Add on-device LLM post-processing for transcriptions

[userFRM]

Summary

• Adds manual per-recording LLM text processing using Apple MLX framework for on-device inference (no cloud API calls)
• Users can clean up filler words, restructure as developer prompts, or format as structured markdown via a wand menu on each recording
• New LLM module with model lifecycle management, curated model registry (0.5B–4B), and conditional compilation for non-MLX builds

Details

New files

• OpenSuperWhisper/LLM/LLMPostProcessor.swift — Singleton managing MLX model lifecycle with generation counter for async race protection, task coalescing for duplicate loads, and Task.detached for off-main-thread inference
• OpenSuperWhisper/LLM/LLMProcessingMode.swift — Processing mode enum (raw/clean/dev/markdown) with system prompts
• OpenSuperWhisper/LLM/LLMModelRegistry.swift — Curated registry of 8 MLX models (5 general purpose, 3 code-specialized)
• OpenSuperWhisper/LLM/DevModePrompts.swift — System prompt for converting spoken developer dictation to structured markdown

Modified files

• ContentView.swift — Per-recording LLM processing via wand icon context menu, shimmer overlay during processing, revert-to-raw support
• Settings.swift — @MainActor on SettingsViewModel, Speech/LLM sub-tabs, model install flow with cancel button and rollback on failure/cancel (race-safe via model-id guards)
• Recording.swift — rawTranscription field, DB migration v3 with backfill
• IndicatorWindow.swift — rawTranscription set on hotkey recordings, removed dead .processing state
• TranscriptionQueue.swift — Passes rawTranscription through on transcription completion
• AppPreferences.swift — LLM preferences (processing mode, model ID, temperature, installed model IDs)
• PermissionsManager.swift — Use AXIsProcessTrustedWithOptions instead of AXIsProcessTrusted()
• run.sh — Dynamic libomp discovery, Metal Toolchain checks
• project.pbxproj — MLX framework references, mlx-swift-lm pinned to upToNextMajorVersion: 2.30.3

Test plan

• Build with MLX enabled — verify LLM sub-tab appears, models can be installed/cancelled
• Build without MLX — verify app compiles, LLM UI shows "unavailable" messaging
• Record via main window — verify rawTranscription is set, wand menu works (clean/dev/markdown)
• Record via hotkey indicator — verify rawTranscription is set, revert-to-raw works
• Cancel model install mid-download — verify previous model is restored, no error toast
• Cancel then immediately install another model — verify no state clobbering (race scenario)
• Process text then revert to raw — verify original transcription is restored
• Drop audio file for transcription — verify rawTranscription is set on completion
• Run ./run.sh on both Intel and Apple Silicon Macs

🤖 Generated with Claude Code