feat(studio): consolidate community fixes and composer UI

[debpalash]

Summary

Consolidates 20 community PRs and completes the composer, sidebar and voice-card UI. Engine switches preserve active renders, and compressed audio remains usable without TorchCodec.

Changes

• Add workflow conditions, reusable Voice Design recipes and MCP tools, dictation vocabulary, profile photos, and compact composer/sidebar controls.
• Fix engine language mapping, model identity and residency, VoxCPM2 attempts, audio decoding, Windows backup cleanup, remote export history, and concurrent job events.
• Stop unsupported Intel Mac installation before downloads and offer remote-backend setup. Preserve cloning gates, all 21 locales, contributor credits, and documentation.
• Supersedes feat(llm-providers): add Cheaper Inference as a named LLM provider #2325, fix(electron): ignore the Enter that commits IME composition #2338, feat(mcp): describe_voice and design_voice tools for agent voice design #2368, Refresh Chinese README for Electron launch #2377, Fix dub assembly failing under torchaudio 2.9 when TorchCodec is absent #2379, feat(workflows): run the Condition step instead of drawing it #2380, fix(exports): gate /export/record and /export/history with require_local #2383, fix(job-store): assign append_event seq atomically to close a race #2384, fix(setup): gate Intel Macs in preflight before any download #2387, fix(tts): speak a dollar amount before sentence punctuation #2390, Fix streaming override sidecar reuse without in-flight eviction #2391, fix(pronunciation): match dictionary keys in scripts written without spaces #2392, fix(longform): name an EPUB chapter after all of its styled heading #2393, feat(dictation): vocabulary hint for Whisper-based dictation engines #2395, I’m working on #2374 by reusing the existing cached TTS engine path f… #2400, fix(design): send free-form voice descriptions as written to free-text engines #2401, fix(db-backup): never list a torn partial snapshot as a backup #2402, fix(audio): preserve streaming preview edges #2409, fix(settings): allow installed packs below reserve #2410 and fix(voxcpm2): bound sidecar generation work #2412. All current heads are included; close them after merge. Make LLM setup verifiable and add CLI, paid LLM, and translation providers #2397, Preserve audio precision and add clear voice quality controls #2406 and Complete voice quality, model-aware languages and workspace integrations #2407 were already merged separately.

Type

• New feature
• Bug fix
• Refactor
• Documentation
• Tests
• CI / Build
• Release prep

Testing

• Local bun run check:electron: 1,143 Electron tests and 2,922 shared tests passed, one skipped; typecheck, desktop/web builds, packaging contract and all 21 locales passed.
• Offline backend validation uses HF_HUB_OFFLINE=1 with a fresh empty HF_HUB_CACHE: targeted suites 735 passed/3 skipped; isolated backend suite 464 passed/8 skipped; later review regression suites 183 passed/15 skipped.
• All 19 audio-loader tests pass, including real MP3/M4A/AAC/Opus decoding, bounded memory transport, compressed-input and decoded-output size limits, temporary-file cleanup, and disk headroom. Oversized output is rejected rather than returned truncated.
• OuteTTS downmix/resampling regressions failed before the fix and pass afterward; affected engine/audio suites: 83 passed, 8 skipped.
• All four browser smokes pass: playback, dubbing, longform layout, and language selection at desktop/compact/mobile widths. Frozen Bun installation and uv lock --check pass.
• Four-platform packaging and fresh runtime installation passed on Windows, Linux, Apple Silicon and Intel Mac. The Intel smoke verifies remote setup on its unsupported local-runtime platform. Subsequent backend-only fixes bound compressed-audio staging and route OuteTTS reference preprocessing through the same decoder.
• Final-head CI passed: 9,039 backend tests, 472 isolated backend tests, 1,144 Electron tests, 2,922 shared tests, browser workflows, and all four OS smokes. Backend results also include 32 skips, 8 expected failures and one non-strict unexpected pass.
• CUDA and ROCm Docker builds and security checks passed on final head d223241a.

Earlier live MPS engine measurements remain in the handoff/history. These Windows checks do not claim new live synthesis for every optional model or a completed MPS ASR matrix. MeloTTS validates its documented manual g2p_en/NLTK prerequisites without downloading during generation.

Checklist

• Tested locally, including fail-before/pass-after regressions.
• Relevant documentation and contributor credits updated.
• No local machine paths, logs or personal environment details committed.
• Version files remain in sync; version 0.5.6 is unchanged.
• Final smoke-matrix fixture checks green on macOS, Windows and Linux.

Review disposition: all concrete CodeRabbit findings and engine/export/backup Greptile findings are fixed. Credential-free server mode deliberately permits consumption and export recording; its history redacts host paths. Configuring a key or PIN rejects anonymous access. Native file operations retain separate authorization. The lazy import-cycle note has no import-time cycle. The sidebar remains collapsible on every platform; macOS uses the existing sidebar controls. The automatic public star refresh follows the explicit owner exception in CLAUDE.md. MCP voice creation uses the existing admitted-caller write boundary, matching clone_voice. Greptile withdrew its final truncation warning after the FFmpeg documentation, muxer code, and nine real size-boundary checks confirmed the guard.

Release cadence

Continuous-to-main: this merge joins rolling source and Docker :latest. Electron rehearsals publish artifacts only. No version bump, release tag, or stable release is requested; those remain owner-controlled under the release checklist.