Offer Eleven v3 alongside the v2 models, each with its own settings

[NeoVand]

What changed

Eleven v3 is now selectable. It was deliberately excluded from the catalog because it could not return character alignment, and read-along highlighting depends on it. That is no longer true — probing the live API shows both eleven_v3 and eleven_v3_conversational return alignment from /with-timestamps at pcm_24000.

The picker in Settings → Voice now groups two generations. Nothing was removed: Flash v2.5, Turbo v2.5, and Multilingual v2 all stay, and Flash v2.5 remains the default — half the credits, and steadiest on the ~280-character passages the segmenter produces. Switching to v3 is one click.

Each model carries its own settings, persisted per model id in the new elevenlabs-model-options setting. You can leave v3 on Creative while Multilingual keeps speaker boost off; the two don't interfere.

Only the knobs a model actually honours are rendered. The support matrix comes from GET /v1/models plus measurement, not from the docs:

Model	Stability	Similarity	Style	Speed	Speaker boost
Eleven v3	3 named points	—	—	—	—
Eleven v3 Conversational	3 named points	—	—	—	yes
Flash v2.5 / Turbo v2.5	0–1	yes	—	yes	—
Multilingual v2	0–1	yes	yes	yes	yes

Notable findings behind that table, all measured against the live API rather than read off documentation:

• v3 accepts speed and ignores it. 0.7 and 1.2 both produced identical 2.88 s audio, where Flash tracked them exactly (4.04 s vs 2.32 s). Exposing a speed slider for v3 would have been a lie, so it is off.
• v3 reports can_use_style: false, and plain v3 reports can_use_speaker_boost: false.
• v3 rejects previous_text/next_text with unsupported_model, so the usual long-form stitching trick is unavailable there.
• v3 caps at 5,000 characters per request against Flash's 40,000.

Why the cache is safe

This is the part most worth a careful look. Existing users must not silently re-generate (and re-pay for) audio they already have.

• While options are untouched, the request sends no voice_settings at all — byte-for-byte what the code did before this PR, which means the voice's own saved settings still apply.
• In that same untouched state elevenLabsRevision() returns the bare model id, so every pre-existing AudioVariantMeta still matches and coverage resolves exactly as before.
• Move any knob and the revision gains a sorted #k=v,… suffix. That flows through speechVariant().revision into variantSignature(), matchesCurrentVariant(), and cacheKey(), so only the tuned variant re-synthesizes.
• The suffix is cache-only. The wire always receives the bare modelId; sending eleven_v3#stability=1 would be a 400.

Setting a knob back to its default collapses the revision to the bare id again and the original cached audio is reused.

Other fixes in this PR

• Character-limit guard. A passage over the model's per-request cap now throws a readable error before the request goes out instead of burning a round trip on a 422.
• $state.snapshot when persisting. Nested $state proxies cannot be structured-cloned into IndexedDB. The first version of this code threw DataCloneError inside a floating promise: the UI updated correctly and nothing persisted, visible only after a reload. Caught by driving the live preview — types and tests were green.

Verification

• npm run lint
• npm run check — 1838 files, 0 errors, 0 warnings
• npm run test:coverage — 464 unit tests, thresholds met
• npm run build
• npm run test:e2e — 44 passed
• Live journey: imported a document, played it on Eleven v3 with Robust stability, confirmed word-level highlighting tracked correctly across passages. Switched to Creative and observed the previously-cached passage re-synthesize with the live request body { model_id: "eleven_v3", voice_settings: { stability: 0 } } — bare model id on the wire, only the field v3 honours.
• Per-model isolation checked against IndexedDB: eleven_v3 held stability: 0 while eleven_multilingual_v2 held speakerBoost: false, simultaneously, surviving reload.

New tests cover stability snapping, clamping, unsupported-knob suppression, the untouched-sends-nothing invariant, per-model request bodies, revision stability, and the character-limit guard.

Review checklist

• Privacy and network behavior are unchanged or explicitly explained. Same endpoint, same key handling, same direct browser-to-ElevenLabs call. The only request-body change is an optional voice_settings object, sent only when the user has tuned something.
• Storage/migration and model-license implications are covered. One additive setting (elevenlabs-model-options), absent-reads-as-{}, no migration needed. No stored audio is invalidated by this PR alone. No new model licenses.
• Keyboard, screen-reader, contrast, and reduced-motion behavior are covered. Sliders are native input[type=range] inside <label>; the v3 stability picker is a role="group" of buttons with aria-pressed, matching the existing model picker; the checkbox row keeps the 44px touch target. Colors reuse the existing settings tokens, so light/dark and contrast are inherited. No new animation.
• Visual changes include desktop screenshots. Not attachable from the CLI — the new panel sits under the model picker in Settings → Voice and reuses the existing .engine-model card styling, so it reads as part of that section rather than as a new surface.

🤖 Generated with Claude Code