Make the voice end-of-turn pause longer and configurable

[arqo123]

Closes #296.

Problem

ConversationMode set the VAD redemption window from a hardcoded, level-derived table: 1300 ms for A1/A2, 1100 ms for B1/B2, 900 ms for C1/C2, 1000 ms without a level. A learner pausing to search for a word regularly exceeds that, so the utterance is closed and sent mid-sentence and the tutor answers a fragment. Nothing in Settings exposes the window, and a level assigned too high makes it shorter still.

Change

Longer automatic values. 1800 ms for A1/A2, 1500 ms for B1/B2, 1200 ms for C1/C2, 1500 ms when the level is unknown. This raises the window for every user, which is the behaviour change in this PR — the scaling-by-level idea is kept, each tier just gets more room.

A user setting. conversation_speech_pause on the user stores the window in milliseconds: 0 means automatic, and 1000 / 2000 / 3000 override it. It is offered in Settings → Conversation next to the existing session settings, validated server-side against that set, and returned by /api/auth/me.

One resolution rule. frontend/src/lib/conversation-vad.ts holds resolveVadRedemptionMs(speechPause, cefrLevel) — the stored value wins when set, otherwise the level decides. The local vadRedemptionMs() in ConversationMode is gone.

Two notes on behaviour:

• useMicVAD reads redemptionMs when the VAD is created, so a changed setting applies to the next conversation, not to a session already running.
• A subscription downgrade resets the conversation quota settings but deliberately leaves this one alone — it is a comfort preference, not a quota.

The adaptive option from the issue (keep the turn open after an obviously incomplete utterance) is not implemented; it needs transcript-level signals that are not available at the point the VAD closes a turn.

Migration

0051_conversation_speech_pause adds the column with server_default="0". It is hand-written, since autogenerate needs the Docker stack; it applies with the usual alembic upgrade head.

Tests

• backend/tests/test_conversation.py — PATCH stores a chosen pause, 0 restores automatic, an unoffered value returns 422, and /api/auth/me exposes the field with its default.
• frontend/tests/lib/conversation-vad.test.ts — automatic tiers, unknown/missing level, explicit override wins, stored values outside the offered set are ignored, and every tier is longer than the previous window.
• frontend/tests/components/ConversationSection.test.tsx — the options render, the chosen pause is sent, automatic is sent back as 0, and saving another setting preserves the stored pause.

Backend: 976 passed, coverage 85%. Frontend: 460 passed, lint and tsc --noEmit clean.

Docs

specs/phase-3-conversation.instructions.md, specs/database-models.instructions.md, specs/api-endpoints.instructions.md, specs/architecture-frontend.instructions.md and the conversation-settings line in AGENTS.md are updated. CHANGELOG.md and specs/version.md are left untouched — releases and version bumps look like maintainer territory, and version edits would conflict across parallel PRs.

Not covered here

The roadmap/ENABLE_CONVERSATION_BARGE_IN discrepancy noticed in the issue is untouched — it is a separate question about whether the roadmap or the flag is out of date.

[artcc]

Thanks for the contribution. The implementation preserves the stable-turn behavior, so speaking while Lingu is responding still does not interrupt the tutor.

Before merging, could you please review the speech-pause selector on narrow mobile screens? The four buttons use a single non-wrapping flex row, and longer German, Spanish, or Russian translations can overflow or become clipped at widths of 320–375 px. A two-column grid on mobile and four columns on larger screens would solve it.

There is also a small correction needed in the PR description: apply_subscription_quotas() resets the existing conversation settings when a subscription is activated, not when it is downgraded.

I will handle the remaining documentation and versioning updates separately.