_TAG and digest bump here. Until then, Linux "Enable CUDA" fails with "missing required libraries", so this can't merge first. #7 conflicts with master too.

[IdrisGit]

@boseq I have resolved the conflict in whisper.cpp#7. Thanks for the heads-up about the release dependency. Once OpenWhispr/whisper.cpp#7 is merged and released, I’ll update WHISPER_CPP_TAG and the Linux CUDA artifact digest in this PR.