feat: transcribe audio/video files with progress, history, and srt/vtt export

[ElbertePlinio]

Closes #30 — PR 2 of 2 (commands + UI; core landed in #32).

What changed

• src-tauri/src/file_job.rs (new): file transcription session — worker thread pipeline (convert → transcribe → optional cleanup → history), pickscribe://file progress events, single-job guard, cancel token, and a panic-safe Drop guard that always removes the per-job temp dir and frees the job slot (recovers a poisoned mutex).
• Commands: transcribe_media_file, cancel_file_transcription, pick_media_file (native open dialog, Rust-side via tauri-plugin-dialog), export_history_entry (TXT/SRT/VTT via save dialog). ffmpeg added to run_doctor.
• UI: window drag-drop with ember overlay, dashboard "drop or browse" affordance, confirm panel with "Clean up with AI" default off, live progress + cancel (indeterminate until whisper's first tick), done/error/cancelled cards, History file badges + per-entry TXT/SRT/VTT export chips (SRT/VTT need segments). Modal follows the app convention: Esc/backdrop dismiss for non-running states, focus lands in the dialog.
• HistoryDb::metrics now excludes file-sourced rows so file transcriptions don't inflate dictation sessions/minutes-saved.
• Silent media emits "no speech detected" instead of saving an empty entry; a failed opt-in cleanup surfaces as "cleanup skipped: …" on the done card instead of silently falling back.

Privacy invariants

Transcription is fully local (spawned ffmpeg + whisper-cli). The cleanup LLM call happens only when the user flips the per-file toggle (default off); local-only mode keeps its existing gating. File jobs never paste, touch the clipboard, or play sounds.

Tested

• cargo test --workspace --locked --all-targets: 79 passed (new: metrics exclusion regression, export content/filename, extension validation, file-job payloads).
• bun run check: 0 errors. bun run test:coverage: 19 tests, coverage ratchet green.
• ffmpeg conversion flags verified live against WAV/MP3/MP4 samples; whisper JSON/progress format verified against the installed whisper-cli.
• Reviewed by a 3-model panel + anti-overengineering gate; all findings fixed or dismissed with rationale.

Not tested yet

• Interactive drag-drop/native-dialog smoke in a real desktop session (needs a live Wayland run; logic extracted to pure TS and unit-tested, wiring type-checked against the Rust contract).

Known deferrals

• Cancel during the optional cleanup step takes effect only after the LLM call returns (temp files still cleaned).
• Multi-file drop takes the first media file silently.
• Word-level timestamps (issue notes segment-level for v1).

[ElbertePlinio]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

Reviewed commit: a59cf2b9d9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".