feat(dictation): Parakeet v3 as a dictation model

[Maxaubert]

Closes #121.

Adds Parakeet v3 as a dictation model. This is a change to Prism too: it is in core/ (core 0.23.0), so Prism gets it through the automatic core bump. Merge Maxaubert/Prism#279 first: Prism's dictationPage gate would otherwise fail on the new row (verified: it passes there on both the old and the new core).

What it does

• Catalog: ggml-parakeet-tdt-0.6b-v3-q8_0.bin (668757119 bytes, SHA-256 4d64e9e9...fd600e), URL pinned to ggml-org/parakeet-GGUF commit 35156454d1a39de06863303dd209fd2bed6ee079, flagged limitedLanguages.
• Engine: parakeet-cli.exe + parakeet.dll from the same b5130 zip and GPU pack; fetch-whisper.mjs and ENGINE list the same files (test holds them).
• Main: one parakeet-cli per pass, the WAV on stdin (-f -), never written to disk; text off stdout (UTF-8). GPU when the pack is installed and enabled, else -ng. A GPU pass falls back to the CPU only when the engine died, Whisper's rule (an error: about the audio or a timeout keeps the GPU). Warm-up is a silent pass at the Whisper warm-up's hooks; partials one at a time through the existing queue. Switching off, a model or pack change and quit kill a pass in flight.
• Download prompt: "This model doesn't support all languages." with Cancel / Download.
• Language picker: while the active model is flagged it shows Auto-detect, disabled, with a globe icon whose tooltip is "This model doesn't support all languages."; the stored language is untouched and comes back with a Whisper model.

Tests

• Typecheck clean; lint 0 errors (12 existing warnings); unit 1236/1236.
• npm run e2e -- dictation: dictation 21/21, dictationPage 36/36, dictationParakeet 11/11 (real engine, q4_0 file, jfk clip: the sentence lands on the prompt, no Enter).
• Measured outside the app: the real parakeet-cli reads the WAV from stdin and gives the exact sentence; the silent warm-up compiled the GPU kernels in 23.7 s on an empty cache, the next pass 0.59 s.

For the owner

• The Parakeet row in the model list carries the NVIDIA mark, as the Whisper rows carry OpenAI's; the pop-up and tooltip name no vendor. The row's note reads "Fast and accurate on any PC, in fewer languages."; say if you want other words or none.
• Not packaged or installed (per the task); hands-on: how it sounds, and the GPU path on your card.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHHaWKR4M5QtW7Wecyuk4t