[MarcusJRLee]

What

• Speech models (iOS): Settings → Speech model lists the built-in Whisper Tiny English plus one-tap Download for Whisper Base English (85 MB), Whisper Small English (270 MB), and Parakeet v3 (741 MB, the Mac default), with progress, Cancel, ✓ for the active model, and swipe-to-remove. Downloads use a new module shared with the Mac (Sources/hardware_controller_model_download): HTTPS-only, commit-pinned URLs, per-file SHA-256, then Rust admission (0062). Import from Files moves to its own section with an explanation.
• History: swipe or detail Delete Dictation (removes audio first), ⋯ → Clear All History, both confirmed; an in-progress recording is untouched.
• Voice Keyboard: system-proportioned keys with letter/number/symbol layers, auto-capitalization, double-space period, hold-to-delete, input clicks; one voice bar (Style, plain-language status, round mic that turns amber while the app records).
• Icons: Center Signal family drawn in code (tools/app_icons/render_app_icons.swift): three pedals (Mac) and a voice waveform (iOS), center lit amber (0063).

Boundaries

• The fetcher is Voice Input's only network code; check_ios_local_only.sh fails if anything but the app links it and build_ios_device.sh fails if an extension contains it. Conventions confine network APIs to the fetcher and the Ollama adapter.

Verification

• scripts/check.sh and scripts/check_ios.sh (183+ unit tests, 5 UI tests including History delete/clear-all) pass.
• Simulator: real Whisper Base download → verified → active → JFK clip transcribed with punctuation; keyboard checked in light and dark, letters and numbers.

🤖 Generated with Claude Code