Add meeting language picker and automatic Whisper language detection

[r3dbars]

Summary

• Add a Meeting language picker with Auto and 100 named Whisper languages for new meetings and imported audio. Dictation is unchanged.
• Add bounded, recording-scoped Whisper language detection. At least two non-overlapping speech-candidate windows must agree, and every sampled observation must be confident; mixed or weak evidence keeps per-segment Auto.
• Preserve the requested language through capture/import queues, recording journals, recovery, retries, and saved Markdown. Explicit selection requires Whisper; Parakeet retains its existing model-specific behavior without a hidden model switch.
• Fix a blocker discovered during live testing: duplicate-instance handoff now presents existing controls directly instead of entering an app-modal “already running” alert that can obstruct Stop. The single-instance lock remains intact.

Stack / merge boundary

This is a draft feature PR, not release approval. Its base is the frozen combined verification snapshot codex/audio-release-verification at a2011b07ddc098ec2acf1a4f215cd21cfd17d92f.

That snapshot combines work already tracked in #1747 (audio-only permissions and Meetings creation menu), #1749 (stop-tail regression coverage), and #1751 (existing-functionality fixes). Using it as the base keeps this diff limited to the two new feature/fix commits instead of duplicating all three dependency PRs.

Do not merge this into the integration snapshot as a substitute for shipping. Land/reconcile those dependency PRs first, then rebase/retarget this PR to main and rerun the relevant gates. No dependency PR is merged by this change.

Verification completed

• App build and signed local launch smoke passed; latest launch-to-interactive result: 651.3 ms.
• Latest fast suite: 15,057 passed, 0 failed.
• Feature baseline: Core suite 1,144 reported, 13 skipped, 0 failures; app/Core integration and deterministic E2E passed; CaptureKit 54, CLI 78, MCP 190 passed.
• Full QA bench at the feature baseline: 13 passed, 0 failed, one separately-covered build skip and one warnings-only empty-artifact check. Overall report remains INCOMPLETE, not a release sign-off. The later nonmodal handoff patch was rebuilt and fast-tested separately.
• Native UI: selected Whisper Large V3 Turbo, verified cached/Ready, opened the language picker, selected Finnish, imported an audio file, and restored Auto. Six sanitized light/dark component renders are committed under .agent-review/visuals/transcription-language/.
• Supervised live Finnish video recording: captured system signal, stopped normally with Command-R, saved Finnish text, and retained audio. No stop timeout, audio gaps, or device switches in this run. The mic also heard the speakers, so this is not echo-cancellation or headphones proof.
• Same captured system track imported twice: Finnish selected → fi / explicit / fi, then Auto → auto / detected / fi. Each saved 73 words / 5 system utterances. No reference transcript was used, so this is functional language/detection proof, not a measured accuracy score.
• The live dual-track run used automaticUncertain and successfully fell back to per-segment Auto. Do not describe that run as a confident recording-wide language lock.
• Failed attempts are not counted as passes: an initial managed-browser playback attempt had no captured system signal; another attempt was interrupted after controls became inaccessible, with recovery files preserved. Native menu clicks also produced stale automation IDs after the patch; the successful Stop used the application's keyboard command.

Independent review

• Full feature diff reviewed independently against the combined base. A retry error-copy problem was corrected.
• Accepted documentation/test finding: clarify that all observed windows must agree, not a two-of-three quorum; added the [fi, fi, nil] → uncertain regression.
• Original-model retention finding not adopted: recovery/retranscription intentionally preserves existing current-model behavior. Explicit language is never silently ignored; switching to Parakeet produces guidance to select Whisper. Automatically restoring a historical model could override an intentional model change or redownload removed files. This limitation is documented.
• The three-file nonmodal handoff fix was independently reviewed against the feature commit: no actionable findings. Source-level tests are not proof of a deliberate second-process collision during active recording.

Remaining release acceptance

• Compare Auto/preselected English, Finnish, and another language against a known transcript or fluent review.
• Mixed-language speech and English introductions; short/quiet/noisy/silent input.
• Headphones/speakers and real-call route checks, with retained playback and normal stop.
• Native relaunch, queued-job, retry/recovery, and model-switch acceptance; dictation remains unchanged.
• Deliberate duplicate-process launch during capture, followed by successful Stop.
• Fresh-install/upgrade, permission denial/recovery, offline/cache/model compatibility, and signed distribution gates on the final combined candidate.

Raw recordings, transcript contents, private logs, and customer data remain local. Nothing has been merged, released, or emailed.